Real-time crop health monitoring in precision agriculture using Geospatial data and deep learning

Abstract

Precision agriculture increasingly relies on advanced technologies to monitor crop health and optimize farming operations. This thesis presents a real-time crop monitoring system that leverages UAV-acquired multispectral imagery, deep learning, and a scalable data streaming infrastructure. High-resolution drone images are processed to generate vegetation indices such as NDVI, which serve as input for U-Net models. These models perform semantic segmentation and classification to detect various crop stress types, including nutrient deficiency, drydown, planter skips, water stress, and weed clusters. To ensure real-time processing and system scalability, the architecture integrates Apache Kafka for streaming UAV imagery and prediction outputs, supporting parallel operation across multiple producers and consumers. A React-based dashboard displays live predictions and metadata, enabling field operators to make timely and informed decisions. The system was validated using the Agriculture-Vision 2021 dataset and tested under simulated UAV deployment. The best segmentation performance was achieved using a U-Net model with RGB+NIR input at 512×512 resolution, reaching a Dice score of 0.70 and classification accuracy of 95%. In contrast, a NDVI-based model offered faster inference (3–5s) with slightly lower accuracy, making it suitable for resource-constrained environments. Kafka demonstrated low-latency image transmission (<2.5s) even with up to 15 parallel producers.