Editorial Board Editor-in-Chief Tunde Akim Omokanye, Peace Country Beef & Forage Association, Canada Associate Editors Esther Shekinah Durairaj, Michael Fields Agricultural Institute, United States of America Ignacio J. Diaz-Maroto, University of Santiago de Compostela, Spain Jose Manuel Brotons Martinez, Miguel Hernandez University, Spain Manuel Teles Oliveira, Universidade Tras os Montes Alto Douro, Portugal Qin Yang, North Carolina State University, United States of America Editorial Assistant Anne Brown, Canadian Center of Science and Education, Canada Reviewers Abhinaya Subedi Alfredo Jiménez-Pérez Ali Jadoo' Alsharafat Allan Lopes Bacha Amir Raza Ana Sanches Silva Andre Ricardo Machi Atinderpal Singh Benjawan Chutichudet Bindiya Sachdev Byju Govindan Caroline Valverde dos Santos Ceciliah Njoki Ngugi César A. Rosales Nieto Christian Keambou Tiambo Chunlin Yang Claude Bakoume Daniel Constantino Zacharias David González Solís Delvis Valdés Zayas Dione Dambrós Raddatz Domingo J. Iglesias Ednalva Duarte Eleonora Nistor Essiagnon John-Philippe Alavo Fabrício Menezes Ramos Filomena Fonseca Gemma Reig Córdoba George Doupis Gláucio Diré Feliciano Glayciane Costa Gois Helio Lima Neto Hiren Bhavsar Hong Tham Dong Hong-Sik Hwang Honoré Muhindo Siwako Ighodaro Ikponmwosa David Inez Naaki Vanderpuye Iuliana Vintila Jean Marc Nacife Jinlong Han Joaquim José Frazão Jose Luis Arispe Vazquez Joseph Kimutai Langat Juan José Villaverde Karel Petrzik Khairul Osman Kuldeep Srivastava Laercio Zambolim Lennin Musundire Levi Mugalavai Musalia Liria Queiroz Luz Hirano Luiz A. S. das Neves Mabhaudhi Tafadzwanashe Marcelo Zolin Lorenzoni Marcio A. S. Pimenta María Serrano Mariana de L. A. Vieira Mario Gutierrez Najla Mezghani Patrick Kelven Mashiqa Rene Murrieta-Galindo Richard N. Onwonga Richard Uwiera Rodríguez-Vicente Verónica Ruojin Zhang Semirames Silva Sixto Marquez Slawomir Borek Songjun Zeng Takashi Osada Tâmara Prado de Morais Thaise Mariá Tobal Thangavelu Muthukumar Tiago Goulart Petrolli Toshik Iarley da Silva Tugay Ayasan Vinícius Politi Duarte Willian Rodrigues Macedo Wuyi Liu Yonny Martinez Lopez jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 2; 2024 Contents Factors Affecting the Rice Yield During the Rainy Season Among Farmers in Southeastern Cambodia 1 Sar Sary, Bun Phearin, Kong Ravuth, Sar Saren, Siek Darith & Peter David Kulyakwave YOLOv5 Model Application in Real-Time Robotic Eggplant Harvesting 9 E. Kahya, F. F. Ozduven & Y. Aslan Canopy Structure Influence the Critical Period for Weed Removal of Three Cassava (Manihot esculenta Crantz) Varieties in Zambia 24 Joseph Nzunda, Nhamo Nhamo, David Chikoye, Kallunde Sibuga & Pheneas Ntawuruhunga Effects of Post-Harvest Processing Methods on the Quality of Radix et Rhizoma Salviae Miltiorrhizae 41 Huyen Thu Nguyen, Hang Thi Cu, My Thi Chu, Nguyen Thi Yen Chi, Quach Anh Kiên & Nguyen Minh Nhat jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 2; 2024 Contents Effect of Extender Supplementation With Caffeine on Arabian Stallion’s Semen Quality After Freezing 48 M. A. M. M. Shehab-El-Deen, M. A. Ali, F. H. S. Alsulaim & I. A. I. Alolayan Reviewer Acknowledgements for Journal of Agricultural Science, Vol. 16, No. 2 56 Anne Brown Journal of Agricultural Science; Vol. 16, No. 2; 2024 ISSN 1916-9752 E-ISSN 1916-9760 Published by Canadian Center of Science and Education 1 Factors Affecting the Rice Yield During the Rainy Season Among Farmers in Southeastern Cambodia Sar Sary1,2, Bun Phearin1, Kong Ravuth1, Sar Saren3, Siek Darith2,4 & Peter David Kulyakwave2 1 Research and Technology Development Centre, National Polytechnic Institute and Ministry of Labour and Vocational Training, Phnom Penh, Cambodia 2 Agricultural Information Institute, Chinese Academy of Agricultural Science, Beijing, China 3 Department of Agriculture, Forestry and Fisheries Takeo, Ministry of Agriculture, Forestry and Fisheries, Phnom Penh, Cambodia 4 Regional Polytechnic Institute Techo Sen Battambang, Ministry of Labour and Vocational Training, Battambang, Cambodia Correspondence: Sar Sary, Research and Technology Development Centre, National Polytechnic Institute and Ministry of Labour and Vocational Training, Phnom Penh, Cambodia. E-mail: sarsary@npic.edu.kh Received: October 18, 2023 Accepted: December 2, 2023 Online Published: January 15, 2024 doi:10.5539/jas.v16n2p1 URL: https://doi.org/10.5539/jas.v16n2p1 Abstract A research study utilized the Cobb-Douglas production function to examine the elements influencing paddy production during the wet season in three rural provinces of Cambodia. This analysis was based on data gathered from a survey of farmers’ households conducted in 2022. The study discovered that the use of fertilizers and herbicides, the size of the family, and income from off-farm sources significantly impacted the output of wet-season paddy. A one percent increase in the use of fertilizer, herbicide, and family size resulted in an increase in rice output by 0.06 percent, 0.04 percent, and 0.05 percent respectively. Furthermore, a one percent increase in the age of the household head, hired labor, and off-farm income led to an increase in rice yield by 0.08 percent, 0.11 percent, and 0.05 percent respectively. The use of seeds, pesticides, household labor, and the education level of the household heads were found to enhance rice yields in southeastern Cambodia. However, these production relationships varied significantly across different regions. The study concluded that higher yields during the rainy season improved the effectiveness of paddy production, primarily due to the increased responsiveness to fertilizer application. Keywords: rice production, farmers in south-eastern, Cobb-Douglas production function 1. Introduction The world’s population is expected to reach 9 billion by 2050, up from 6.1 billion in 2000 (FAOSTA, 2015). This rapid increase puts a lot of pressure on food supply and food security, requiring a better understanding of agricultural growth and productivity to meet the growing demand for food. However, agriculture is especially sensitive to the impacts of climate change, making it more vulnerable than other sectors. Climate change could reduce crop yields by more than 25 percent (Sokvibol et al., 2016). Agriculture is the backbone of Cambodia’s economy. Most Cambodians live in rural areas and depend largely on subsistence farming for their livelihoods. The government has implemented several reforms in recent years to improve the efficiency and effectiveness of its extension and advisory services (Samoeurn & Suresh, 2018). The agricultural sector contributed 24.4% of the country’s GDP in 2022, while the services and industries sectors contributed 37% and 38.5%, respectively. The reason for the lower contribution of the agricultural sector is the growth of the other two major economic sectors, namely industry and services. It is important to note that between 2017 and 2021, the contribution of the agricultural sector decreased from 24.9% to 24.4%, mainly due to the rapid growth of the industry, construction, and service sectors (MAFF, 2021). Rice production is a key component of Cambodia’s agricultural sector and plays a crucial role in driving national economic growth, earning it the nickname “white gold” among the Cambodian people (Sokvibol et al., 2016). The Royal Government’s strong support, along with the active participation of farmers, has been instrumental in increasing rice production. A notable change has been observed among Cambodian farmers, who have shifted from subsistence farming to commercially oriented production in jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 2 response to better market demand and higher prices. Wet rice production dominates Cambodian rice production, accounting for 80% of the total rice population and occupying approximately 85% of the harvested land area. Wet-season rice depends heavily on rainfall due to limited irrigation capabilities. Farmers usually start rice production in May and continue harvesting until December or later, depending on the rice variety. The Ministry of Agriculture, Forestry, and Fisheries classified the wet season into five types in 2015: early, medium, late, upland, and floating rice, based on the cultivation-to-harvest timeframe and topographical conditions. Early rice varieties generally have a shorter growing period than medium rice varieties, while late rice varieties require the longest time to mature. Floating rice is typically cultivated in flooded areas surrounding the Tonle Sap Lake and the Mekong River, while upland rice is commonly grown in mountainous regions of the Plateau/Mountain area (Rido 2015). This study aims to analyse the factors influencing rice yield during the wet season among farmers and explore its potential as a policy tool for rural development. The study focuses on farming households operating in the southeastern region of Cambodia, encompassing Takeo, Kandal, and Kampong Speu provinces. The insights gained from this study are intended to benefit both farmers and policymakers, guiding efforts to enhance Cambodian rice production. The rest of this paper is organized into five sections: introduction, research methodology, data and description, results and discussion, and conclusion. 2. Method 2.1 Empirical Model The study made use of the Cobb-Douglas production function, with the aid of STATA software, to identify the significant factors influencing the input in wet season rice. The Cobb-Douglas function, a widely accepted model for illustrating the relationship between input and output, was originally proposed by Knut Wicksell (1851-1926) and later validated through statistical evidence by Charles Cobb and Paul Douglas (Coelli et al., 1997 as cited Nhat & Tansuchat, 2015). Several researchers have conducted studies in this field. For example, Pisedh et al. (2011) investigated the challenges of dry season rice production under the irrigation scheme of Tapeing Thmor Water reservoir. Yu and Fan (2009) examined rice production response in Cambodia, while Tun and Kang (2015) analyzed the factors affecting rice production efficiency in Myanmar. Ahmed et al. (2017) conducted a comparative study on factors influencing rice yield in Niger State of Nigeria and Hainan of China, and Koirala et al. (2014) researched the determinant of rice productivity and technical efficiency in the Philippines. Furthermore, Sokbibol et al. (2017) compared Cambodian rice production technical efficiency at the national and household level. Rido (2014) studied factors affecting the cost efficiency of Cambodian rice farming households, and Sokchea and Richard (2013) researched ways to improve rice productivity and farmers’ income in Cambodia. Rada et al. (2016) studied the profit efficiency of rice farmers in Cambodia. Rohmad and Praptiningsih (2016) analyzed the efficiency of production factors utilization in upland rice farming, and Hossain and Majumder (2015) measured the efficiency of the Cobb-Douglas production function with additive and multiplicative errors. Despite extensive research in this field, no study has specifically focused on the factors affecting rice yield in the wet season. Therefore, this study aims to fill this gap and contribute to increasing rice production in Cambodian farming. In the study, the Cobb-Douglas production function was applied as follows: Y�=�AK�1L�2 (1) where, Y = the total output; L = labor input; K = capital input; A = is constant; �1 and �2 are the coefficients to be estimated for labor and capital, respectively. Equation (1) is nearly always treated as a linear relationship by making a logarithm transformation, which yields: lnY�=�lnA�+��1lnK�+��2lnL (2) According to Equation (2) with independent variables L and K to i become: lnY�=��0�+��1lnK�+��2lnL�+�…�+��ilnXi (3) And decoding Equation (3) according to this study we have: lnY�=��0�+��1lnseed�+��2lnfert�+��3lnpest�+��4lnweed�+��5lnhom�+ �6lnhired�+��7lnage + �8lnfamily�+��9lnedu�+��10lnincome�+ �i (4) where, lnseed: logarithm of seed; ������: logarithm of fertilizer; lnpest: logarithm of pesticide; lnweed: logarithm of weedicide; lnhom: logarithm of household labor; lnhired: logarithm of hired labor; lnage: logarithm of the age jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 3 of household head; lnfamily: logarithm of family size; lnedu: logarithm of education of household head; �� � ���: logarithm of income off-farm job. The coefficient �1, �2, �3, … �i are the elasticity yield to input L, K,… and i. The sum of elasticity �1�+��2�+��3+�…�+��i, provides the returns to scale of the farms in question. It means if: �1�+��2�+��3�+�…�+��i�=�1, the production operates under constant returns to scale. �1�+��2�+��3�+�…�+��i�>�1, the production operates under increasing returns to scale. �1�+��2�+��3�+�…�+��i�<�1, the production operates under decreasing returns to scale. 