Fusion of Landsat 8 and Sentinel-2 data for mangrove phenology information extraction and classification

Abstract

Scientific and accurate monitoring of mangroves is the basis and premise for the protection of marine coastal and transitional ecosystem. However, mangroves are mainly distributed in the intertidal zone, result in large-scale manual monitoring a tough task. Although remote sensing technology can map mangrove within for long time and large area, the existing studies have some shortcomings. On the one hand, mangroves are often distributed in tropical and subtropical regions, where long-term coverage of effective optical remote sensing data is difficult to obtain due to weather conditions. On the other hand, mangroves are easily confused with other terrestrial vegetation by only using spectral information. In this paper, we choose the Sundarbans located in the Ganges River Delta as the study area. Landsat 8 OLI and Sentinel-2 MSI data in 2016 are obtained based on GEE (Google Earth Engine) to conduct mangrove extraction in this research. Firstly, the relation between the two sensors for the same index is constructed based on a least square regression model, which is used to reconstruct the time series data. In this phase, EVI (Enhanced Vegetation Index) and LSWI (Land Surface Water Index) are selected according to the separability criterion. Secondly, Savitzky-Golay filtering is applied to the time series data of the two indices, and 13 phenology metrics are extracted. Finally, these metrics of the two indices are cascaded, and Random Forest (RF) is used to extract the area of mangrove. Fusing the Landsat 8 OLI and Sentinel-2 MSI can effectively improve the quality of time series data. Compared with the classification results based on single sensor data, the overall accuracy is improved by 1.58%. Phenology information can significantly enhance the separability between mangrove and other vegetation, with a 1.92% improvement of overall accuracy compared with the classification results using time series data directly. Considering both EVI and LSWI indices can greatly improve, the classification effect, with 14.11% and 9.69% improvements compared with using a single index. Therefore, the method in this paper can effectively extract mangroves, with the overall accuracy and Kappa coefficient reaching to 91.02% and 0.892 respectively. This research takes full account of the deficiency of optical remote sensing monitoring, biological characteristics and geographical characteristics of mangroves, which can extract the area of mangroves more objectively and accurately. Compared with other similar studies, the differences and characteristics of this study are: (1) Jointly using EVI and LSWI time series to describe the phenological information of mangroves can effectively differentiate mangroves and evergreen forests; (2) We introduce phenology information into mangrove classification using remote sensing for the first time, and verify the feasibility of using phenology information to monitor the range of mangrove. The method proposed in this paper may be benefit for scientific and accurate monitoring of global or regional mangroves.