Histogram based object classification using 1D Convolution Neural Networks

Keywords: Deep Learning, Remote Sensing, Histogram, 1D CNN, Object-based classification
Abstract Type: Paper Abstract

Abstract

Classification of high-spatial-resolution (HSR) remotely sensed (RS) data using object-based image analysis (OBIA) has shown promising results over the years. Similarly, the analysis of RS data using deep learning (DL) is also increasing. However, research that combines both DL and OBIA has been limited. The main limitation is that DL requires the input data in a fixed shape but in OBIA, the input data are irregularly shaped image segments. Hence, transforming the irregularly shaped input segments to regularly shaped deep learning features is the primary task while performing classification. The other limitation of DL is that the volume of the training data must be large for the algorithm to learn different signatures. This study proposes a novel method that uses histograms of the object as an input for the deep learning algorithm. In this method, an object is represented as a histogram with equal-size bins in a one-dimensional array. This one-dimensional array is classified using one-dimensional convolution neural networks (CNN). The data used for the study is high-resolution WorldView-2 imagery with 8 multi-spectral bands. In this study, the number of training segments used to define the algorithm is less, and the overall accuracy that was obtained by classifying the image using DL and histograms of segments is 98%.