AI-Empowered UAV-IoT System for Smart Agriculture

Accurate soil monitoring in rural areas increasingly relies on Unmanned Aerial Vehicle (UAV)-assisted remote sensing, where data collection is sparse and unreliable, and learning must be performed under communication, energy, and connectivity constraints. This work proposes a lightweight Distributed Multi-sensor Transformer (DMT-Former) cooperation framework built on Multi-Head Split Learning (MHSL), in which each worker UAV learns a compact representation from local multi-station soil-sensor sequences, while a master UAV fuses these representations to capture cross-sensor and cross-UAV interactions under communication constraints. To handle incomplete soil measurements, low-rank recovery is applied during data preparation and uncertainty-aware training is used to improve robustness under variable soil dynamics. Experiments on the TERENO and REMEDHUS datasets show that the proposed MHSL-DMT-Former achieves a cumulative in-situ soil-moisture RMSE of 0.034 m3/m3 over a 24-step prediction horizon on TERENO. The proposed model obtains Robustness Index values of 0.75 for the TERENO dataset and 0.81 for the REMEDHUS dataset, improving robustness by about 5 points compared to our previous CNN–LSTM-based MHSL design, while requiring about 3.01 Wh of onboard computational energy and 176.39 MB of worker–master communication overhead per retraining cycle. These results demonstrate favorable accuracy–robustness–energy–communication trade-offs for heterogeneous UAV-assisted agricultural monitoring.