2.2 Data Description This study involved the examination of data collected from a household survey carried out in 2022 across three provinces in Cambodia. A total of 240 farmer households were selected for the study using a random sampling technique. The data collection process was spearheaded by the authors, who were assisted by postgraduate students from AII-CAAS and graduate students from the National Polytechnics Institute of Cambodia (NPIC). The data collected covered various aspects of rural farmers, including household conditions, farming income, daily expenditure, inputs for rice production, and information on agricultural technology. To facilitate face-to-face interviews with household farmers and stakeholders, the authors reached out to local authorities, such as the chief of wards and chief of communes. The inputs for wet season rice included seeds, pesticides, herbicides, fertilizer, irrigation, household labor, hired labor, and transportation, among others. The increase in rice inputs positively impacted wet-season rice production, leading to a yield increase of about 7.637 million tons and 2.315 million tons for dry-season rice production in 2016. This surge in rice production was largely attributed to the support provided by the Royal Government of Cambodia, relevant ministries and institutions, development partners, national and international organizations, sub-national authorities, and the participation of farmers. To reduce bias in sample selection, the study only included farming households, while mixed farmers, paddy producers, and other crops were excluded from the data analysis. Data modification and filtering were carried out to ensure that the unit of measure of each variable was consistent with the academic goals and that the data quality was satisfactory. 3. Results and Discussion 3.1 Statistical Summary The statistical summary of variables in rice inputs and rice yield was estimated using multiple regression analysis. Table 1 includes variables such as seeds, pesticides, weedicides, fertilizer, household labor, and hired labor. Household characteristics such as the age of the family head, education level of the household head, family size, and off-farm income of the household head were also considered. The table displays the average, standard deviation, minimum, and maximum value of each variable. On average, the output of wet rice was about 1.7 tons, with a minimum and maximum of approximately one ton and 2.5 tons, respectively. According to MAFF (2015), the rice yield in the wet season was around 3.21 tons per hectare, which is larger than the result of this study. This discrepancy could be because the MAFF report covered the whole country, while this study focused only on three provinces. Farmers used an average of about 175 kg/ha of variety in the wet season, with a minimum and maximum of approximately 42 kg/ha and 400 kg/ha, respectively. According to the Cambodian Agriculture Value Chain Program CAVAC (2016), farmers typically used an average of 134 kg for wet season varieties, or up to 400 kg/ha. Furthermore, Sothy et al. (2017) showed that dry-season rice varieties required 322.1 kg, while wet-season varieties required 122.2 kg. jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 4 Table 1. The statistical Summary inputs of the wet rice per hectare Variables Units Obs Mean Std. Dev. Min Max Rice outputs Kg/ha 240 1,699 274 1,000 2,500 Seed Kg/ha 240 175 88 42 400 Pesticides ml/ha 240 2,149 1,012 500 6,600 Weedicides ml/ha 240 1,184 659 480 3,500 Fertilizer Kg/ha 240 201 71 100 450 Household labor person/ha 240 8 3 3 17 Hired labor person/ha 240 7 2 1 13 Age_HHhead year 240 49 12 22 88 Education_HHhead level (0-5) 240 2 1 0 5 Family size person 240 5 1 2 9 Income off-farm KHR/ha 240 2,886 2,897 - 4,000 The study utilized multiple regression analysis to estimate the variables in rice inputs and rice yield. The average application of pesticide was approximately 2,149 ml/ha, with a range of about 500 to 6,600 ml/ha for the wet season rice. The average application of weedicide was roughly 1,184 ml/ha, with a range of around 480 to 3,500 ml/ha. When compared, the average usage of pesticides and weedicides in the wet season was lower than that applied in the dry season rice. The average application of fertilizer was around 201 kg/ha, with a range of about 100 kg to 450 kg/ha. Our results showed that farmers used more fertilizer compared to the research of Dary et al. (2016), which revealed that farmers apply about 153 kg/ha of chemical fertilizer for dry season rice, while only 66 kg/ha in the wet season. During the study period, the average household labor was nearly 8 people per hectare, with a range of about 3 to 17 people. Sok Vibol et al. (2016) demonstrated that the average adult working day per year of family members was around 108 days in 2013, increased to 110.5 days in 2014, and then decreased by 106.7 days in 2015. The statistics summary revealed that the average age of the household’s head was 49 years old, ranging from 22 to 88 years old. The average education level was 2, implying that most of the farmers’ household heads obtained education at primary school (grades 1-6 in the Cambodian education system). The results also displayed that the average family size of peasant households in Takeo, Kampong Speu, and Kandal is around 5 persons per household, ranging from 2 to 9 persons per household. The average income from the off-farm job is approximately 2,886 thousand riels, with a maximum of 4,000 thousand riels. Based on research, farmers with small land cultivated l are not able to support family members throughout the year. Consequently, most farmers take different actions to generate income to support their household family. It is not just farming activities. The peasants find various works to increase their household income. 3.2 The Estimated Parameters of Rice in the Wet Season Table 2 shows that all estimated parameters of the variables align with expectations. Fertilizer had a positive coefficient and was significant at the one percent level. Weedicides, irrigation systems, family size, and off-farm income also had positive coefficients but were significant at the five percent level. Hired labor and the age of the household head were significant at the ten percent level. The seed input had a negative coefficient but was significant at the ten percent level, implying a contrasting relationship between varieties and rice output during the research period. The study also showed no significant correlation between pesticide input and wet-season paddy output. Similarly, the education level of the household head had a negative coefficient and showed no significance at any level, indicating no correlation between the education of the household head and rice yield during the study. The Cobb-Douglas production function was used to estimate the parameters of inputs affecting the wet season paddy. The results showed that increasing the quantity used of fertilizer, weedicides, hired household labor, age of the household head, and off-farm income could lead to an increase in household farming rice output. For instance, a one percent increase in the input of seed would decrease the rice yield of the wet season by approximately 0.07 percent. A one percent increase in the input of weedicides will increase the rice yield by about 0.04 percent. The rice yield will increase by approximately 0.06 percent if there is a one percent increase in the input of fertilizer. A one percent increase in the input of hired labor will increase the rice output by about jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 5 0.11 percent. An increase of one percent in the input of the age of the household head will increase the rice output by roughly 0.08 percent. A one percent increase in family size would result in a 0.05 percent increase in wet paddy production. The rice output will increase by approximately 0.05 percent if there is a one percent increase in off-farm job income. Table 2. The parameter of Cobb-Douglas for wet-season paddy Variable Coefficient Standard Error t-ratio P > t Constants �0 6.5708 0.5095 12.9000 0.000*** lnseed �1 -0.0720 0.0394 -1.8300 0.074* lnpesticides �2 -0.0281 0.0585 -0.4800 0.633 lnherbicides �3 0.0434 0.0845 4.5100 0.031** lnfertilizer �4 0.0587 0.0865 8.6800 0.002*** lnhousehold_labor �5 -0.0503 0.0778 -0.6500 0.521 lnhired_labor �6 0.1114 0.0724 1.5400 0.053* lnage_household_head �7 0.0797 0.0677 1.1800 0.065* lnfamily_size �8 0.0485 0.0601 2.8100 0.024** lneducation �9 -0.0248 0.0436 -0.5700 0.571 lnincome_off �10 0.0469 0.0196 4.3600 0.024** Prob > F = 0.0355 R-squared = 0.3284 Adj R-squared = 0.1745 Root MSE = 0.1386 Note. ***, ** and * indicate significance at 1%, 5% and 10% respectively. The majority of Cambodian farmers can grow rice just once a year due to insufficient irrigation and effective water management practices. The rainy season is the primary period for rice cultivation, accounting for 80 percent of the total crop. Irrigation facilities are primarily utilized for the dry season rice and to supplement the wet season paddy if necessary (Smith & Hombuckle, 2013). Yu and Diao (2011) argued that Cambodia has tremendous potential to boost rice production given its abundant agricultural land and water resources. If the rice fields are irrigated, production will increase. Regarding fertilizer application, Yu and Fan (2009) demonstrated that the percentage of farming households using fertilizers rose for both seasons of rice production. On average, about 78 percent of rice in the wet season plots used fertilizers, and this percentage increased to 88 percent in 2007. In summary, the use of fertilizer has the most significant impact on paddy rice production, followed by herbicides, household size, non-farm income, seed, hired labor force, and the age of the household head. The production of wet-season rice is affected when farmers use higher volumes of these inputs. 