CropGuard: An AI-Assisted Crop Health Detection and Smart Farming Rover Prototype

Early detection of foliar crop diseases is critical to safeguarding agricultural yields and minimizing economic losses, particularly for smallholder farms that lack on-demand access to professional plant pathologists. CropGuard is an autonomous agricultural rover platform combining multi-angle optical sensing, host-based edge computing, and localized mist treatment. In this paper, we document and rigorously evaluate the computer-vision core of the prototype: a lightweight YOLO11n object detector fine-tuned from COCO-pretrained weights across 16 crop-health classes (healthy foliage plus 15 fungal, bacterial, viral, and pest-induced conditions). The experimental dataset comprises 2,573 multi-crop leaf images (8,931 annotated bounding boxes) partitioned into a 2,344-image training split (8,476 instances) and a 229-image validation split (455 instances). On the validation split, the best-performing checkpoint (epoch 21 of 22 completed) achieves a mean precision of 0.5427, mean recall of 0.5778, [email protected] of 0.6082, and [email protected]:0.95 of 0.4624 across the 15 evaluable classes. One class (TWO_SPOTTED_SPIDER_MITES) contains zero validation instances and cannot be evaluated. On an Intel Core i7-14700HX CPU, the standalone YOLO11n forward pass requires 32.72 ms per frame (30.56 FPS), while an end-to-end processing pipeline incorporating CLAHE contrast enhancement and bounding-box drawing requires 107.23 ms (9.33 FPS). Physical benchtop integration of the 1:8 four-wheel-drive rover chassis, quad ESP32-CAM optical array, dual-relay ultrasonic mist actuation, and bilingual operator dashboard demonstrates responsive closed-loop plant treatment.