SepHRNet: Generating High-Resolution Crop Maps from Remote Sensing imagery using HRNet with Separable Convolution

Accurate mapping of crop production is essential for ensuring food security, efficient resource management, and sustainable agricultural practices. Recent advances in deep learning have shown significant potential in achieving high accuracy for crop production mapping using high-resolution satellite imagery. However, effectively capturing complex and variable crop patterns remains a challenge. In this study, a novel deep-learning model has been proposed that combines High-Resolution Networks (HRNet) with Separable Convolutional layers in the shallow layers to represent spatial patterns of crops at low computational cost. A multi-head attention mechanism further enhances the model by capturing long-term temporal dependencies in the encoded image representations. The HRNet model serves as a backbone, extracting high-resolution features from crop images. Subsequently, a CNN decoder generates the final crop map from the aggregated feature representation. The proposed model achieves a notable classification accuracy of 97.5% and an Intersection over Union (IoU) of 55.2% in generating crop maps. Evaluation on the Zuericrop dataset demonstrates that our approach outperforms several state-of-the-art models, including U-Net++, ResNet50, VGG19, InceptionV3, DenseNet, EfficientNet, and ms-ConvSTAR. This research underscores the potential of Deep Learning for large-scale Earth Observation Systems.