Remote sensing parameters optimization for accurate land cover classification

Sustainable natural resources management requires considerable accurate land cover information given the evident climate change impacts and human disturbances on wetlands. It is characterized by the convergence of numerous materials and energies, resulting in fragmented landscapes and frequent land cover changes. To address the challenges posed by the complexity of landforms, diversity of land cover types, and non-linearity of remote sensing image features in traditional remote sensing image classification methods, this paper proposes a feature parameter selection method based on the Gini index of random forests, with a 10% threshold decision. The aim is to identify the optimal combination of remote sensing feature parameters. Firstly, spectral features, texture features, thermal features, elevation features, and principal component features are selected to form a stack of remote sensing images. Then, multiple decision trees are set up to cross-validate the contributions of the features, and the feature ranking is determined based on the normalized mean importance of the features. Finally, a threshold is set to select the remote sensing feature parameters that meet the requirements, and the process is iterated. Experiments are conducted using Sentinel-2 remote sensing images covering the Yancheng Nature Reserve in Jiangsu province. The results show that the remote sensing feature parameters selected by this method have good representativeness. Compared with CART, SVM, KNN, and RF methods that only use band information, the proposed method produces clearer boundaries and more accurate category attributes in the classification results, with an overall accuracy of 96.20% and a Kappa coefficient of 0.9556. This research can provide technical support for regional spatial planning and sustainable development. © 2024 SinoMaps Press. All rights reserved.