Attention-based 3D convolutional neural network for crop boundary detection in high-resolution satellite image time series

Advancements in satellite missions have dramatically improved the monitoring of vegetation and agricultural activities through high-resolution Satellite Image Time Series (SITS), providing enhanced insights into crop dynamics and boundary identification. However, traditional UNet-based Convolutional Neural Networks (CNNs), though effective for crop mapping, often struggle to capture the full spatio-temporal complexities inherent in these datasets, particularly when it comes to detecting less distinct boundaries. To address these challenges, a novel attention-based residual 3D UNet architecture has been developed, incorporating a spatial-temporal attention mechanism that enhances the networks ability to represent spatial and temporal features. This attention mechanism is strategically implemented in the decoder, where it gathers information from both the encoder and the previous layer within the decoder. This dual-source integration allows the model to focus more effectively on relevant crop boundaries during training, assigning greater weight to these crucial areas while reducing the emphasis on non-crop regions. The residual 3D UNet architecture adeptly handles the intricate spatial-spectral-temporal correlations present in SITS, enabling more accurate and simultaneous modelling of both spatial and temporal information. The proposed method is evaluated on an area with small-scale crop fields in Germany using Sentinel-2 SITS data collected over several months, this approach demonstrated superior performance in boundary detection compared to existing state-of-the-art methods, particularly in scenarios where boundaries are less clearly defined.