An Automatic Method for Water Extraction From High Spatial Resolution GF-1 Imagery Based On A Deep Learning Algorithm

Abstract

Traditional remote sensing methods for water extraction generally rely on complex spectral analysis and subjective thresholds, which do not satisfy the requirements for automated high-accuracy water detection from high spatial resolution images. In this paper, a novel framework of Convolutional Neural Networks (CNNs) is proposed for automatic water extraction based on Densely Connected Convolutional Networks (DenseNet) and a U-net structure. The proposed CNN model is trained on 1300 GF-1 images and tested on another 50 GF-1 images. Six of the GF-1 images are used for experiments to demonstrate the accuracy of the results and compare them with those from NDWI and supervised classification methods. The results show that our method has better computation efficiency and classification accuracy. The new framework has remarkable advantages for automatic water extraction from high-resolution images, especially in maintaining the integrity of small water bodies and in overcoming interference from clouds, shadows in built-up-areas, and mountainous areas with snow cover.