Shallow Parallel CNNs for contextual remote sensing image classification

Abstract In this paper we present a new neural network structure that can better learn to classify remote sensing images of moderate and high spatial resolution where the main source of information about desired objects are the pixels themselves and the tight neighborhood. It enhances the pixel-based classification process by incorporating the contextual information in its surroundings.The proposed algorithm is an arrangement of small Shallow Parallel Convolutional Neural Network layers, SP-CNN, that are centered, each of them, over training or test pixels. Each SP-CNN drives information from the pixel to be classified and its contextual neighborhood. Depending on the information that may be found in the context of each pixel, the size and the number of SP-CNNs to be used in the proposed structure can be learned or optimized.The proposed method has been applied to Sentinel-2 (10 m resolution) and Pl\'eiades data (0.5 m resolution) and gave superior results when compared to 1-D CNN and other pixel-based methods such as KNN, SVM, and RF.The proposed method shows its suitability when the training and test data are pixel-based and in same time the spatial domain has valuable information to be incorporated in the classification process.