Improving land-cover classification accuracy with a patch-based convolutional neural network: data augmentation and purposive sampling

The unit of classification in land-cover mapping is generally divided into two main categories: pixel and object. When it comes to medium-resolution images, a pixel has generally been used as a unit of classification because the object-based approach is often not as effective due to its coarse resolution. Recently, however, the patch-based approach for land-cover classification has shown higher accuracy levels than the pixel-based approach by exploiting the informative features from neighboring pixels. In this study, the light convolutional neural network (LCNN) was used as a patch-based classification algorithm, and two methods to further improve the classification accuracy for patch-based algorithms were addressed. First, data augmentation by flipping and rotation was applied to LCNN to check if its classification accuracy can increase. Second, the purposive sampling, which considers the heterogeneity of a map, was applied to LCNN. This study shows that the classification accuracy of LCNN can be further improved by data augmentation and purposive sampling and thus confirms that the patch-based approach has a distinct advantage over the pixel-based approach.

I. Introduction

Land-cover classification is one of the most important applications of remote sensing. Crucial information on urban and environmental studies can be obtained by accurate land-cover maps, such as urban planning [1], carbon-cycle monitoring [2], and urban heat islands [3]. However, land-cover maps consist of misclassification and these errors might affect downstream studies and mislead the results [4]. Therefore, there has been enormous effort to improve the accuracy of land-cover map for a long time in the remote-sensing society [5], [6].