Detectron2-enhanced mask R-CNN for precise instance segmentation of rice blast disease in Tanzania: Supporting timely intervention and data-driven severity assessment

Rice blast, caused by Magnaporthe oryzae, poses a significant threat to rice production in Tanzania and across Africa, affecting food security and farmers' livelihoods. Traditional inspection methods are slow and often overlook early symptoms, leading to delayed responses. Although progress has been made with deep learning diagnostics, many approaches still depend on whole-image classification or broad bounding boxes, lacking the pixel-level detail needed to assess infection severity. This study introduces a Mask R-CNN instance segmentation model developed within the Detectron2 framework to accurately detect and segment blast lesions (BL), blast-infected leaves (BIL), and healthy leaves (HL) at the pixel level. In addition to detection, the model quantifies the lesion severity by computing the proportion of infected leaf area, supporting informed evaluation and improved disease management decisions. Built on a ResNet-50 backbone with a Feature Pyramid Network (FPN), it achieved a mean average precision (mAP) of 89.4 %, with an AP50 of 94.6 % and an AP75 of 90.5 %. The model exhibited consistent performance across object scales, achieving an AP of 81.31 % for small objects and 86.06 % for large objects. Furthermore, testing on unseen images (images not used in the training process) demonstrated strong generalization, with detection confidence above 99 % and accurate masks that provide reliable severity scores. By enabling pixel-level severity assessment without expensive sensors or UAVs, this study offers a practical and affordable solution for disease monitoring in resource-constrained farming communities. It equips Tanzanian smallholder farmers with timely, accessible tools for effective blast detection and data-driven decision making.