4. Conclusions The findings above indicate that factors such as fertilizers, herbicides, family size, and off-farm job income significantly influence the output of wet-season paddy. A one percent uptick in the use of fertilizers, herbicides, and family size could enhance rice production by 0.06, 0.04, and 0.05 percent respectively. Similarly, a one percent rise in the age of the household head, hired labor, and off-farm job income could boost rice yield by 0.08, 0.11, and 0.05 percent respectively. Elements like seeds, pesticides, household labor, and the education level of the household head also contribute to rice production in Cambodia, although the relationships vary by region. Simulations show that high yields in the rainy season, even when doubled due to high fertilizer responsiveness, enhance the efficiency of paddy production. These insights have significant implications for boosting rice production in Cambodia. Farmers stand to gain the most from improvements in agricultural productivity and technology. Cambodian farmers need to concentrate on the agricultural sector to drive growth in rice production and alleviate poverty in rural areas. Given that most of Cambodia’s poor reside in rural areas and depend on agriculture, high agricultural growth will ensure food security by increasing supply, reducing prices, and raising the incomes of poorer farm households. To facilitate this jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 6 response and achieve food security, agriculture, which has been neglected, has been included on the political agenda. Firstly, there is considerable potential for enhancing rice production in Cambodia. If appropriate resources (like fertilizer and irrigation) and infrastructure (like electricity markets, agricultural extension, and education) are provided, it is possible to elevate Cambodian rice output to match that of its neighboring countries. Given the high awareness of fertilizer, farmers could significantly increase their yield and revenue from more market sales. Secondly, the introduction of advanced technologies and crop diversification should be tailored to local conditions. However, poor road and market conditions hinder local producers from capitalizing on the comparative advantage of rice production. Greater investment in infrastructure would enable farmers to access the latest market information and transport their produce to Phnom Penh and other regional markets. Investments in rural roads have been shown to yield high returns in terms of poverty reduction in developing countries. Improving rural roads will enable rural people to access essential services. References Ahmed, A. G., Xu, S., Yu, W., & Wang, Y. (2017). Comparative study on factors influencing rice yield in Niger State of Nigeria and Hainan of China. International Journal of Agricultural and Food Research, 6(1), 13-23. https://doi.org/10.24102/ijafr.v6i1.724 CAVAC (Cambodian Agriculture Value Chain Program). (2016). Farmers’ Behavior on Using Rice Seed in Cambodia. Phnom Penh: Cambodian Agriculture Value Chain Program. CDRI. (2011). Irrigation Water Productivity in Cambodian Rice Systems. Coelli, T. J., Rao, D. S. P., O’Donnell, C. J., & Battese, G. E. (1997). An introduction to efficiency and productivity analysis. Springer Science & Business Media. https://doi.org/10.1007/978-1-4615-5493-6 Dary, P., Sokcheng, S., & Pirom, K. (2017). Synergies and Trade-offs with Intensification of Rice and Livestock Production in Cambodia (Special Report, June 16, 2017). CDRI. Fan, S. (2008). Public Expenditure, Growth and Poverty in Developing Countries: Issues, Methods, and Finding. Jonh Hopkins Publications, Baltimore. FAOSTAT. (2015). FAOSTAT Online Database 2015. Food and Agriculture Organization of the United Nations. Hossain, M. M., & Majumder, A. K. (2015). On the measurement of efficiency of Cobb-Douglas production function with additive and multiplicative errors. Department of Statistics, Jahangirnagar University, Savar, Dhaka, Bangladesh. https://doi.org/10.19139/soic.v3i1.120 Koirala, K., Mishra, A. K., & Mohanty, S. (2014). Determinants of rice productivity and technical efficiency in the Philippines. Department of Agricultural Economics and Agribusiness, Louisiana State University, USA. MAFF. (2015). Annual Report for Agriculture Forestry and Fishery 2014-2015. The Kingdom of Cambodia. MAFF. (2021). Annual report for agriculture forestry and fisheries 2020-2021 and direction 2021-2022. The Kingdom of Cambodia. Nhat, V., & Tansuhat, R. (2015). Analysis of rice production and contribution to Cambodian economic growth. The Empirical Econometrics and Quantitative Economics Letters, 4(2), 17-30. Pisedh, V., Christophe, J. D., Seng, M., & Songly, U. (2011). Challenges of dry season rice production under irrigation scheme of Tapeing Thmor water reservoir. International Journal of Environment and Rural Development, 2(2). Rada, K., Nanseke, T., & Chomei, Y., (2016). Profit efficiency of rice farmers in Cambodia. The differences between organic and conventional farming. Journal of Sustainable Development, 9(6), 34-45. https://doi.org/10.5539/jsd.v9n6p34 Rido, T. (2015). Productivity and efficiency of rice production: the implication for poverty alleviation in Cambodia (PhD dissertation, Graduate School of International Development, Nagoya University, Japan). Rohmad, B., & Praptiningsih, G. A. (2017). Efficiency analysis of production factors utilization in upland rice farming in NRLS Conference Proceedings. KnE Life Sci, 2017, 180-187. https://doi.org/10.18502/ kls.v3i5.991 Samourn, K., & Suresh, C. B. (2018). Agricultural Extension in Cambodian: An Assessment and Options for Reform (IFPRI Discussion Paper 01706, January 2018). jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 7 Smith, D., & Hornbuckle, J. (2013). A Review of rice productivities in Cambodia and water measurement using direct and indirect methods on a dry season rice crop. Technical Report to ACIAR, Canberra, CSIRO Sustainable Agriculture Flagship, Australia. Sokchea, A., & Richard, C. (2013). Improving rice production and farmers’ income in Cambodia: An econometric estimating using data from the APSIM simulator. Sokvibol, K., Hua, L., & Linvolak, P. (2016). Technical Efficiency Analysis of Cambodian Household’s Rice Production. Global Journal of Human-Social Science, 16(3), 33-44. Sokvibol, K., Hua, L., & Linvolak, P. (2017). Comparison of Cambodian rice production technical efficiency at the national and household levels. Sothy, E., Sokcheng, S., Chhun, C., & Pirom, K. (2017). Rice policy study: Implication of Rice policy Changes in Vietnam for Cambodia’s Rice policy and Rice Producers in South-Eastern Cambodia. Cambodia Development Resource Institution. Tun, Y. Y., & Kang, H. J. (2015). An Analysis of the factors affecting rice production efficiency in Myanmar. Journal of East Asian Economic Integration, 19(2), 167-188. https://doi.org/10.11644/KIEP.JEAI.2015. 19.2.295 Yu, B., & Diao, X. (2010). Cambodia’s Agricultural Strategy: Future Development Options for the Rice. Yu, B., & Fan, S. (2009). Rice production response in Cambodia. International Food Policy Research Institute. Acknowledgments The completion of this research owes a great deal to the backing of the Chinese Academy of Agricultural Sciences (CAAS) and the National Polytechnic Institute of Cambodia (NPIC). The authors extend their heartfelt thanks to the postgraduate students of AII-CAAS and the graduate students of NPIC for their unwavering collaboration in gathering data. Special appreciation is also expressed for Prof. Xu Shiwei and Prof. Yu Wen, whose assistance in data analysis was invaluable. Authors Contributions Not applicable. Funding The article was supported by the National Polytechnic Institute of Cambodia Project (Number: 6110-61106). Competing Interests No potential conflict of interest was reported by the author(s). Informed Consent Obtained. Ethics Approval The Publication Ethics Committee of the Canadian Center of Science and Education. The journal’s policies adhere to the Core Practices established by the Committee on Publication Ethics (COPE). Provenance and Peer Review Not commissioned; externally double-blind peer reviewed. Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. Data Sharing Statement No additional data are available. Open Access This is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/4.0/). jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 8 Copyrights Copyright for this article is retained by the author(s), with first publication rights granted to the journal. Journal of Agricultural Science; Vol. 16, No. 2; 2024 ISSN 1916-9752 E-ISSN 1916-9760 Published by Canadian Center of Science and Education 9 YOLOv5 Model Application in Real-Time Robotic Eggplant Harvesting E. Kahya1, F. F. Ozduven2 & Y. Aslan3 1 Control and Automation Technology Programme, Department of Electronic and Automation, Vocational College of Technical Sciences, Tekirda� Nam�k Kemal University, Tekirda�, Turkey 2 Greenhousing Program, Department of Plant and Livestock Production, Vocational College of Technical Sciences, Tekirda� Nam�k Kemal University, Tekirda�, Turkey 3 Freelance Senior Software Developer, Tekirda�, Turkey Correspondence: E. Kahya, Control and Automation Technology Programme, Department of Electronic and Automation, Vocational College of Technical Sciences, Tekirda� Nam�k Kemal University, Tekirda�, Turkey. Tel: 90-506-718-8820. E-mail: ekahya@nku.edu.tr Received: November 24, 2023 Accepted: December 22, 2023 Online Published: January 15, 2024 doi:10.5539/jas.v16n2p9 URL: https://doi.org/10.5539/jas.v16n2p9 Abstract Deep learning studies in agricultural automation have accelerated in recent years due to its benefits such as increasing product efficiency and reducing labor force. Deep learning is a powerful tool for automation in agriculture with applications ranging from disease identification and crop yield detection to fruit ripeness classification. It helps to automate various processes in agriculture and to perform time-consuming tasks in a shorter time. It quickly processes the data required for robotic harvesting systems and makes it available to the system. In this study, a machine learning study was carried out to be used in the robotic harvesting of eggplant fruit, which is a product that can take time to select and collect in the agricultural area where it is cultivated. YOLOv5 (nano-small-medium and large models) was used for the deep learning method. All training and test metric values of the models were analyzed. It was determined that the most successful model was the model trained with YOLOv5m algorithm on images of 640 × 640 size with 12 Batches and 110 Epochs. The results of the model values were analyzed as “metrics/precision”, “metrics/recall”, “metrics/mAP_0.5” and “metrics/mAP_0.5:0.95”. These are key metrics that measure the detection success of a model and indicate the performance of the relevant model on the verification dataset. It was determined that the metric data of the “YOLOv5 medium” model was higher compared to other models. The YOLOv5m model gave the highest score with F1 score of 85.66%, precision of 95.65%, recall of 96.15%, and mAP at 0.5:0.65 of 78.80%. Hence, it was understood that “Model 3” was the best detection model to be used in robotic eggplant harvesting to separate the eggplant from branch. Keywords: deep learning, eggplants, classification, YOLOv5 1. Introduction Deep learning is a subfield of artificial intelligence that focuses on the development of neural networks and deep neural architectures to solve complex tasks. This method aims to process and learn data using mathematical model-based systems called artificial neural networks. It has the ability to learn based on large amounts of data and can be used to solve complex problems (Çetiner et al., 2022). It is stated that artificial neural networks consist of multi-layered structures. Each layer of this multi-layered structure can learn features at successive levels by processing input data. Each layer learns more abstract and higher-level features than the previous one. They are called “deep” because they have multiple hidden layers between the input and output layers. Each layer in a deep neural network enables the model to learn complex patterns and representations by incrementally extracting higher-level features from the input data. Deep learning can model complex relationships in data sets by means of the multi-layered structure of artificial neural networks. Each layer of this multilayer structure increases the representation power by transforming the input data into higher-level features. Thus, deep learning models can automatically learn more complex and abstract features and recognize patterns in data sets (Akta�, 2022). jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 10 Deep learning is widely used in various fields and applications. One of the important areas where deep learning is used is medical imaging. It is used in medical imaging, especially MRI analysis. These