HLEL: A wetland classification algorithm with self-learning capability, taking the Sanjiang Nature Reserve I as an example

Abstract

Wetland ecosystems have been severely damaged by atmospheric pollution and other environmental problems for decades. Reasonable protection, restoration, and utilization of wetlands have become priorities for environmental protection. Obtaining high-precision wetland map products is essential for ecological research as well as formulation and implementation of conservation policies. This study proposes a hierarchical labeling ensemble learning (HLEL) algorithm with self-learning capabilities to achieve high-precision wetland land cover and land use (LULC) classification by adapting self-learning hierarchical results. This study evaluated the classification performance of HLEL, random forest (RF), support vector machine (SVM), artificial neural network (ANN), and Gaussian Naive Bayes' (GNB) with GF-6 WFV multispectral images as the database and Sanjiang National Nature Reserve I in China as the study area. The results showed that the HLEL algorithm has an advantage over a single algorithm when using the feature set filtered by ReliefF, even if the complete feature set is used without feature selection. The overall accuracy of the classification results is improved by 0.4%, 5.8%, 1.3%, and 9.0% when compared to the four single algorithms RF, SVM, ANN, and GNB, respectively. In robustness tests, HLEL consistently outperformed the other algorithms for different training sample set sizes, with overall accuracy improvements of 0.4–6.9%, 5.8–11.5%, 1.3–4.4%, and 9.0–14.0% compared with RF, SVM, ANN, and GNB, respectively. The overall accuracy standard deviation of HLEL is 1.22%, proving that HLEL has low sensitivity to the size of the training data. The HLEL algorithm demonstrates exceptional precision when applied to diverse datasets. The overall accuracy of HLEL in Sentinel 2, Landsat 8, and GF-6 WFV increased by 3.1% to 22.7%, 3.1% to 39.0%, and 0.4% to 9.0%, respectively, when compared to the overallaccuracy of the basicclassifiers. The results demonstrate the feasibility of HLEL, an algorithm with high accuracy and data robustness, in applying wetland LULC classification and providing a new solution for wetland cover mapping.