Integrating Google Earth imagery with Landsat data to improve 30-m resolution land cover mapping

Abstract

© 2019 Land use and land cover maps provide fundamental information that has been used in different kinds of studies, ranging from climate change to city planning. However, despite substantial efforts in recent decades, large-scale 30-m land cover maps still suffer from relatively low accuracy in terms of land cover type discrimination (especially for the vegetation and impervious types), due to limits in relation to the data, method, and design of the workflow. In this work, we improved the land cover classification accuracy by integrating free and public high-resolution Google Earth images (HR-GEI) with Landsat Operational Land Imager (OLI) and Enhanced Thematic Mapper Plus (ETM+) imagery. Our major innovation is a hybrid approach that includes three major components: (1) a deep convolutional neural network (CNN)-based classifier that extracts high-resolution features from Google Earth imagery; (2) traditional machine learning classifiers (i.e., Random Forest (RF) and Support Vector Machine (SVM)) that are based on spectral features extracted from 30-m Landsat data; and (3) an ensemble decision maker that takes all different features into account. Experimental results show that our proposed method achieves a classification accuracy of 84.40% on the entire validation dataset in China, improving the previous state-of-the-art accuracies obtained by RF and SVM by 4.50% and 4.20%, respectively. Moreover, our proposed method reduces misclassifications between certain vegetation types, and improves identification of the impervious type. Evaluation applied over an area of around 14,000 km2 confirms little improvement for land cover types (e.g., forest) of which the classification accuracies are already over 80% when using traditional machine learning approaches, yet improvements in accuracy of 7% for cropland and shrubland, 9% for grassland, 23% for impervious and 25% for wetlands were achieved when compared with traditional machine learning approaches. The results demonstrate the great potential of integrating features of datasets at different resolutions and the possibility to produce more reliable land cover maps.