Water Body Extraction from High-resolution Remote Sensing images Based on Scaling EfficientNets

Abstract

Effectively and accurately extracting water bodies from high-resolution remote sensing images is an important yet extremely challenging task due to ambiguities raised by spectral and texture similarities between water bodies and distractors such as shadows buildings, mountains and vegetations. Recent developed deep learning techniques have provided researchers powerful tools to defeat traditional hand-engineered features by learning powerful hierarchical representations, however, with a cost of very high model complexity and computational resource requirement. To conquer this issue, in this paper, we propose to build a deep learning model for water extraction based on the EfficientNet-B5. We evaluate possible variations of such design and compare them with several widely used approaches on GF-2 and Sentinel-2 satellites. Experimental results demonstrate that our model obtains much better performance than SVM and U-Net, which clearly shows the effectiveness and robustness. We hope this work will facilitate further research in this field and inspire researchers in the community to develop better models.