Precision pesticide application in solanaceous crops using transfer learning explainable artificial intelligence and drone based spatial zone mapping

Blanket pesticide spraying remains the default response to foliar disease in much of smallholder and mid-scale agriculture, despite three decades of precision-agriculture research arguing that inputs should be applied only where they are needed. This paper reports an case study of a drone-compatible precision-spraying pipeline built around three pepper, potato, and tomato the three Solanaceous crops for which labelled leaf imagery was available with 15 classes out of which 12 disease categories and 3 healthy classes. A MobileNetV2-0.75 classifier is fine-tuned with a two-phase transfer-learning schedule and evaluated over five independently seeded runs, giving a mean validation accuracy of 67.8% ± 2.0% and weighted F1 of 67.5% ± 2.9% (N = 5). A single, more thoroughly profiled run reached $$82%$$ accuracy and 0.82 weighted F1 on its own held-out split. Under an identical training protocol, MobileNetV2-0.75 reached $$86.2%$$ accuracy and ResNet50V2 reached $$79.5%$$ . Grad-CAM explanations are evaluated both qualitatively and, for the first time in this line of work, quantitatively via the deletion/insertion faithfulness protocol with deletion AUC 0.26 ± 0.15 , insertion AUC 0.47 ± 0.18 , $$N=8$$ ). A dose-weighted tile-grid spray planner converts per-tile disease confidence into drone waypoints; a ground-sample-distance heuristic applied to a typical 20 m survey altitude recommends a 32× 32 tiling, sixteen times finer than the 8× 8 default used for demonstration purposes, which it discuss as concrete evidence. Synthetic noise, blur, and brightness/contrast perturbations show the classifier degrading sharply under Gaussian noise and blur but remaining comparatively stable under illumination change. paper identify the dataset’s restriction to three crop species and laboratory imagery as the central limitation bounding any claim of field readiness.