URBAN LAND COVER CLASSIFICATION with VERY HIGH RESOLUTION SATELLITE IMAGERY by MACHINE LEARNING

Abstract

Land cover classification is an important topic of satellite image analysis. In recent years, machine learning image classification has been increasingly used in land mapping and feature extraction showing advantages in many cases. In this paper we used a machine learning method for urban land cover classification using WorldView-2 imagery with 8 multispectral bands at 2-m resolution and a 0.5-m panchromatic band. The WorldView-2 imagery was acquired on 26 Jan 2020 covering a region of interest (ROI) in Singapore from 103°48'22.921"E to 103°51'48.901"E and 1°24'20.486" N and 1°21'6.606"N was chosen in this study. After radiometric correction and spectral-preserving pan-sharpening, the pan-sharpened multispectral reflectance was used for classification. Eight classes (Tree, Grass, Cloud, Water, Building, Bare soil, Shadow and Road/Paved) were defined. A thousand samples, each with 11 spectral features wee extracted from the imagery. The features were 8 spectral bands from the pan-sharpened multispectral image plus NDVI = (NIR1 - Red)/(NIR1 + Red), REVI = (RedEdge-Red)/(RedEdge + Red) and NYVI = (NIR2 - Yellow)/(NIR2 + Yellow). Eighty percent of the samples were used to train a Convolutional Neural Network (CNN) model with depth 5 and kernel size 5 x 5 pixels. The remaining 20% were used for validation and an accuracy of 92.1% was achieved. The trained CNN was applied to classify the whole ROI and the result was compared to a ground truth map. Two regions were selected for accuracy assessment. One was dominated with man-made objects (Buildings and Road/Paved), the other mainly contains natural objects (Tree, Grass, Water and Bare soil). The overall accuracy was 94%. This high accuracy strongly suggests that the machine learning method is a good approach for urban land cover classification using very high resolution satellite imagery.