Random forest classification of land use in hilly and mountaineous areas of southern China using multi-source remote sensing data

Land use has been critical to global environmental change and structure adjustment, particularly to the sustainable development of land resources. However, there are complex terrains, broken distribution of ground objects, as well as the cloudy and rainy weather in hilly and mountainous areas of southern China. High-resolution optical remote sensing data is still lacking for the effective and accurate extraction of land use information. Therefore, the use of multi-source remote sensing data can achieve complementary advantages between remote sensing data and classification accuracy. The Sentinel series of remote sensing satellites launched by the European Space Agency (ESA) can provide new data sources for land-use change research. Multi-dimensional features can be adopted for the land use classification using the Sentinel-2A with red edge characteristics and Sentinel-1 with the nearly fog-free performance. Taking the reaches of Dongjiang River in Jiangxi Province of China as the study area, 9 schemes were designed in the Random Forest (RF) classification of land use to explore the effect of red edge, radar and terrain features on the extracting accuracy in hilly and mountainous areas of South China. In this study, the satellite images from the Sentinel-1, Sentinel-2 and digital elevation model (DEM) were combined to extract 27 feature indices, and then to construct 6 feature variable sets. The RF and Recursive Feature Elimination (RFE) were coupled to rank the importance of feature variables for the optimal one. The classification data from the RF feature selection was compared with the Support Vector Machine (SVM) and K-Nearest Neighbor (KNN). The results showed that the Sentinel-2A spectral features extraction presented the lowest overall accuracy and Kappa coefficient of land use classification in the study area, when the feature variables were not optimized. The addition of red edge, radar and topographic features effectively improved the classification accuracy, when the spectral features, vegetation and water indices were taken as basic schemes. Specifically, the overall accuracy increased by 0.77, 1.79, and 4.27 percentage points, respectively, while, the Kappa coefficient increased by 0.94, 2.18, and 5.2 percentage points, respectively. The topographic features more contributed to the extraction of orchard and cultivated land information in the study area. The RF and recursive feature elimination were combined to optimize all the feature variables from 21 to 13 with an overall accuracy of 0.937 2 and Kappa coefficient of 0.923 4, while maintaining the optimal classification accuracy. There were relatively significant contribution rates of spectral and red edge features variables, which were 26.09% and 23.55%, respectively. The vegetation and topographic indices were then followed in the importance of feature variables. The RF classification depended mainly on the short infrared band of B12, Relative Normalized Difference Vegetation Index (RNDVI) and Ratio Vegetation Index (RVI).The overall accuracy of RF was 0.937 2, 5.75% and 6.6% higher than that of SVM and KNN, respectively, whereas, Kappa coefficient was 0.923 4, 7.1% and 8.15% higher than SVM and KNN, respectively, indicating that the RF classification accuracy was superior to SVM and KNN with the same features. Therefore, the RF classification using the multi-source data can provide a promising technical support and theoretical reference for the extraction of land use in the hilly and mountainous regions of South China. © 2021, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.