Abstract
Quinoa is a resilient, nutrient-rich crop with strong potential for cultivation in marginal environments, yet it remains underutilized and under-researched, particularly in the context of automated yield estimation. In this study, we introduce a novel deep learning approach for quinoa panicle detection and counting using instance segmentation via Mask R-CNN, enhanced with an EfficientNet-B7 backbone and Mish activation function. We conducted a comparative analysis of various backbone architectures, and our improved model demonstrated superior performance in accurately detecting and segmenting individual panicles. This instance-level detection enables more precise yield estimation and offers a significant advancement over traditional methods. To the best of our knowledge, this is the first application of instance segmentation for quinoa panicle analysis, highlighting the potential of advanced deep learning techniques in agricultural monitoring and contributing valuable benchmarks for future AI-driven research in quinoa cultivation.