A novel feature selection criterion for wetland mapping using GF-3 and Sentinel-2 Data
Jinqi Zhao, Zixuan Wang, Qingjie Zhang, Yufen Niu, Zhong Lu, Zheng Zhao
Abstract
Accurate land-cover information extracted from multi-source satellite data on the spatial distribution of coastal wetlands is important for wetland mapping, while also being crucial for effective wetland management and conservation measures. On the one hand, the use of more features allows for better expression of the wetland land-cover types. On the other hand, more features easily lead to data redundancy and complex computation. Feature selection is a compromise between optimal features and computational efficiency, and is a suitable approach for wetland mapping. However, the exiting methods on feature selection are limited by possible information loss, overfitting, and high complexity. To solve these problems, a novel feature selection method for wetland mapping using GF-3 and Sentinel-2 data is proposed in this paper. Firstly, to solve the information loss and risk of overfitting, the LDJ (Laplacian score [ L S], distance correlation [ D C], Jeffries-Matusita [ J M] distance) algorithm is proposed to achieve the fusion of different criteria with adaptive weighting factors, to obtain the preliminary features. Moreover, recursive feature elimination based on cross-validation (RFECV) is used to select more critical features for the classifier through the process of recursively removing features . In order to verify the effectiveness of the proposed LDJ _RFECV algorithm, random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) machine learning classifiers were utilized for the wetland mapping. The proposed LDJ_RFECV algorithm exhibits the capability of attaining an outstanding performance with reduced feature dimensions. In addition, the proposed approach decreases the processing time by 40% when compared to the original RFECV algorithm.