Scientists have shown that a multimodal knowledge‑distillation model, Soil Moisture Distillation (SMD), reduces prediction error for soil moisture. In a benchmark on the MIS‑ME dataset, SMD cuts mean absolute percentage error (MAPE) for soil patches by 2 % and for crop images by 1.2 % compared to current state‑of‑the‑art baselines.

The method trains a single student network that processes both meteorological tables and either soil‑patch or crop images. Two teacher models, each trained on one modality alone, generate high‑quality feature representations. During training, the student aligns its own features with those of the teachers using a feature‑distillation loss, while simultaneously minimizing the usual mean‑squared‑error (MSE) loss against ground‑truth soil‑moisture values.

Experiments followed a 65 : 15 : 20 train/validation/test split. Compared with normal fusion and the MIS‑ME baseline, SMD achieved the lowest MAPE and improved other metrics such as root‑mean‑square error and coefficient of determination (R²).

These results highlight that combining meteorological data and visual cues through knowledge distillation leads to more accurate soil‑weather predictions, but further testing is needed before the model can be deployed in operational settings.