APPLICATION OF DEEP LEARNING ALGORITHM FOR CROP PEST AND DISEASE DETECTION WITH OFFLINE MOBILE APPLICATION DEPLOYMENT

Crop pests and diseases cause major yield losses and are difficult to diagnose quickly in the field. Leveraging widespread smartphone cameras and advances in convolutional neural networks (CNNs), this study develops a multi-crop, offline mobile system that delivers accurate, on-device diagnosis with actionable guidance for smallholder settings. Images from public repositories (PlantVillage, Mendeleev) and the Cashew–Cassava–Maize–Tomato (CCMT) dataset were consolidated and split 70/15/15 into training/validation/testing. A lightweight router CNN first categorizes each input as maize, potato, tomato, or other; crop-specific specialist CNNs then provide fine-grained pest/disease classification. Trained models were exported to TorchScript and embedded in an Android application that performs fully on-device inference and displays Integrated Pest Management (IPM) actions (symptoms, first-aid, cultural/biological/chemical options). The router achieved 96.95% validation accuracy (N = 11,698) and 96.76% test accuracy (N = 11,689). With a routing threshold τ = 0.88, the system recorded a 42.90% route rate, 98.52% precision when routing, 0.63% overall mis-route rate, and 57.11% abstain rate. Per-class routing coverage on test data was maize 88.88%, potato 85.93%, tomato 92.34%, and other 0.64%, reflecting a conservative policy that prioritizes safety over incorrect forwarding. End-to-end inference ran fully offline with sub-second to ~1.00 s latency on representative Android devices. Coupling predictions with a local IPM-oriented knowledge base converts recognition into practical field guidance, indicating the feasibility of dependable, low-connectivity crop diagnostics on commodity smartphones.