Optimizing Mask R-CNN for Enhanced Quinoa Panicle Detection and Segmentation in Precision Agriculture

Manal EL AKROUCHI^1,2*Manal Mhada¹Dachena Romain Gracia¹Malcolm John Hawkesford³Bruno Gérard¹

• ¹College of Agriculture and Environmental Sciences, University Mohammed VI Polytechnic, Ben Guerir, Morocco
• ²School of Collective Intelligence, University Mohammed VI Polytechnic, Rabat, Morocco
• ³Department of Sustainable Soils and Crops, Rothamsted Research, Harpenden, United Kingdom

The burgeoning field of precision agriculture has required the development of advanced methods for crop yield estimation. Quinoa, often labeled an underutilized crop, has been under-investigated despite its dietary benefits and cultivation potential in harsh environments. Our research presents a novel approach for quinoa panicle detection and counting using instance segmentation through Mask Region Convolutional Neural Network (Mask R-CNN). The use of instance segmentation in this context is a novelty, designed to detect and differentiate individual quinoa panicles, allowing for more precise yield estimation. To our knowledge, this is the first attempt to elucidate the role of deep learning in improving quinoa yield prediction. In this study, we propose an improved version of Mask R-CNN based on EfficientNet b7 and Mish function activation. A comprehensive comparative analysis of backbones revealed that our proposed methodology performed well in detecting and counting panicles. This study underscores the potential of leveraging advanced deep-learning techniques for automated and precise yield estimation in crops like quinoa. The insights from benchmarking various Mask R-CNN backbones will guide future research, significantly contributing to the underexplored area of AI-driven quinoa yield prediction.