RS-DCNN: a novel distributed convolutional-neural-networks based-approach for big remote-sensing image classification

Abstract

Developments in remote sensing technology have led to a continuous increase in the volume of remote-sensing data, which can be qualified as big remote sensing data. A wide range of potential applications is using these data including land cover classification, regional planning, catastrophe prediction and management, and climate-change estimation. Big remote sensing data are characterized by different types of resolutions (radiometric, spatial, spectral, and temporal), modes of imaging, and sensor types, and this range of options often makes the process of analyzing and interpreting such data more difficult. In this paper, which is the first study of its kind, we propose a novel distributed deep learning-based approach for the classification of big remote sensing images. Specifically, we propose Distributed Convolutional-Neural-Networks for handling RS image classification (RS-DCNN). The first step is to prepare the training dataset for RS-DCNN. Then, to ensure a data-parallel training on the top of the Apache Spark framework, a pixel-based convolutional-neural-network model across the big data cluster is performed using BigDL. Experiments are conducted on a real dataset covering many regions of Saudi Arabia and the results demonstrate high classification accuracy at a faster speed than other state-of-the-art classification methods.