Multi-temporal remote sensing and artificial intelligence methods for crop mapping and yield estimation

The increasing demand for sustainable agricultural intensification and global food security requires accurate crop mapping and reliable yield prediction. Multispectral satellite imagery, particularly Sentinel-2, provides consistent high-resolution data for monitoring spatial and temporal variations in crop health and productivity. However, subtle spectral differences among crops and redundancy in high-dimensional time-series data make classification and yield estimation challenging. This thesis investigates the integration of Sentinel-2 spectral bands and vegetation indices with advanced machine learning (ML) and deep learning (DL) approaches for crop classification and yield prediction in complex agricultural landscapes of Northern Italy. The first part of the study evaluates optimal feature selection strategies using Random Forest (RF) importance and Principal Component Analysis (PCA). Selected features were tested with Extreme Gradient Boost (XGB), RF, Support Vector Machine (SVM), and Pixel-Set Encoder with Temporal Self-Attention (PSETAE). Results indicate that RF-based feature selection applied to spectral bands, combined with XGB, achieved the highest classification accuracy, highlighting the importance of robust feature optimization. Building on these findings, hybrid deep learning architectures were developed to model spatial and temporal dependencies in Sentinel-2 time series. Four models, 1D CNN-LSTM, 1D CNN-GRU, 2D CNN-LSTM, and 2D CNN-GRU, were evaluated using combined spectral and vegetation index inputs. The 2D CNN-GRU model achieved the best performance (overall accuracy 99.12%; F1-macro 99.14%), demonstrating the advantage of integrated convolutional and recurrent frameworks. Finally, the research extends to field-scale potato yield prediction using a Grouped TreeSHAP Stability Selection (GTSS) method to reduce redundancy and mitigate overfitting. Two tailored deep models, LiteTemporalConv and MS-ConvBiGRU-Attn, outperformed conventional ML baselines, achieving R² values up to 0.84. Overall, the study demonstrates that combining Sentinel-2 data with optimized ML and DL frameworks provides a scalable, data-driven approach for crop classification and yield prediction, supporting precision agriculture and sustainable resource management.