AI Drone for Crop Disease Detection Using Deep Learning

Agriculture in India, particularly in states like Maharashtra, faces constant threats from plant diseases that can wipe out 20–40% of crops annually, leading to severe income losses for small and marginal farmers who often lack access to expert agronomists or expensive monitoring tools. Conventional methods involve manual field scouting — walking row by row, examining leaves for spots, wilting, or discoloration — which is extremely time-consuming, physically demanding, error-prone (especially for subtle early symptoms), and impractical for farms spanning even a few acres. To address this real-world problem affordably, our team developed Agro Drone AI, an end-to-end intelligent crop monitoring framework using low-cost drone technology combined with state-of-the-art AI. We specifically selected the Dynalog DR-DG600C GPS drone (a budget-friendly model priced around ₹9,000–₹12,000 depending on variants and sellers like Flipkart/Amazon/ZoneAlpha, weighing under 250g so no DGCA registration is required for educational use) for image acquisition. This drone features a claimed 4K (often interpolated/upscaled from 1080p native) camera with 120° wide-angle lens, adjustable tilt (up to 90°), 5GHz WiFi FPV for live view, GPS for stable hovering and return-to-home, follow-me/orbit/waypoint modes, and flight times of 12–20 minutes per battery (longer with dual-battery Pro versions). Captured aerial images — which frequently suffer from motion blur, low contrast due to altitude/sun angle, compression artifacts, or wind-induced shake on a lightweight consumer drone — are first enhanced using Re-al-ESRGAN (a powerful GAN-based super-resolution model that realistically reconstructs fine details without introducing unnatural artifacts). The sharpened images are then fed into the DeiT-small (Data-efficient Image Transformer) model, fine-tuned on the PlantVillage dataset, for multi-class disease classification (healthy vs. specific diseases like bacterial spot, early blight, leaf mold, etc.) with confidence scores and basic severity esti-mation. Our experiments (using PlantVillage for training/benchmarking + some self-captured/simulated aerial views from the Dynalog drone) demonstrated clear improvements: super-resolution boosted visibility of subtle symptoms (e.g., tiny vein yellowing or powdery mildew specks), and DeiT's global attention mechanism han-dled aerial perspectives better than local-feature-focused CNNs. This low-budget pipeline offers a practical path for early disease detection in precision agriculture, reducing manual labor, minimizing broad-spectrum pes-ticide use, and empowering farmers/cooperatives in resource-constrained areas like Vidarbha.