Edge-Ready KNN for Orange Disease Diagnosis: A Sustainable, Low-Compute Pipeline with Hybrid Handcrafted Features

This study proposes an edge-ready and sustainability-driven pipeline for orange disease diagnosis that fuses hybrid handcrafted features (HOG-LBP-HSV) with a K-Nearest Neighbors (KNN) classifier. Designed for deployment on resource-limited devices, the method achieves 82.82% (95% CI: [81.5%, 84.1%]) accuracy and a macro-F1 score of 0.816 (95% CI: [0.805, 0.827]) on a three-class dataset (healthy, citrus canker, melanose). Class-wise recall is particularly strong for healthy (95.9%) and melanose (97.2%), while error analysis highlights under-detection in citrus canker due to overlapping lesion features. Importantly, operational benchmarks confirm feasibility for on-device applications, with ~16 ms/image inference time and ~15.9% CPU utilization. Compared with baseline models such as SVM and Random Forest, KNN offers comparable predictive power at significantly lower computational cost. Beyond accuracy, the work emphasizes deployment-centric evaluation, proposing misclassification-aware strategies to improve canker recall. By providing a low-compute, real-time diagnostic tool, this work contributes to the UN's Sustainable Development Goals (SDGs), particularly SDG 2 (Zero Hunger) by enabling early detection to boost yields by up to 20-30% in smallholder contexts (as per FAO estimates)and SDG 12 (Responsible Consumption and Production), by promoting resilient and sustainable agricultural practices for smallholder farmers.