methods show promise in improving clinical applications and have been applied in tasks such as image segmentation, disease classification, and anomaly detection (Lundervold et al., 2019). Another area where deep learning finds wide application is the automation of industries. Ayadi et al. (2022) emphasized the use of deep learning-based soft sensors to increase automation flexibility in industries. Furthermore, deep learning is used in the field of artificial intelligence, especially in natural language processing (NLP). Deep learning models such as repetitive neural networks (RNNs) and convertors have been successful in tasks such as machine translation, sentiment analysis and text generation. These models have revolutionized the field of NLP by capturing complex linguistic patterns and semantic relationships. Deep learning has made significant contributions to computer vision. Convolutional neural networks (CNNs), a type of deep learning model, have achieved remarkable results in tasks such as object detection, image classification, and face recognition. CNNs enable accurate and efficient analysis of images and videos by automatically learning hierarchical representations of visual data (LeCun et al., 2015). Deep learning is used in many applications in agriculture. One of the most common applications is the detection of diseases and pests in plants. Deep learning models have been used to develop image recognition systems that can identify diseases and pests in crops based on leaf or plant images (Zhang et al., 2020; Liu et al., 2020). These models can help farmers detect and diagnose plant diseases early, enabling timely intervention and preventing crop losses. Another application of deep learning in agriculture is precision agriculture. By analyzing data from various sources such as satellite imagery, weather data and soil sensors, deep learning models can provide insights and recommendations to optimize crop management practices. This includes tasks such as crop prediction, irrigation scheduling and nutrient management (Ampatzidis, 2018; Jin et al., 2020). Deep learning is also used to increase the efficiency of agricultural operations. It can be used for weed detection and classification, thus, targeted and precise weed control precautions can be taken (Cicco et al., 2017). Another use of deep learning in agriculture is classification. Models have been used for automatic fruit and vegetable classification, allowing faster and more accurate classification according to quality and ripeness (Aji et al., 2019). Deep learning models can enable robots and drones to perform tasks such as autonomous harvesting, crop monitoring, and autonomous spraying (Ampatzidis et al., 2017). These technologies can increase productivity, reduce labor costs and minimize the use of agricultural chemicals. Deep learning is not only used in precision agriculture applications. It has also found a place in livestock management. By analyzing data from sensors and cameras to monitor animal behavior, health, and welfare, it can help farmers detect abnormalities, predict disease outbreaks, and optimize feeding and rearing practices (Umar et al., 2022). In general, deep learning has the potential to revolutionize agriculture by enabling data-driven decision-making, increasing efficiency and reducing environmental impact. However, successful application of deep learning in agriculture requires handling challenges such as data collection and processing, model interpretability, and ethical issues (Dara et al., 2022; Ryo et al., 2022). Eggplant is a vegetable that is widely grown in the world and in almost every region of our country. It is a plant that has a significant share not only as a summer vegetable but also in greenhouse cultivation in our country. Correctly determining the harvesting time of eggplant, one of the important vegetables of Turkish cuisine, will provide ease of marketing and will also affect its shelf life. In this context, determining the harvesting time of eggplant correctly and harvesting at the right time is one of the important parameters. 2. Material and Method 2.1 Material Eggplant belongs to the Solanaceae family and its homeland is known to be India-Burma and Assam. When vegetable cultivation in the world and Turkey is considered, eggplant (Solanum melongena L.) is among the most produced, consumed and economically high species (E�iyok, 2012). Eggplant is rich in vitamins, minerals, has a high antioxidant capacity and is rich in phenolic acids. It is called ‘egg-plant’ because its fruit shape and color look like an egg (Sao et al., 2010). The fruit shape, fruit color, and fruit size show a very large variation in eggplants. Eggplant fruit forms are long, medium long and round. There is a big difference between the harvesting size and the size of the fruit whose seeds have ripened. When a fruit that has reached harvesting ripening is left as a seed, it reaches 4-6 times the harvest size and weight (Uzun et al., 2000, Vural et al., 2000). The harvesting process of eggplant involves several aspects that will affect fruit quality and yield. Researches have shown that the fruit ripening stage at harvesting time also affects fruit quality (Passam et al., 2010). The harvesting criteria of eggplant are affected by several factors such as harvest season, fruit ripeness and environmental conditions. It is known that the harvest season is related to the amount of phenolic acid in eggplant, indicating that harvest timing may affect the nutritional quality of the fruit (Gürbüz et al., 2018). In addition to the developmental stage of the plant and harvest time, fruit type, shape and size are also important jas.ccsenet. items in d recommen generally (Msogoya at the appr has been o and the im 2012). The quality (Sa 2.2 Method 2.2.1 Labe Labeling i increases detection m For this re greenhous Figure 1. After the m The labelin Roboflow of image-b graphical selected fo the class i shown in F org determining t nded to harvest estimated at 7 et al., 2014). T ropriate stage observed that t mportance of c e limited shelf antacatalina et d eling in deep learnin the training a model on a dat eason, the part es in Tekirda� Figure 1. Sa markings were ng process wa is a web platf based artificia user interface or the model to information an Figure 2. the optimum t the fruits bef 70-90 days fro The harvesting of ripening (X the position of onsidering the f life of eggpl al., 2016). ng is the proce and classificat taset, the objec ts containing th � Naip Village amples of imag e made, it was as performed on form designed al intelligence e of the websi o be used. Whe nd coordinate Journal of A harvest time fore the skin h om the date o g method of eg Xu et al., 2020 f the fruit on t ese factors dur lant after harv ess of assignin tion performan cts targeted to he eggplant im were marked ges taken from s assigned to t n Roboflow, a to simplify an projects. Thi ite. Within the en the text file information o Agricultural Sci 11 (Ferreira-Sant hardens and the of planting, de ggplant is very 0). In terms of the plant and t ring the harves vesting require ng images in th nce of deep be detected m mage in each o with a boundi m producer gre the “eggplant” a popular webs nd accelerate t s marking and e scope of the of any image of the marking ience tos et al., 20 e original vari epending on v y important. Eg f the physiolog the time of har sting process h s careful post- he data set to learning mode must first be lab of the 52 plant ng box area. E enhouses in Te ” class, which site frequently the data prepar d labeling pro e study, the Y containing the g and labeling 021). In Afric iety color chan varieties and w ggplants are ty gical quality of rvest may affe has been emph -harvest handl the correct cla els. In order beled in the da ting images ta Examples of im ekirda� Naip V is the object used in object ration and mo ocess is easily YOLO deep le e labeled objec g will be liste Vol. 16, No. 1; can eggplant, nges. This peri weather condi ypically handpi f eggplant see ect the seed qu hasized (Reis e ling to mainta asses. This pro to train an o ataset to be tra aken from prod mages are show Village class it belong t detection proj odel training ph y done through earning model ct classes is en ed. Label scre 2024 it is iod is itions icked ds, it uality, et al., in its ocess object ained. ducer wn in gs to. jects. hases h the was tered, en is jas.ccsenet. 2.2.2 Train In the stud developed that has be et al., 202 detection a and uses th fast detect used in ma objects by detection o YOLOv5l agriculture in UAV im detection agriculture the same a a robotic targeting developed performan field, Tsai indicated t given as a better perf detection accuracy o tested the Universal detection a terms of a application learning tr 2.2.3 Initia To start tra model on org ning Model Se dy we carried d by the CNN een widely stu 3). YOLOv5 i algorithm in co he PyTorch fra tion speed and any fields. In f adding a layer on defective U and found tha e and remote s mages and dete method based e, YOLOv5 wa accuracy as YO sprayer system specific objec d for ground o nce indicators s (2023) grade the efficiency an example of formance in in network for r of YOLOv5 w impact of var Adversarial P algorithm that accuracy and ns. In the stud raining. ation of Trainin aining the mod the computer election d out, the YOL method, was p udied and deve is one of the v omputer vision amework inste high accuracy fabric defect d r of ultra-smal UAVs, Nepal e at YOLOv5l h sensing. Batin ected false ears d on improved as used for au OLOv4 was ac m for real-tim cts was indica object recogni such as precisi d AB-YOLOv of YOLOv5 in another usage ntelligent defec robust vehicle as also been ev rious cyber-att Perturbations o t has been dev precision. Its dy, YOLOv5n/ ng del that will de is to open the Journal of A Figure 2 LOv5 family, preferred. YO eloped in vario versions of the n. It is conside ead of the Dark y (Sozzi et al., detection, Shi e ll detection hea et al. (2022) co had the best p et al. (2023) p s caused by oc d YOLOv5 an utomatic cluste chieved (Sozzi me tobacco rec ated (Nasir et ition in remot ion, recall and v5 and PB-YO n this applicati e area. In their ct detection in e detection an valuated in the tacks such as L on the accura veloped and ap s speed and d /s/m and l (nan etect the eggpl e Python execu Agricultural Sci 12 . Labeling scre developed as OLOv5 deep le ous fields. It al e YOLO (You ered a lightwei knet framewor 2022). Due to et al. (2022) im ads to YOLOv ompared the ac performance. Y proposed an im cclusion condit nd indicated i er detection in i et al., 2022). cognition and t al., 2023). X te sensing ima d mAP. Anothe OLOv5 for rib ion. The resear r study, they e n asphalt pavem nd obtained sig e context of cy L-BFGS, FGS cy of YOLOv pplied in vario detection perfo no-small-medi lant on the see utable editor. A ience een s an open sou earning model lso has high de Only Look O ight version co rk (Liu et al., 2 o this feature, mproved the d v5s. In the cont ccuracy and sp YOLOv5 was mproved YOLO tions. Zhao et its efficiency white grape v In precision a spraying, and Xue et al. (2 ages and evalu er use of Yolov b fracture dete rch conducted emphasized th ments. Kumar gnificant resu yber-attacks on SM, C&W, BI v5. In general ous fields and ormance make ium and large) edling, the nex After opening urce of the YO is an object d etection speed Once) algorithm ompared to the 2022). YOLO YOLOv5 deep detection accur text of autonom peed of YOLO also used in o Ov5 method to al. (2021) dev in UAV imag varieties, and ra agriculture, YO d its efficiency 022) used a uated the dete v5 is in medic ction in chest d by Carvalho e hat the YOLOv r et al. (2023) ults on benchm n the Internet. Z IM, PGD, One l, YOLOv5 is d has shown pr e it a popular ) models were xt step after run the editor, the Vol. 16, No. 1; OLO model fa detection algor d and sensitivit m, a