INTEGRATED UAV-IOT SENSING AND MACHINE LEARNING FOR REAL-TIME CROP WATER STRESS DETECTION AND PRECISION IRRIGATION

Abstract

Water scarcity and outdated agricultural scheduling threaten global food security, accounting for nearly 70% of freshwater withdrawals worldwide. While Agriculture 4.0 paradigms show promise through the union of Unmanned Aerial Vehicles (UAVs) and Internet of Things (IoT) sensor networks, practical field deployment lacks unified, statistically verified pipelines. This study develops a reproducible framework that fuses high-spatial resolution UAV multispectral/thermal imagery with high-temporal-resolution IoT telemetry across the soil-plant atmosphere continuum. Over two growing seasons, we tested this architecture using a randomized complete block design (RCBD) spanning 12 distinct field plots (four crop types, three replicates). Distributed LoRaWAN nodes captured real-time soil moisture and microclimate variables at 5-minute intervals, matching RTK-georeferenced imagery gathered during UAV flights. We synchronized these spatiotemporal layers through ordinary Kriging and raster alignment before running feature engineering steps. To predict the Crop Water Stress Index (CWSI), we trained Random Forest, XGBoost, and LightGBM models, verifying our outputs against physical leaf water potential and stomatal conductance readings. Rigorous verification was handled via time-series cross-validation and Bayesian hyperparameter tuning. Our field results show the system can detect crop stress 5 to 7 days before visual signs appear. Furthermore, the predictive pipeline cleared an accuracy threshold of (R^2 = 0.924\), slashing irrigation water use by 28.7% without hurting crop yields. This framework offers a validated template for modern agricultural decision support systems.