| Abstract [eng] |
This research project addresses the estimation of crop growth stages on the BBCH scale from satellite data as a regression task. Ground-truth measurements collected with the Smart Agrometer for corn (14 observations, BBCH 10–51) and winter wheat (BBCH 21–69, two growing seasons) were combined with Sentinel-1 SAR backscatter (VH, VV) and Sentinel-2 multispectral reflectance via the Copernicus Data Space API, with a ±2-day temporal tolerance and spatial averaging over field polygons. Fourteen input features were produced, including raw spectral bands and three vegetation indices (NDVI, NDWI, RVI). Three models of increasing complexity were trained and compared: Ridge regression, a feed-forward neural network (FFNN), and a Long Short-Term Memory (LSTM) network, with hyperparameters tuned using Bayesian optimization. For corn, the LSTM achieved an MSE of 0.23 on an independent test set, against 13.37 for the FFNN and 70.03 for Ridge regression. For winter wheat, evaluated by 6-fold GroupKFold cross-validation with field-level separation, the LSTM reached R² = 0.639, RMSE = 8.18, and MAE = 5.45, outperforming the FFNN (R² = 0.451) and Ridge (R² = 0.205). Permutation Importance, SHAP, and LIME revealed crop-specific feature hierarchies: NDVI and NDWI dominated for corn, while RVI together with near-infrared and red- edge bands dominated for winter wheat. An XAI-driven feature selection experiment (reducing 14 features to 9) improved the FFNN (R² 0.451 → 0.520) but slightly degraded the LSTM. The main limitations are the small dataset size and the underrepresentation of late-season stages (BBCH ≥ 60), which led to systematic underestimation at advanced maturation phases. The results confirm that sequence-based deep learning is the most suitable approach for continuous BBCH regression from fused radar and optical satellite data. |