popular o e previous ver v5 is known f p learning mod racy of small t mous landing Ov3, YOLOv4 other fields suc o detect wheat veloped a whea ges. In the fie apid detection OLOv5 was us y in detecting YOLOv5 net ection effect u cine. In the me X-ray images et al. (2022) ca v5 architecture used the YOL mark datasets. Zhang et al. (2 e Pixel Attack s a versatile o romising resu r choice for m preferred for nning the YOL e train.py prog 2024 amily rithm y (Li object sions for its del is target point , and ch as t ears at ear ld of with ed in g and work using edical s and an be e had LOv5 The 2022) , and object lts in many deep LOv5 gram, jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 13 which is in the main directory and provides the YOLOv5 training, was run. The last step is to run the Python program and customize it with parameters values. Within the study, the parameters and regulations in the code written below were preferred. python train.py --img 640 --batch 12 --epochs 110 --data dataset.yaml --weights yolov5n.pt python train.py --img 640 --batch 12 --epochs 110 --data dataset.yaml --weights yolov5s.pt python train.py --img 640 --batch 12 --epochs 110 --data dataset.yaml --weights yolov5m.pt python train.py --img 640 --batch 12 --epochs 110 --data dataset.yaml --weights yolov5l.pt --img: The pixel size at which the images to be trained will be reduced by the YOLOv5 model. Its default value is 640 × 640, and it was chosen here in this way. --batch: The number of data point packets to be used by the display card at a time while training the model. --epochs: The number of times all training data is shown to the trained network and the weights are updated while training the model. --data: The path to the .yaml file containing the general path and class information of the file containing the dataset. --weights: The location of the weight file containing the training coefficients to be used in training the model. By running these lines of code, the training process of the model was started. The program first checks the YOLOv5 files. The training process is carried out during the determined number of cycles (epoch). 2.2.4 Evaluation Indicators True Positive (TP) indicates the number of positive images that are correctly categorized as positive. True Negative (TN) indicates a specific sample number that the model correctly identified a negative sample as actually negative. False Positive (FP) details the number of samples that a negative sample was incorrectly identified as a positive sample by the algorithm. False Negatives (FN) indicates the number of samples that the algorithm incorrectly categorized a positive sample as negative. � Accuracy: This metric is used when the classification problem has a balanced class distribution (similar amount of data in each class). If the class distribution is unbalanced, the problem of capturing the class with a low number of classes may be encountered. Accuracy�=� TN�+�TP TP�+�FP�+�TN�+�FN (1) � Error Rate: It is the rate of frequency of incorrect classifications/predictions in the problem. Error Rate�=� FN�+�FP TP�+�FP�+�TN�+�FN or (1�- Accuracy) (2) � Precision: It is the success rate of positive class (1) predictions. It indicates how many of the predicted positive classes (classes predicted as 1) are actually positive. Precision�=� TP FP�+�TP (3) � Recall: It is the correct prediction rate of the positive class (1). It is the metric value that shows how many of the predicted positive classes have been predicted correctly. Recall�=� TP TP�+�FN (4) � F1-Score: It is the harmonic average of precision and recall values. It retains the effect of both Precision and Recall values. F1 Score�=� 2�× Precision Precision�+�Recall (5) � Mean Average Precision: This metric is the precision and recall product of detected bounding boxes. The MAP value scale varies between 0 and 1. The higher the value, the better the result. MAP is found by calculating jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 14 the average precision (AP) for each class separately and then calculating the average over the class. The result is accepted as true positive if the mAP value is above 0.5. mAP = 1 C + P(k)�R(k)T k=1 (6) 3. Research Results F1 Score, Precision and Recall value graphs were examined according to the error matrix metrics of YOLOv5 algorithms. F1 Score, precision, recall and loss function graphs are given in Figures 3, 4, 5 and 6, respectively. Figure 3. F1 Score graph Model 3: When the graph was examined, it was seen that the model showed a general increasing trend over time. The F1 score is the harmonic average of the precision and recall metrics and is an indicator of overall performance. The increasing trend indicated that the overall performance of Model 3 improved. Figure 4. Precision graph Model 3: The Precision graph indicated that the model generally exhibited a high precision score and this score showed a slight increasing trend over time. Besides, the graph indicated that most of the positive predictions of Model 3 were correct. 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 10 1 10 5 10 9 YOLOv5s YOLOv5n YOLOv5m YOLOv5l 0 0.5 1 1.5 2 2.5 3 3.5 4 1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 10 1 10 5 10 9 YOLOv5s YOLOv5n YOLOv5m YOLOv5l jas.ccsenet. It can be o over time. that the gr certain situ org observed that th It meant that raph still conta uations or whe 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 1 5 9 he model cons the developme ained some fluc en encountered (A) Size: 640 9 13 17 21 25 YOLOv Journal of A Figure stantly improve ent and adjustm ctuations indic d with certain d 0 × 640, Batch: 29 33 37 41 45 v5s YOL Agricultural Sci 15 5. Recall grap ed and was ab ment of the M cated that there data subsets. 12, Epoch: 110, 49 53 57 61 65 LOv5n YO ience ph le to make mo odel 3 proceed e was variabili Algorithm: YO 69 73 77 81 85 OLOv5m ore correct pos ded accurately ity in the mode OLOv5n 85 89 93 97 10 1 10 5 YOLOv5l Vol. 16, No. 1; itive classifica y. However, the el’s performan 10 5 10 9 2024 ations e fact nce in jas.ccsenet.org (B) Size: 640 (C) Size: 640 Journal of A 0 × 640, Batch: 0 × 640, Batch: 1 Agricultural Sci 16 12, Epoch: 110, 12, Epoch: 110, ience Algorithm: YO Algorithm: YOL OLOv5s LOv5m Vol. 16, No. 1; 2024 jas.ccsenet. It was see during the COMPAR The metric Table 1. C Model Model 3 Model 2 Model 1 Model 4 Model Model 3 Model 2 Model 1 Model 4 Model Model 3 Model 2 Model 1 Model 4 According -YOLOv5 the model result was -YOLOv5 and recall org en that Model training proce RISON OF MO c data of Mode omparison of m Metr 0.975 0.939 0.948 0.966 Trai 0.031 0.028 0.319 0.024 val/b 0.012 0.017 0.017 0.010 g to the result v m: Medium m ’s performanc found as 0.78 s: The Small m value of 0.961 (D) Size: 640 3 errors gener ess and its pred ODEL ALGOR el “3” and the model algorith rics/mAP_0.5 545 961 83 631 n/box_loss 1798 8442 967 4745 box_loss 2554 7055 7151 0914 values in Table model achieved ce as it averag 802. Precision model came se 154. Journal of A 0 × 640, Batch: Figure 6. L rally decreased dictions were c RITHMS: difference of o hms Difference (M 0.03584 0.02715 0.00914 Difference (M 0.003356 -0.000169 0.07053 Difference (M -0.004501 -0.004597 0.00164 e 1, d the highest s ges the model’ n and recall val econd with a m Agricultural Sci 17 12, Epoch: 110, Loss Function g d over time. T closer to the tru other models to Model 3) Me 0.7 0.6 0.6 0.8 Model 3) Tr 0.0 0.0 0.0 0.0 Model 3) va 0.0 0.0 0.0 0.0 score of mAP_ ’s sensitivity a lues were also mAP_0.5: 0.95 ience , Algorithm: YO graph This meant tha ue values. o these data ar etrics/mAP_0.5 78802 68131 64466 81199 rain/obj_loss 014735 016246 017904 012433 l/obj_loss 0086365 0098648 01639 0074921 _0.5:0.95, whi at various IoU high for this m 5 value of 0.68 OLOv5l at the model g re given in Tab 5:0.95) Diffe 0.10 0.14 -0.02 Diffe -0.00 -0.00 0.002 Diffe -0.00 -0.00 0.00 ich is a good o U thresholds. T model. 8131, precisio Vol. 16, No. 1; generally got b ble 1. erence (Model 3 671 336 2397 erence (Model 3 01511 03169 2302 erence (Model 3 012283 020025 1144 overall indicat The 0.5 mAP_ n value of 0.8 2024 better 3) 3) 3) tor of _0.95 2769 jas.ccsenet. -YOLOv5 of 0.96154 -YOLOv5 value of 0. However, our applic classificati many obje examined, was seen t 3.1 Trainin Training re org n: The Nano m 4. l: The Large m .92794 and rec these ranking cation, the ra ion losses in tr ect classes. Wh it was seen th that the model ng Result esult screensho Size: 64 110, Alg Size: 64 110, Alg model had a m model showed call value of 0. s may vary de anking will c raining and va hen the train/b hat the model with the least ots are shown 40 × 640, Batch gorithm: YOLO 40 × 640, Batch gorithm: YOLO Journal of A mAP_0.5:0.95 v the lowest pe .99081. epending on th change. Train/ alidation data, p box_loss, train with the least loss values in in Figure 7. h: 12, Epoch: Ov5n h: 12, Epoch: Ov5s Agricultural Sci 18 value of 0.644 erformance wit he specific use /cls_loss and play an import n/obj_loss, val/ t loss values in box_loss and ience 66, precision v th a mAP_0.5: e case. If recal val/cls_loss tant role in mo /box_loss, val n the training val_loss in the value of 0.870 :0.95 value of ll is prioritized parameters, w odels that requ /obj_loss valu set was “Mod e validation da Vol. 16, No. 1; 29 and recall v f 0.81199, prec d over precisio which express uire the detectio ues in Table 1 del 4”. Howev ata was “Mode 2024 value cision on in s the on of were ver, it l 3”. jas.ccsenet. 4. Discuss Deep learn productivi due to this used toma groups of sizes. As a the YOLO various mo VGG16, In disease typ close to 99 they used study, they (2023) car Phan et a experimen determined in their fru number of conducted (2022) use period for harvesting respectivel org Size: 64 110, Alg Size: 64 110, Alg Figure 7. Va sion ning is used es ty are achieve s productivity atoes, potatoes products. The a result of the Ov5 model. The odels of deep nception V3, V pes for eggpla 9%. Hu et al. pepper, eggpla y found that th rried out a cla al. (2023) used nts performed d the rate of fin uit detection s f cherry tomat d experiments u ed YOLOv2 a tomatoes. Liu g system study ly. According 40 × 640, Batch gorithm: YOLO 40 × 640, Batch gorithm: YOLO alidation Batch specially in cla ed in the proce contributes to s, eggplants an ey trained a da comparison o ey found the av learning in th VGG 19, Mob ant. According (2023) chose ant and tomato he average sen assification an d YOLOv5 an with 200 epo nding damage ystem studies toes on the br using YOLO a and YOLOv3 u et al (2022) u y and determin to these valu Journal of A h: 12, Epoch: Ov5m h: 12, Epoch: Ov5l h” prediction m assification an esses of identi o the developm nd peppers in ataset consistin f deep learnin verage precisi heir disease det bileNet, NasNe g to the test se the YOLOv5 o as materials nsitivity value w nd identificatio nd CNN deep ochs, batch 12 d tomatoes as using the dee ranch with 93 as a training m models in the used the DA-M ned the precis es, they confir Agricultural Sci 19 markings result nd discriminati ifying and har ment of select n their study o ng of a total of g models, they on of the YOL tection system etMobile and R et results, they deep learning for the detecti was 95% and on study for p p learning mo 28 and image 94% for YOL ep learning mo 3% accuracy w model in their st eir harvest pre Mask RCNN de sion, recall an rmed the suita ience ting from the tr ion in agricult rvesting agricu tive harvesting on disease det f 16580 image y emphasized LOv5 model as m for eggplant. ResNet50. The y emphasized t g model as the ion of seedling the detection s potatoes with Y odel to classif size of 224 LOv5m. Sa et a odel. Rahnemo with deep lear tudy of countin ediction system eep learning m nd F1-score va ability of the raining of the m tural systems. ultural product g systems. Hid tection with d es with 100 ep that they foun s 94.2%. Haqu . The deep lea ey created the that they achie e method in th g branch nodes speed was 0.0 YOLOv3 deep fy tomatoes. A × 224 for the al. (2023) used oonfar et al. (2 rning model. W ng mango frui m study based model in their au alues as 0.993 model. Abeyr Vol. 16, No. 1; models High accuracy ts. Increasing dayah et al. (2 deep learning pochs and 16 b nd the best resu ue et al. (2022) arning models training set fr eved a success heir study in w s. As a result o 19 sec. Wang p learning me As a result o e training set, d the R-CNN m 2023) predicte Wang et al. (2 t on trees. Lee d on the flow utomatic aspar , 0.971 and 0 rathna et al. (2 2024 y and yield 2022) on 4 batch ult in used were om 9 s rate which of the et al. thod. f the they model d the 2019) e et al. ering ragus .982, 2023) jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 20 examined which deep learning model was suitable for robotic harvesting systems. The compared models were YOLOv4, YOLOv5 and YOLOv7 deep learning models. They emphasized that the models to be used in such systems were YOLOv5 and YOLOv7. In the modeling study, the best model for eggplant was selected to be used in robotic harvesting systems. When the modeling results were compared with previous studies, it was seen that they showed parallelism according to the study criteria. In all studies, YOLOv5 was found to be the best model. Differences were determined in the sub -models. This difference was due to the structure of the products used, epoch, batch and image processing pixel values. It was determined that the modeling was suitable for robotic harvesting. 