DensePPMUNet-a: A Robust Deep Learning Network for Segmenting Water Bodies From Aerial Images

Abstract

The identification of water bodies from aerial images using semantic segmentation networks can provide accurate information for ecological monitoring, flood prevention, and disaster reduction. Outliers on aerial images might reduce interclass separability and thus cause discontinuous prediction of water bodies. The fusion of global context information is helpful to solve this problem. However, the existing global prior representation does not provide sufficient information for identifying a large number of multi-scale objects and outliers. In this study, a dense pyramid pooling module (DensePPM) was introduced to extract global prior knowledge with a dense scale distribution. The ablation experiments showed that the models using the DensePPM had higher values of IoU, F1-score, and Recall than that using pyramid pooling module (PPM), showing that the proposed module could capture more global context information of outliers under multi-scale scenarios. A robust deep learning network named DensePPMUNet-a based on the DensePPM was then proposed for segmenting water bodies from aerial images. The comparative experiments with different datasets demonstrated that the DensePPMUNet-a outperformed U-Net, CE-Net, MultiResUNet, ResUNet-a, PSPNet, LANet, DeepLabV3, MANet, and FactSeg.