Jiemin Wu

Land cover information depicting the complex interactions between human activities and surface change is critically essential for nature conservation, social management, and sustainable development. Recent advances have shown great potentials of remote sensing data in generating high-resolution land cover maps, but it remains unclear how different models, data sources, and inclusive features affect the classification results, which hinders its applications in regional studies requiring more accurate land cover data. Informing these issues, here we developed a robust framework to improve the mapping results of 10 m resolution land cover classification in Guangdong Province, China using thousands of manually collected samples, multisource remote sensing data (Sentinel-1, Sentinel-2, and Luojia-1), machine learning algorithms, and a free cloud-based platform of Google Earth Engine. Results showed that an overall accuracy of 86.12% and a Kappa coefficient of 0.84 could be achieved for land cover classification in Guangdong for 2019. We found that random forest models achieved better performance than classification and regression trees, minimum distance, and support vector machine models. We also found that features derived from Sentinel-1 data and Sentinel-2 spectral indices greatly contributed to the classification process, while the feature of Luojia-1 data was not as much important as other configurations. A comparison between our results and several existing land cover products in terms of classification accuracy and visual interpretation further validated the effectiveness and robustness of the proposed framework. Our experiments and findings not only systematically elucidate the role of classification methods and data sources in deriving more accurate and reliable land cover maps but also provide certain guidelines for future land cover mapping from regional to global scales.

Sentinel-2 data is of great utility for a wide range of remote sensing applications due to its free access and fine spatial-temporal coverage. However, restricted by the hardware, only four bands of Sentinel-2 images are provided at 10 m resolution, while others are recorded at reduced resolution (i.e., 20 m or 60 m). In this paper, we propose a parallel residual network for Sentinel-2 sharpening termed SPRNet, to obtain the complete data at 10 m resolution. The proposed network aims to learn the mapping between the low-resolution (LR) bands and ideal high-resolution (HR) bands by three steps, including parallel spatial residual learning, spatial feature fusing and spectral feature mapping. First, rather than using the single branch network, the parallel residual learning structure is proposed to extract the spatial features from different resolution bands separately. Second, the spatial feature fusing is aimed to fully fuse the extracted features from each branch and produce the residual image with spatial information. Third, to keep spectral fidelity, the spectral feature mapping is utilized to directly propagate the spectral characteristics of LR bands to target HR bands. Without using extra training data, the proposed network is trained with the lower scale data synthesized from the observed Sentinel-2 data and applied to the original ones. The data at 10 m spatial resolution can be finally obtained by feeding the original 10 m, 20 m and 60 m bands to the trained SPRNet. Extensive experiments conducted on two datasets indicate that the proposed SPRNet obtains good results in the spatial fidelity and the spectral preservation. Compared with the competing approaches, the SPRNet increases the SRE by at least 1.538 dB on 20 m bands and 3.188 dB on 60 m bands while reduces the SAM by at least 0.282 on 20 m bands and 0.162 on 60 m bands.