Mapping 30-m Resolution Land Cover of China Based on Full Convolutional Neural Network

Abstract

AbstractLarge-scale land cover maps are essential geoscience information for many applications. However, the cost of the current mapping process greatly limits the update frequency of large-scale land cover products. Moreover, the inconsistency among land cover maps also hindered further long time-series land cover change analysis and research. Therefore, we proposed a novel framework to efficiently generate new land cover maps using historical land cover products. The framework mainly includes the following three innovations: (1) To solve the influence of clouds and seasonal inconsistency, use Google Earth Engine to composite the multi-temporal remote sensing images to generate consistent cloud-free images; (2) To avoid the repeated collection of training samples, training data were generated by integrating multiple historical products; (3) To process massive generated training data and improve mapping accuracy, the data-driven deep fully convolutional network model is used to achieve end-to-end land cover mapping. Based on the proposed mapping approach, the 30-m resolution land cover map of China in 2015 was automatically completed with improved accuracy, which shows the potential for frequent large-scale land cover product integration and updating.KeywordsLand cover mappingGoogle earth engineDeep learningFull convolutional network