5. Conclusion The deep learning method has been an important tool in robotic harvesting applications of many products in agricultural automation. Robotic harvesting applications appear as an important method that can contribute to increasing agricultural productivity and reducing labor force. Studies on eggplant harvesting using deep learning methods have a significant potential to increase efficiency and optimize harvesting processes in the agricultural sector. For this reason, in this deep learning model study, it was determined which model gave the best results. In the study, it was determined that the YOLOv5m model was the most successful model to be used in robotic eggplant harvesting. All models were trained with 640 × 640 images. Metric values such as “metrics/precision”, “metrics/recall”, “metrics/mAP_0.5” and “metrics/mAP_0.5:0.95” of the models created with 12 Batch, 110 Epoch were examined. As a result of the comparisons, it was seen that the “YOLOv5 medium” model had higher metric values than the other models. The YOLOv5m model gave the highest score with F1 score of 85.66%, precision of 95.65%, recall of 96.15%, and mAP at 0.5:0.65 of 78.80%. In this study, the potential of deep learning, especially the YOLOv5m model, in automating the robotic eggplant harvesting process was revealed. According to the results, it was contributed to which model would be more efficient in robotic harvesting applications. References Abeyrathna, R. M. R. D., Nakaguchi, V. M., Minn, A., & Ahamed, T. (2023). Recognition and counting of apples in a dynamic state using a 3D camera and deep learning algorithms for robotic harvesting systems. Sensors, 23(8), 3810. Aji, W. S., & Hawari, K. (2019). A study of deep learning method opportunity on palm oil ffb (fresh fruit bunch) grading methods. Proceedings of the 2019 Ahmad Dahlan International Conference Series on Engineering and Science (ADICS-ES 2019). Akta�, H. (2022). Classification of pistachios based on outer shell color using deep learning. Ni�de Ömer Halisdemir University Journal of Engineering Science, 11(3), 461-466. Ampatzidis, Y. (2018). Applications of artificial intelligence for precision agriculture. Edis, 2018(6). Ampatzidis, Y., Bellis, L. D., & Luvisi, A. (2017). Ipathology: Robotic applications and management of plants and plant diseases. Sustainability, 9(6), 1010. Ayadi, R., El-Aziz, R. M. A., Taloba, A. I., Aljuaid, H., Hamed, N. O., & Khder, M. A. (2022). Deep learning– based soft sensors for improving the flexibility for automation of industry. Wireless Communications and Mobile Computing, 1-10. Carvalho, R. B. D., Guedes, E. B., & Figueiredo, C. M. S. (2022). Detecção inteligente de falhas em pavimentações asfálticas com redes neurais convolucionais regionais. Anais Do III Workshop Brasileiro De Cidades Inteligentes (WBCI 2022). Çetiner, H., & Çetiner, �. (2022). Classification of citrus diseases with convolutional neural network based deep learning model. Bitlis Eren University Journal of Sceince, 11(1), 352-368. Cicco, M. D., Potena, C., Grisetti, G., & Pretto, A. (2017). Automatic model based dataset generation for fast and accurate crop and weeds detection. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Dara, R., Fard, S. M. H., & Kaur, J. (2022). Recommendations for ethical and responsible use of artificial intelligence in digital agriculture. Frontiers in Artificial Intelligence, 5. E�iyok, D. (2012). K��l�k ve Yazl�k Sebze Yeti�tiricili�i. Meta Bas�m (p. 404). Bornova-�zmir. jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 21 Ferreira-Santos, P., Duca, A. B., Genisheva, Z., Silva, B., Biasio, F. D., Botelho, C. M., & Teixeira, J. A. (2021). Extracts from red eggplant: impact of ohmic heating and different extraction solvents on the chemical profile and bioactivity. Frontiers in Sustainable Food Systems, 5. Gürbüz, N., Ulu���k, S., Frary, A., Frary, A., & Do�anlar, S. (2018). Health benefits and bioactive compounds of eggplant. Food Chemistry, 268, 602-610. Haque, M. R., & Sohel, F. (2022). Deep network with score level fusion and �nference-based transfer learning to recognize leaf blight and fruit rot diseases of eggplant. Agriculture, 12(8), 1160. Hidayah, A., Radzi, S. A., Razak, N. A., Saad, W. H. M., Wong, Y. C., & Naja, A. (2022). Disease detection of solanaceous crops using deep learning for robot vision. Journal of Robotics and Control, 3(6), 790-799. Hu, J., Li, G., Mo, H., Lv, Y., Qian, T., Chen, M., & Lu, S. (2023). Crop node detection and internode length estimation using an improved YOLOv5 model. Agriculture, 13(2), 473. Jin, X., Yu, X. H., Wang, X. Y., Bai, Y. T., Su, T., & Kong, J. (2020). Deep learning predictor for sustainable precision agriculture based on internet of things system. Sustainability, 12(4), 1433. Kumar, S., Jailia, M., Varshney, S., Pathak, N., Urooj, S., & Elmunim, N. A. (2023). Robust vehicle detection based on improved you look only once. Computers, Materials & Continua, 74(2), 3561-3577. LeCun, Y., Bengio, Y., & Hinton, G. E. (2015). Deep learning. Nature, 521(7553), 436-444. Lee, U., Islam, M. P., Kochi, N., Tokuda, K., Nakano, Y., Naito, H., … Ahn, D.-H. (2022). An automated, clip-type, small internet of things camera-based tomato flower and fruit monitoring and harvest prediction system. Sensors, 22(7), 2456. Li, F., Chen, Y., Hu, M., Man-lin, L., & Wang, G. (2023). Helmet-wearing tracking detection based on strongsort. Sensors, 23(3), 1682. Liu, L., Liu, Y., Gao, X., & Zhang, X. (2022). An immersive human-robot interactive game framework based on deep learning for children’s concentration training. Healthcare, 10(9), 1779. Liu, X., Wang, D., Li, Y., Guan, X., & Qin, C. (2022). Detection of green asparagus using improved mask R-CNN for automatic harvesting. Sensors (Basel), 28, 22(23), 9270. Lundervold, A. S., & Lundervold, A. (2019). An overview of deep learning in medical imaging focusing on mri. Zeitschrift Für Medizinische Physik, 29(2), 102-127. Msogoya, T. J., Majubwa, R. O., & Maerere, A. (2014). Effects of harvesting stages on yield and nutritional quality of African eggplant (Solanum aethiopicum) fruits. Journal of Applied Biosciences, 78(1), 6590. Nasir, F. E., Tufail, M., Haris, M., Iqbal, J., Khan, S. G., & Khan, M. T. (2023). Precision agricultural robotic sprayer with real-time tobacco recognition and spraying system based on deep learning. Plos One, 18(3), e0283801. Nepal, U., & Eslamiat, H. (2022). Comparing YOLOv3, YOLOv4 and YOLOv5 for autonomous landing spot detection in faulty uavs. Sensors, 22(2), 464. Passam, H. C., Theodoropoulou, S., Karanissa, T., & Karapanos, I. C. (2010). Influence of harvest time and after-ripening on the seed quality of eggplant, Scientia Horticulturae, 125, 518-520. Phan, Q.-H., Nguyen, V.-T., Lien, C.-H., Duong, T.-P., Hou, M. T.-K., & Le, N.-B. (2023). Classification of tomato fruit using YOLOv5 and convolutional neural network models. Plants, 12(4), 790. Rahnemoonfar, M., & Sheppard, C. (2017). Deep count: Fruit counting based on deep simulated learning. Sensors, 17(4), 905. Reis, R. d. G. E., Guimarães, R. M., Vieira, A. R., Gonçalves, N. R., & Costa, V. H. (2012). Physiological quality of osmoprimed eggplant seeds. Ciência E Agrotecnologia, 36(5), 526-532. Ryo, M., Schiller, J., Stiller, S., Palacio, J. C. R., Mengsuwan, K., Safonova, A., …Wei, Y. (2022). Deep learning for sustainable agriculture needs ecology and human involvement. Journal of Sustainable Agriculture and Environment, 2(1), 40-44. Sa, I., Ge, Z., Dayoub, F., Upcroft, B., Perez, T., & McCool, C. (2016). DeepFruits: A fruit detection system using deep neural networks. Sensors, 16(8), 1222. jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 22 Santacatalina, J., Soriano, J., Cárcel, J. A., & García-Pérez, J. (2016). Influence of air velocity and temperature on ultrasonically assisted low temperature drying of eggplant. Food and Bioproducts Processing, 100, 282-291. Sao, A., & Mehta, N. (2010). Heterosis and inbreeding depression for fruit yield and its components in brinjal (Solanum melongena L.). Dirasat Agricultural Sciences, 37(1), 36-45. Shi, Z., Zhao, J., Shi, Y. S., Wang, Y. F., & Shunqi, M. (2022). Research on improving YOLOv5s algorithm for fabric defect detection. International Journal of Clothing Science and Technology, 35(1), 88-106. Sozzi, M., Cantalamessa, S., Cogato, A., Kayad, A., & Marinello, F. (2022). Automatic bunch detection in white grape varieties using YOLOv3, YOLOv4, and YOLOv5 deep learning algorithms. Agronomy, 12(2), 319. Tsai, H., Lu, N., Liu, K., Lin, C., & Wang, J. (2023). Cascading ab-YOLOv5 and pb-YOLOv5 for rib fracture detection in frontal and oblique chest x-ray images. IET Computer Vision, 17(7), 750-762. Umar, M. A., Sani, B. M., & Suleiman, U. (2022). An overview of machine and deep learning technologies application in agriculture: opportunities and challenges in Nigeria. SLU Journal of Science and Technology, 4(1&2), 84-96. Uzun, S., Demir, Y., & Cemek, B. (2000). Plastik seralarda uygulanan de�i�ik havaland�rma sistemi ve plastik tiplerinin baz� sebzelerin büyüme, geli�me ve verimine etkisi. O.M.