Land Cover Classification for Satellite Images Through 1D CNN

Land cover classification of satellite imagery can provide significant information for many applications, including surface analysis, environmental monitoring, building reconstruction, etc. Land cover classification has been generally performed using unmixing-based or shallow/deep learning approaches, among which the unmixing-based approaches suffer from stability issues due to the complex intrinsic properties of the data, deep learning-based approaches like 2D CNN requires large labeled training set which is often unavailable in satellite images and small ground truth collection leads to spatial discontinuities (as shown in Fig. 1), making 2D CNN approaches unviable. In this paper, we first propose a 1D convolution neural network-based framework applied to each pixel in the spectral domain where we extract descriptive local features for improved classification. Experimental results demonstrate superior classification accuracy through comparison with traditional unmixing-based and neural network methods using just limited number of training samples.