Assessing generalization of deep learning models for crop classification under climatic variability in Denmark

This study explores the impact of climatic variability on the generalization capabilities of a deep learning model for pixel-level crop classification using multi-temporal Sentinel-1 SAR data in Denmark. With agriculture accounting for 61% of Denmark’s land area, accurate and timely crop mapping is essential for providing insights into crop distribution, offering valuable information to advisors and authorities to support large-scale agricultural management, and address challenges posed by changing climatic conditions.

Our study leverages a novel deep learning architecture that combines a 3-D U-Net with a conv-LSTM module to effectively capture both spatial and temporal dependencies in crop growth patterns. We consider 14 crop types over an eight-year period (2017–2024) and growth season (May to August), with ground truth data derived from Denmark’s Land Parcel Identification System (LPIS). Our analyses reveal that climatic variables such as precipitation, temperature, and humidity significantly influence model performance across years. Notably, extreme years like 2018 (characterized by drought and high solar radiation) and 2024 (marked by record precipitation) challenge the model’s ability to generalize effectively. By correlating inter-annual model accuracy trends with climatic data, the study demonstrates the necessity of incorporating environmental context into AI-driven agricultural monitoring systems.

We also evaluate the benefits of training the model on multi-year datasets to enhance robustness against climatic variability. Our findings reveal that leveraging temporal diversity improves model performance but highlights persistent difficulties in generalizing to outlier years with extreme climate conditions. While training on multi-year datasets helps capture a broader range of crop phenological variations, the results underscore that this approach alone is not sufficient, and underscores the importance of integrating auxiliary data, such as local climatic variables, to enable models to better adapt to evolving crop growth patterns influenced by changing environmental conditions.

This work represents one of the most comprehensive evaluations of deep learning for crop classification, spanning eight years and covering over 1.5 million hectares of agricultural land. By linking model performance to climatic variability, this study provides critical insights for improving the generalization capabilities of deep learning models in precision agriculture. These findings not only pave the way for enhanced crop monitoring under diverse climatic scenarios but also emphasize the potential of integrating climate-resilient AI technologies to address global agricultural and environmental challenges.