Deep Learning for Extracting Water Body from Landsat Imagery

There are regional limitations in traditional methods of water body extraction. For different terrain, all the methods rely heavily on carefully hand-engineered feature selection and large amounts of prior knowledge. Due to the difficulty and high cost in acquiring, the labeled data of remote sensing is relatively small. Thus, there exist some challenges in the classification of huge amount of high dimension remote sensing data. Deep Learning has a good capacity of hierarchical feature learning from unlabeled data. Stacked sparse autoencoder (SSAE), one deep learning method, is widely investigated for image recognition. In this paper, a new water body extraction model based on SSAE is established. At first, current useful features (NDWI, NDVI, NDBI and so forth) are collected to construct unique feature matrix for each pixel. Next, a Feature Expansion Algorithm (FEA) is designed by taking account of the influence of neighboring pixels to expand feature matrixes. Setting the expansion features as inputs, SSAE is trained to extract water body. The experimental results showed that the proposed model outperformed Support Vector Machine (SVM) and traditional neural network (NN). Meanwhile, the proposed FEA explored more distinct features of water body so that the accuracy of water body extraction was improved to a great extent.