Atrous convolutions and ConvLSTM fusion for multi-scale feature learning in EuroSAT satellite imagery

Satellite imagery classification plays a crucial role in land cover monitoring, urban planning, and environmental conservation. This paper presents a deep learning architecture designed to improve multi-scale feature learning for satellite image classification on the EuroSAT dataset. The model combines a ResNet50 backbone with atrous convolutions to capture multi-scale contextual information across multiple spectral bands, essential for accurately differentiating complex land-cover classes. The Convolutional Block Attention Module (CBAM) refines the spatial and channel features, focusing on details critical to classification. The Convolutional Long Short-Term Memory (ConvLSTM) module is incorporated as a recurrent spatial refinement stage operating over latent feature representations. Its gated update mechanism enables iterative feature modulation while preserving spatial structure, providing a alternative to purely feed-forward refinement strategies. The model achieves a mean Top-1 accuracy of approximately 97% across multiple random seeds, with a peak accuracy of 98%.