An IoT-Integrated Deep Learning Framework for Weed Segmentation and Yield-Impact Estimation Using Mask R-CNN with a ResNet-50 Backbone and Regression Analysis

Abstract

Weed infestation is among the most serious biotic stresses of cereal and horticultural production and the economic losses caused by weed infestation are exacerbated by the manual, calendar-based weed management practices which are still widely employed on smallholder farms. This research suggests and tests an Internet-of-Things (IoT) enabled deep-learning approach that combines pixel-based weed segmentation and a regression-based yield-impact model to facilitate site-specific weed management. The field imagery included 5 400 tiles (from 1 500 georeferenced RGB captures) obtained in 5 experimental plots in 2 growing seasons using unmanned aerial vehicles and edge connected field cameras. The trained Mask R-CNN detector with a ResNet-50 + Feature-Pyramid-Network (FPn) backbone was compared with U-Net, DeepLabV3+, YOLOv5-seg, YOLOv8-seg, and Mask R-CNN with a ResNet101-DC5 backbone via a five-fold cross validation protocol. The proposed model achieved a mean Average Precision at IoU 0.50 (mAP@0.5) of 93.4 % (95 % CI: 92.6-94.1 %), an Intersection-over-Union of 91.7 %, precision of 94.2 % and recall of 92.1 %, which were superior to all baselines with the same training set under the same setting. The outputs from the segmentation were transformed to two field metrics that are ecologically relevant — weed density (plants · m⁻²) and coverage ratio — which were then used as predictors in an ordinary-least-squares regression of measured grain yield. The fitted model was quite significant (R² = 0.89, adjusted R² = 0.87, p < 0.001, RMSE = 218 kg ha⁻¹) and was similar to the classical hyperbolic loss-yield law in its linear regime. If used as a prescription map generator on the edge gateway, the framework achieved a 41.6 % reduction in the volume of herbicides applied as compared to blanket spraying without compromising the projected yield by more than 2 % from the control plot. The results show that combining high fidelity instance segmentation with parsimonious statistical yield models is an affordable path towards decision-ready, real-time precision-agriculture systems.