Developing Long Time Series 1-km Land Cover Maps from 5-km AVHRR Data Using a Super-Resolution Method

Abstract

Dynamic land cover (LC) information is an essential part of environmental and ecological research. Therefore, acquiring dynamic LC data with high spatial resolution has attracted a great deal of attention in the remote sensing community. Nevertheless, the high-temporal resolution satellite data tend to have a coarse spatial resolution, and satellite data with high temporal resolution are often relatively low. Obtaining LC with high spatiotemporal resolution is extremely challenging. The super-resolution method can help researchers achieve this goal, and the recently developed neural-network-based deep learning algorithms have great potential for use as an alternative solution. This study proposes a focal loss temporal convolutional long short-term memory (FL-T-ConvLSTM) model for super-resolution LC classification research. It first trains the deep FL-T-ConvLSTM network to establish a transformation between low-resolution quantitative remote sensing parameters and high-resolution quantitative remote sensing parameters and then engages in nonlinear mapping with a high-resolution LC map. A long-term series 1-km super-resolution LC classification model based on deep learning was established and applied to the Beijing-Tianjin-Hebei region. Based on this method, a long-term series of 1-km LC maps from 1982 to 2019 can be obtained. The test accuracy and field validation accuracy of the model reached 90.1% and 86.8% when using reliable test samples and field test samples, respectively. This study provides a method for obtaining high-resolution LC classification products from low-resolution quantitative remote-sensing products.