农业人工智能 变量施肥

综述2020-2026年文献，阐述人工智能如何支撑精准农业实现资源优化。

Estimating swath width is essential for improving fertilizer distribution uniformity in precision fertilizer application. This study presents a rapid vision-based framework for swath-width prediction using a hybrid CNN–Transformer architecture trained on synthetic video sequences generated from a physics-informed simulation of fertilizer particle motion incorporating gravity and aerodynamic drag. The proposed method provides an alternative to tray-based calibration, ground-deposition analysis, and individual particle tracking, which are labor-intensive or unreliable under high particle overlap. The dataset included 297 synthetic videos sampled into 25 sequential frames, consisting of 162 baseline scenarios with variations in disc diameter, disc speed, and disc height, and 135 additional scenarios incorporating wind speed and wind direction. Spatial features were extracted using a ResNet-18 backbone, combined with temporal positional encoding, and processed by a Transformer encoder. Model performance was evaluated using held-out synthetic samples, independent EDEM simulations, laboratory collector-bin measurements, and real image sequences captured under controlled conditions. On held-out synthetic test data, the generalized model achieved an MAE of 0.529 m, RMSE of 0.695 m, R 2 of 0.996, MAPE of 8.06%, bias of − 0.142 m, and MaxAE of 1.51 m. Compared with EDEM simulations, it achieved an MAE of 0.535 m, RMSE of 0.594 m, R 2 of 0.954, MAPE of 10.61%, bias of − 0.40 m, and MaxAE of 0.99 m. Laboratory validation using real image sequences also showed close agreement between baseline-model predictions and measured swath widths. These results demonstrate the feasibility of the approach, although field-scale validation under terrain slope, dust, vegetation occlusion, and variable illumination is still required.