Enhancing Crop Classification in Emilia-Romagna (Italy) Using Transformer-Based Multi-Source Data Fusion with Thermal Observations

Abstract. This study explores the potential of integrating multi-source remote sensing data—including Sentinel-1 synthetic aperture radar (SAR) imagery, Sentinel-2 optical imagery, and Landsat 8 thermal data—for crop classification in Emilia-Romagna (Northern Italy). Using satellite imagery and agricultural surveys, we constructed a temporal dataset covering 2020 with 27 biweekly time steps. After filtering out underrepresented crop types with insufficient samples for machine learning training, nine crop types remained. We implemented four deep learning models using TensorFlow: Dense Neural Network (DNN), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Transformer. Our results indicate that removing underrepresented crops significantly improves classification performance, leading to an overall accuracy of approximately 91%. Incorporating Landsat 8 thermal data further enhanced accuracy, with the Transformer model achieving a peak accuracy of 92.08%. A crop-specific analysis revealed that temperature observations notably improved classification for crops with distinct thermal signatures (e.g., sugar beets, corn), whereas limited improvement was observed for spectrally similar cereals (e.g., wheat, barley). Overall, the Transformer model demonstrated exceptional ability in capturing spatial-temporal dependencies in multivariate time-series data. These findings underscore the advantages of integrating multi-source satellite data including thermal infrared and leveraging attention-based neural networks for large-scale agricultural monitoring and resource management.