Ü. Ziraat Fakültesi Dergisi, 15(1), 46-53. Wang, G., Yang, W., Liu, Y., Yang, X., Wang, Q., Yang, S., … Li, H. (2023). Potato malformation identification and classification based on improved YOLOv3 algorithm. Electronics, 12(21), 4461. Wang, Z., Walsh, K., & Koirala, A. (2019). Mango fruit load estimation using a video based mango YOLO—Kalman Filter—Hungarian algorithm method. Sensors, 19(12), 2742. Xu, Q., Hu, K., Zhang, H., Han, H., & Li, J. (2020). Organic vegetable cultivation reduces resource and environmental costs while increasing farmers’ income in the north china plain. Agronomy, 10(3), 361. Xue, J., Zheng, Y., Dong-Ye, C., Wang, P., & Yasir, M. (2022). Improved YOLOv5 network method for remote sensing image-based ground objects recognition. Soft Computing, 26(20), 10879-10889. Zhang, Q., Liu, Y., Gong, C., Chen, Y., & Yu, H. (2020). Applications of deep learning for dense scenes analysis in agriculture: A review. Sensors, 20(5), 1520. Zhang, Y., Cui, J., & Liu, M. (2022). Research on adversarial patch attack defense method for traffic sign detection. Communications in Computer and Information Science, 199-210 Zhao, J., Zhang, X., Yan, J., Qiu, X., Yao, X., Tian, Y., … Cao, W. (2021). A wheat spike detection method in uav images based on improved YOLOv5. Remote Sensing, 13(16), 3095. https://doi.org/10.3390/rs13163095 Authors Contributions The authors declare that they have contributed equally to the article. Funding This research received no external funding. Competing Interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Informed Consent Obtained. Ethics Approval The Publication Ethics Committee of the Canadian Center of Science and Education. The journal’s policies adhere to the Core Practices established by the Committee on Publication Ethics (COPE). Provenance and Peer Review Not commissioned; externally double-blind peer reviewed. jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 23 Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. Data Sharing Statement No additional data are available. Open Access This is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/4.0/). Copyrights Copyright for this article is retained by the author(s), with first publication rights granted to the journal. Journal of Agricultural Science; Vol. 16, No. 2; 2024 ISSN 1916-9752 E-ISSN 1916-9760 Published by Canadian Center of Science and Education 24 Canopy Structure Influence the Critical Period for Weed Removal of Three Cassava (Manihot esculenta Crantz) Varieties in Zambia Joseph Nzunda1,2,3, Nhamo Nhamo1,4, David Chikoye1, Kallunde Sibuga2 & Pheneas Ntawuruhunga1 1 International Institute of Tropical Agriculture (IITA), Southern Africa HUB, Lusaka, Zambia 2 Sokoine University of Agriculture (SUA), Morogoro, Tanzania 3 Tanzania Agricultural Research Institute (TARI), Mtwara, Tanzania 4 International Center for Biosaline Agriculture (ICBA), Dubai, United Arab Emirates Correspondence: Nhamo Nhamo, International Center for Biosaline Agriculture (ICBA), P.O. Box 14660, Al Ruwayyah 2, Academic City, Dubai, United Arab Emirates. Tel: 971-4-304-6300. Email: n.nhamo@biosaline.org.ae Received: November 5, 2023 Accepted: January 4, 2024 Online Published: January 15, 2024 doi:10.5539/jas.v16n2p24 URL: https://doi.org/10.5539/jas.v16n2p24 Abstract Cassava (Manihot esculenta Crantz) is an important crop for food, feed and income security. Cassava productivity is limited by poor weed management. Field trials were conducted in Zambia to determine the Critical Period for Weed Removal (CPWR) on 3 cassava varieties (Chila, Mweru and Nalumino), with contrasting canopy structure, using a split-plot design in randomized blocks. Nine weeding treatments, i.e., control, 21, 42, 63, 84, 105, 126, 147, 168 days after planting (DAP), were applied on two sets of weeding regimes. In one set, weeds were allowed to grow followed by a weed free period while in the second, plots were kept weed-free followed by a period of natural weed infestation at Kabangwe and Kaoma. Cassava varietal means were in the order Chila (10,199 kg ha-1) > Nalumino (9,047.6 kg ha-1) > Mweru (8,429 kg ha-1). Chila, a branching cassava variety, significantly out-yielded (P < 0.05) other varieties. Fresh cassava root yields were higher at Kabangwe (23,270 kg ha-1) compared to Kaoma (21,347 kg ha-1). The CPWR was determined to be 60 DAP (48-73 DAP), at both sites. Yield differences among weeding treatment ranged between 18% and 75%. The determined CPWR is a determinant of weed management strategy for branching cassava varieties. The branching canopy architecture smothered weeds and hence is considered an important cassava varietal attribute. The yields in the current study are doubled the regional yield average of 8000 kg ha-1 and four times the Zambian average of 5000 kg ha-1. Keywords: field management practices, weed interference, biotic stress factors, root and tuber crops, Southern Africa 1. Introduction Cassava (Manihot esculenta Crantz) is an important food, feed, energy and income security crop grown mainly for its starchy roots that can be processed into various products (Balagopalan, 2002; Scott, 2021). It is an important source of food calories for > 800 million people. Cassava provides double the amount of energy (250 kcal) compared to sorghum (114) or wheat (110) per capita per day (El-Sharkawy, 2012). It is a perennial crop belonging to the Euphorbiaceae family that has been cultivated for 50 years (Altieri et al., 2012; Godfray et al., 2010; Hirschmann & Vaughan, 1983). The crop is highly adapted for growth in the tropics between latitudes 30° north and south of the equator, at elevations between 0-2,000 m above sea level (asl). Cassava can be grown under semi-arid or humid eco-zones with annual precipitation ranging from 500 mm to > 2,000 mm, because of its wide adaptation (El-Sharkawy, 2012). Cassava production and utilization has been projected to increase by between 50-80% in Sub-Saharan Africa (SSA) due to its resilience to climate impacts (El-Sharkawy, 2006; FAO, 2013; Jarvis et al., 2012; Scott et al., 2000). However, the practice of growing cassava on marginal lands with minimal external inputs by smallholder farmers threatens future production in SSA. Production per unit area across most regions in SSA has remained extremely low, i.e., 8.3 t ha-1 (FAOSTAT, 2022) while the realist potential to produce yields exceeding 30 t ha-1 exist (Biratu et al., 2018). jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 25 In Zambia, cassava is important for food security, it ranks second after maize in daily consumption, is grown by over half a million smallholder rural households and is served in a range of dishes (Chitundu et al., 2006; ITC, 2010). The crop is mostly grown in the five provinces of Luapula, Northern, Western, Copperbelt, and Western provinces, where it is regarded as a staple food (Alene et al., 2013). The production of cassava has been growing and now exceed one million metric tonnes (MT) annually (Kabwe, 2014). Due to the increasing prominence, its production has expanded in recent years most of which (92%) is consumed locally (Haggblade et al., 2007) and contributes to social and economic development opportunities (Chipeta & Bokosi, 2013). However, cassava production faces various biotic challenges that constrain production on smallholder farms. Weed infestation is one of the major biotic stresses affecting yield, drudgery, and cost of production (Gianessi, 2013). Weed infestation in cassava production systems, which often receives less attention from researcher, is likely to cause large yield losses if not managed (Ekeleme et al., 2016). Weeds have been estimated to cost more than US$ 420 million and cause yield losses greater than 0.65 million tonnes in Sub-Saharan countries annually (Rodenburg & Johnson, 2009). Relying on manual weeding was found to increase drudgery among farmers and took between 173 and 376 labour hours per hectare (Ogwuike et al., 2014) hence the need for better weed management strategies. Appropriate weed management has been found to lead to high yield gains compared to other crop management practices, i.e., as high as 91.6% on crop yields (Nhamo et al., 2014) and reduce the total time spent on by at least 31% (Johnson et al., 2019). Valadatilde et al. (2013) reported that weed emergence, weed population density and abundance are the important weed factors influencing crops while crop factors such as time of planting, canopy, stage of crop development and competitive ability can increase chances of higher crop yields. Weed management practices have been found to influence the seed weed bank in the soil over time (Ekeleme et al., 2016). Therefore, management of both weed and crop factors is important in designing integrated weed management practices for crops such as cassava. In cassava, weed pressure results in intensive competition and when left to cover substantial portions of the growth cycle can cause root, stem and leaf yield losses, e.g., Khanthavong et al. (2016) reported a range of between 46-95% of produce losses. Reducing weed residence on cassava fields has been used as a practical method of minimizing the impact of weeds on the crop yields. Therefore, determination of the specific timing of applying weeding operations is important for successful cassava production. In Zambia, weed control methods commonly used by smallholder farmers are hand slashing, hand pulling and hoeing. These manual methods are inefficient because of drudgery, time consuming, expensive, and high labour demands during peak production periods (Islami et al., 2017; Maniyam et al., 2013). Alternative weed management practices such as chemical control (herbicides), cultural methods and integrated weed management practice have been tried elsewhere with variable results in cassava production systems (Ola et al., 2021a; Quee et al., 2016). Similarly, mechanical weed control tools that are gender neutral and labour saving have not attained the desired impact due to high initial cash outlay and hence low uptake by smallholder farmers (Johnson et al., 2019; Rodenburg et al., 2015). Access and knowledge of use of technologies remains a barrier for many smallholder farmers as evidenced by the number of accidents from chemical and mechanical weed control methods (Malambo et al., 2019). Nevertheless, appropriately timed weeding can contribute to drudgery reduction. To establish the specific time for weed removal, the peak period of weed competition must be determined (Shukor et al., 2009; Silva et al., 2013). Research on the determination of the appropriate timing of control, management and effective removal of weeds, dates back to around 6000BC and peaked post herbicide discovery in the 1940 and became intensive in the 1960s before the green revolution of the 1970s (Timmons, 1970). This work is the genesis of the integrated weed management (IPM) paradigm which became popular in the late 20th century (Young, 2012). The “critical period of weed control” is a more commonly used phrase in literature and has been aligned with the ground-breaking work on refining weed management timing in many crops. Critical period of weed control has several definitions including (a) the period during which weed control is necessary to avoid yield penalty (Nazarko et al., 2005), (b) the period of time between the minimum time point weed-free (MTPWF) and the maximum time point under weed-infestation (MTPWI). Where MTPWF, refers to the time after which the crop must be kept weed-free to prevent crop yield loss from the competition while MTPWI is the time up to which the crop must be kept weed-free from the beginning to avoid crop yield loss from emerging weeds thereafter (Singh et al., 1996). Further definitions of the critical period of weed control include (c) an interval in the life cycle of the crop when it must be kept weed-free to prevent yield loss (Van Acker et al., 1993), and (d) the shortest time spent in the crop growth cycle when weeding will result in highest economic returns (Hasanuzzaman, 2015). Based on the current understanding of integrated weed management, we define the critical period of weed removal in cassava based on the time of equal interference determined from weed-free and the weed-infested best fit polynomial jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 26 equations from cassava root, stem and leaf yields measured at equal time intervals during the growth cycle from planting through to harvesting. Thus, CPWR is the threshold time interval within which the application of weeding will effectively reduce crop-weed competition to a level that supports viable crop production at no yield penalty. We use the terms critical period of weed removal (CPWR) in place of the critical period of weed control (CPCR) in this publication. The knowledge of the CPWR is critical in the determination of appropriate timing of manual, mechanical or chemical weed management for economic and environmental sustainability. Cassava is susceptible to weed interference during the early growth stages, i.e., before full canopy development and this time period, depending on the variety and climatic conditions, ranges between 8 to 16 weeks after planting. Slow canopy development entails that the crop requires a prolonged period for the crop aerial parts to effectively shade the area under the crop thereby making early smothering of weed and hence weed suppression a challenge (Gianessi, 2013). Weed proliferation during early growth stages of cassava that encompass canopy development and early tuberization impact both yield and yield components (Onochie, 1975). Cassava variety, canopy architecture and branching behavior, and phenology are important determinants of the CPWR. Thus, determination of CPWR is a key component of integrated weed management for cassava production and it is useful for making decisions on the need and timing of weed control (Silva et al., 2013). The CPWR helps to make decisions on the necessity and selection of best time for weed control (Knezevic et al., 2002). The main objective of this study was to determine the critical period for weed removal (CPWR) in three improved cassava varieties, namely Mweru, Nalumino and Chila that are commonly grown in cassava growing areas of Kabangwe and Kaoma districts in Zambia. The research tested the hypotheses that; CPWR is longer in branching varieties than non-branching varieties due to early canopy development and more weed smoothing from shading effects, and fresh root yield and yield components respond positively to appropriately timed weed removal. 