Materials and Methods

Land use classification is a key tool for understanding and monitoring both natural and human-modified landscapes. It supports environmental protection, sustainable agricultural management, and urban planning. Its importance is also reflected in the Land Use, Land-Use Change and Forestry (LULUCF) framework, which is closely linked to European Union climate policies, carbon stock changes, greenhouse gas emissions, and ecosystem resilience. Satellite remote sensing provides an efficient and cost-effective way to monitor land use over large areas. However, accurately separating similar land cover types, especially in agricultural areas with small parcels and strong seasonal changes, remains a difficult task. This study investigates the use of deep learning methods to improve land use classification using Sentinel-2 imagery, supported by suitable preprocessing steps and additional ancillary data.
We compiled a time series of Sentinel-2 imagery, computing monthly median reflectance values for each spectral band, resulting in 12 composite images. To ensure data quality, individual images with more than 20% cloud cover were excluded before median calculation. Further filtering was applied at the pixel level using the Normalized Difference Snow Index (NDSI) and the Normalized Difference Water Index (NDWI), setting a <30% threshold to remove pixels associated with snow cover and water bodies. The study was conducted in the Eastern Macedonia and Thrace Region, utilizing 2022 Land Parcel Information System (LPIS) data provided by the National Paying Agency of Greece as ground truth, including the field boundaries and crop type labels.
We evaluated multiple deep learning models for land use classification, including and Temporal CNN Bidirectional Long Short-Term Memory (BiLSTM) Bidirectional GRU (BiGRU) Multi-Head Attention (MHA) and Transformer Encoder. Performance was assessed using the Classification Report from the scikit-learn library, which provided key metrics such as precision, recall, F1-score, and support. By analyzing results at the crop type level, we identified underperforming classes and refined the classification strategy by merging related categories, ultimately enhancing overall model accuracy.

Results and Evaluation

As the dataset is highly imbalanced among crop-type classes, we utilized the F1-score to select the best performing model which was temporal CNN achieving the value of 0.85, which is comparable to the reported State-of-the-Art at various works.

Conclusion

Overall, the study shows that multitemporal deep learning applied to Sentinel-2 parcel sequences provides a robust and scalable framework for operational crop type classification.