Large-scale extraction of check dams and silted fields on the Chinese Loess Plateau using ensemble learning methods

Abstract

Check dams are an effective soil and water conservation measure which is widely constructed in the world. A large number of check dams have been installed on the Chinese Loess Plateau during the last 70 years, which have played an important role in sediment control and soil conservation. However, high-resolution and large-scale mapping of the location of the check dam and silted field is absent, thus hampering the efficient management of the check dams and quantitative evaluation of their eco-environmental benefits. In this study, we present a methodological framework to extract silted fields and estimate the location of the check dam at a pixel level by using three ensemble learning methods combined with a novel resampling procedure to lessen the influence of class imbalance. Our results indicate that the distribution of check dams and silted fields is a typical imbalanced binary classification problem that the amount of silted field samples only accounted for 4% of the total randomly collected samples, which has a significant impact on the accuracy of both model training and validation. By using the random under-sampling method, the optimal imbalance ratio was determined for each model, combined with the optimal modeling parameters and 23 features, to train and validate the model. The validation results on the testing set show that the F1-score of Random Forest, Extreme Gradient Boosting, and EasyEnsemble model for the silted field is 0.7501, 0.7664, and 0.7754, respectively. The feature importance analysis shows that three macro-terrain features and multi-temporal spectral indices contributed mostly to the accurate extraction of silted fields, among which the multi-temporal vegetation cover change index has the highest feature importance for all models. Applying the tuned model to the whole Wuding River catchment, we produced a 10m-resolution silted fields and check dams map showing that there are ca. 10,500 check dams correspondingly forming ca.283.3 km² area of silted fields. This study provides an important and efficient methodological framework for quick mapping of check dams and silted fields at a high-resolution and a large scale with addressing the imbalanced classification problems.