2. Materials and Methods 2.1 Description of the Study Sites 3.1.1 Location The study was conducted between 2016 and 2018 at two locations: (1) at IITA-Zambia research farm at Kabangwe (S15°.3 , E 28°3 ) in Chongwe district and (2) at ZARI-Longe experimental farm (S14°8 , E 24°.9 ) in Kaoma district in Western Province, Zambia. The altitude for the two site was not significantly different (about 16.6 m) with Kabangwe at 1188.6 m.a.s.l and Longe at 1163.3 m.a.s.l. 2.1.2 Soils To characterise the sites, soil samples were collected at the onset of the experiment and analysed in the Zambia Agricultural Research Institute (ZARI) central laboratories at Chilanga in Lusaka. Results of the chemical analysis showed that soils from the sites at Kaoma were sandy clay, with pH 5.4 (CaCl2), organic carbon 0.4% (Walkley Black, Anderson & Ingram, 1993), while those at Kabangwe were sandy loam with pH 5.6 and organic carbon 0.4%. These soils were classified as medium to slightly acidic. Sandy soils are dominant in Zambia and constitute about 70% of agricultural land (Yerokun, 2008). 2.1.3 Rainfall The total rainfall amounts received during the 2016/2017 cropping season at Kabangwe research station and Kaoma experimental site were 906.2 mm and 1221.9 mm respectively. Both sites (Kabangwe and Kaoma) received most rainfall in January compared to other months (316.6 mm and 366.2 mm) respectively. Both sites, Kaoma and Kabangwe, had adequate rainfall for growth of Cassava. The long-term mean monthly rainfall, maximum and minimum temperature for the study sites is presented in Figure 1. jas.ccsenet. Figure Both Kaba November moisture r shifts in th through th recurring p located in the amoun The rainfa production minimum developme 2.1.4 Temp The minim between 1 was record the minim cassava gr Optimal p and 35 °C minimal a summarize crop can g field. 2.2 Experi The exper main plot total of ni weed free Land was used for p into the so org e 1. Mean mon angwe and Ka r and finishes a rich airmasses he north-south he Mozambiqu phenomenon a Southern Afric nt of rainfall re all distribution n. Earlier repo annual rainfa ent and above peratures mum and max 8.7 °C and 26 ded in October mum temperatu rowing except hotosynthetic C for cool clim at temperature ed the effect of grow, 25-29 °C imental Design riment was laid factor and the ne (9) weedin and weed infe ploughed and planting the tri oil at a slanted nthly rainfall (m aoma sites exp around April. and closely re h direction. Ea ue channel also and is often as ca (Davis-Redd ceipts and tem n during the ca orts have show all of about 50 ground bioma ximum temper 6.4 °C at Kaom r. The mean m ure (10 °C) wa t during the M rates and cass mate and betwe es below 15 f temperature o C optimal cassa n d out using a e weeding man ng treatments w ested managem d harrowed usi ials. Planting w position while Journal of A mm), for the pe Kabangwe a perience a unim The timing an elated to the m sterly winds th o influence ra ssociated with dy & Vincent, mperature varia assava growth wn successful 00 mm distrib ass growth (El- rature recorded ma with the lo maximum temp as observed in May, June and sava storage ro een 30-40 °C °C (DPP, 201 on cassava gro ava growth, an Split-plot desi nagement as th were studied. ment systems. T ing a tractor in was conducted e keeping the r Agricultural Sci 27 eriod between and Kaoma in Z modal tropical nd intensity of movement of t hat are influen infall events. A extreme even 2017). The top ations (Libanda period, soils cassava produ buted across th -Sharkawy et a d in Kabangw owest average prature (35 °C) May. Both sit July period w oots bulking oc for hot climat 10; El-Sharka owth as follows nd 30-40 °C hi ign in random he subplot fac The varieties The experimen n November 2 d by burring a est of the cutti ience 2001 and 202 Zambia rainfall patter f the tropical ra the intertropica nced by the wa An El Nino S nts such as dro pography of th a et al., 2020). and temperatu uction under s he crop growi al., 1992). we ranged betw s recorded in ) in Kabangwe te experience s where night tem ccur when temp tic regions, wh awy et al., 19 s: < 15 °C plan ighest photosy mized complete ctor. Three cas and weeding ntal units were 2016. Fresh ca a minimum of ing above the g 0 for the two r rn which starts ains are influe al convergence arm Indian Oc Southern Oscil oughts and flo he specific sites ures were favo andy loamy o ing period to ween 19.7 °C May and June e was observe suitable averag mpertaures dro mperatures aver hile overall gro 92). Work of nt growth is inh ynthesis rates a e blocks with t sava varieties treatments we replicated thre assava cuttings f three nodes o ground. Vol. 16, No. 1; research sites a s invariably ar nced by the C e zone (ITCZ) cean currency lation (ENSO) ods in this cou s also contribut ourable for cas or loamy sands support both and 27.4 °C e while the hig ed in October w ge temperature op to below 12 age between 2 owth is reduc f Alves (2002) hibited, > 17 °C are attainable i the varieties a were used wh ere studied un ee times. s 30 cm long on cassava cut 2024 at ound ongo ) that flow ) is a untry tes to ssava s and tuber C and ghest while es for 2 °C. 25 °C ed to ) has C the n the as the hile a der a were ttings jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 28 3.2.1 Main Plot Factor Three (3) varieties, Chila, Mweru and Nalumino, constituted the main plot factor in this experiment. Plant stature varies among the varieties and the branching characteristic affects the impact of cassava canopy on weeds and weed development in the field. The varieties also have different adaptation with some varieties having a wide adaptation to environmental growth conditions. Table 1 summarizes the characteristics considered in selecting the varieties for use in the study. Table 1. Characteristics of three cassava varieties used in the critical period of weed removal experiment at Kabangwe and Kaoma from 2016 to 2018 Variety Taste Architecture Year released Mweru Sweet Upright architecture 2000 Nalumino Bitter Branching architecture 1993 Chila Bitter Semi-branching architecture 2000 Source : Chitundu et al. (2006); Chiona et al. (2016). 2.2.2 Sub-plot Factor A total of nine (9) treatments constituted the sub-plot factor in this study. The weeding treatments were executed at 21, 42, 63, 84, 105, 126, 147, 168 days after planting (DAP) and a control treatment was made up of weed free or weed infested plots that were kept throughout a period of 168 days. Growth parameters were measured at the same time intervals and at crop harvest. Weeds were allowed to interact (compete) with the crop for a period of about 6 months and this period constituted the study period when weed management treatments were applied. Earlier studies had shown that weed management during the initial vegetative and tuberization periods largely impacts the total marketable fresh root yield of cassava and its yield components. Table 2 shows the two sets of nine (9) weeding treatments used in the study under the weed free and weed infested regimes. Table 2. Treatment description for the weed free plots and weed infested plots in a study to determine the critical period of weed removal at Kabangwe and Kaoma in Zambia # Set one (1) Weed Free Plots (WF) Treatment The plots were kept free from weeds for the first: 1 21 DAP 21 Days After Planting and then left with weeds (weed infested) until harvest 2 42 DAP 42 Days After Planting and then left with weeds (weed infested) until harvest 3 63 DAP 63 Days After Planting and then left with weeds (weed infested) until harvest 4 84 DAP 84 Days After Planting and then left with weeds (weed infested) until harvest 5 105 DAP 105 Days After Planting and then left with weeds (weed infested) until harvest 6 126 DAP 126 Days After Planting and then left with weeds (weed infested) until harvest 7 147 DAP 147 Days After Planting and then left with weeds (weed infested) until harvest 8 168 DAP 168 Days After Planting and then left with weeds (weed infested) until harvest 9 Control WF Control Free from weeds all the time (Weed Free) until 168 DAP # Set two (2) Weedy Infested plots (WI) Treatment Plots were infested with weeds for the first: 1 21 DAP 21 Days after planting and then weed free until 168 DAP 2 42 DAP 42 Days after planting and then weed free until 168 DAP 3 63 DAP 63 Days after planting and then weed free until 168 DAP 4 84 DAP 84 Days after planting and then weed free until 168 DAP 5 105 DAP 105 Days after planting and then weed free until 168 DAP 6 126 DAP 126 Days after planting and then weed free until 168 DAP 7 147 DAP 147 Days after planting and then weed free until 168 DAP 8 168 DAP 168 Days after planting and then with weeds until the time of harvesting 9 Control WI Control (Weedy Infested) infested with weeds until harvest Note. T = Treatment. WF = Weed free plots, WI = Weed infested plots. jas.ccsenet.org Journal of Agricultural Science Vol. 16, No. 1; 2024 29 2.2.3 Plot Management Cultural plot management practices were applied from planting through to harvesting to ensure uniformity and minimal effects of pests and diseases. A uniform plant establishment was maintained by replacing the non-sprouting cassava cutting with in the first 21 days and eliminating varietal mixtures. Diseased plants were rouged out to avoid the spread. The positive control treatment which constituted weed-free plots throughout the 168 days while the negative control had weeds for the same period were carefully managed as they were important in the determination of the CPWR. 3.3 Variables and Data Collection Cassava yield and yield parameters were recorded at 3, 6 and 24 months during the experiment at Kabangwe and Kaoma. Cassava yield variables collected from aerial plant portions were leaf weight and stem weight; variables for soil part were root girth, root length and fresh root weight. The cassava storage root weights were the main factor which were used to determine critical period for weed removal. Prior to sampling, the plant stands were counted for each plot. During the final crop harvest at 24 months, yield data was collected from net plots sizes of 16 m2 (4 × 4 rows). Sampling proceeded by removing the whole plant and separate measurements taken for leaves, stems, and cassava roots. The roots were cleaned of soil before weighing them on a digital portable scale. Fresh subsamples were then collected for further analysis in the laboratory. 2.3.1 Root Weight Cassava root tubers were harvested by removing the edible below-ground structure from rest of the plant root systems. To maintain