Accurate and timely crop yield prediction and forecasting are important for improving agricultural productivity and supporting informed management decisions. In this study, we developed and evaluated a framework for estimating in-season potato yield at the field scale using Sentinel-2 satellite time series. A Grouped TreeSHAP Stability Selection (GTSS) was first applied to identify a compact, phenology-aware subset of spectral bands and vegetation indices, thereby reducing redundancy and mitigating overfitting in small-data settings. Two deep learning architectures tailored for limited training data were then introduced: LiteTemporalConv, a lightweight temporal convolutional network, and MS-ConvBiGRU-Attn, a hybrid encoder combining multi-scale convolutions, bidirectional GRUs, and an attention mechanism. Both models were benchmarked against widely used machine learning methods, including Random Forest, Support Vector Machine, Extreme Gradient Boost, Partial Least Squares, as well as standard deep learning baselines (CNN and GRU). Results showed that the proposed models outperformed both machine learning and conventional deep learning baselines, with LiteTemporalConv achieving the highest accuracy under 10-fold cross-validation (R² = 0.84; RMSE = 3.18 t ha⁻¹; rRMSE = 6.36%) and MS-ConvBiGRU-Attn yielding similarly strong performance (R² = 0.82; RMSE = 3.53 t ha⁻¹; rRMSE = 7.03%). By comparison, the best baseline, XGB, achieved an R² of 0.79 with an rRMSE of 9.8%. The two best-performing models were further evaluated on an independent spatial dataset to assess their generalization beyond the training region. In an additional experiment, both deep learning models trained on mid-season observations showed predictive stability for late-season yield estimation. Overall, the results highlight the importance of targeted feature selection and lightweight encoders for yield modeling in data-scarce conditions.
Tufail et al. (Sun,) studied this question.