An Integrated Approach of DeepLabV3+ with an Ensemble Backbone for Precise Water Bodies Segmentation in Remote Sensing Imagery

The segmentation of water bodies in very high-resolution satellite and aerial images is a crucial task in remote sensing, essential for applications such as environmental monitoring, urban planning, and disaster management. Traditional methods often struggle to cope with the intricacies and high variability of these high-resolution images. In this paper, we address these challenges by leveraging the DeepLabV3+ model, a state-of-the-art architecture for semantic segmentation, enhanced through the integration of an ensemble of feature extraction backbones. Our approach employs ResNet50V2, MobileNetV3, and EfficientNetV2 as backbones in the encoder part of the DeepLabV3+ model, utilizing a weighted averaging ensemble technique to combine their strengths. This ensemble method is designed to further boost the segmentation performance by refining the accurate extracted features and preserving the information contained in these extracted features. We chose the backbones for their synergy: ResNet50V2 complementing performance, MobileNetV3 complement computational efficiency, and EfficientNetV2 complement scalability. The results show substantial improvements in key performance metrics, with the ensemble model achieving measures around 94% accuracy and a mean IoU of about 85%. Moreover, the precision as well as recall metrics showed substantial improvement which further justifies that the weighted averaging technique helped to combine the output of the implemented backbone. This study also highlights the benefits of ensemble-based methods for refining segmentation results for very high spatial resolution satellite data.