Advancing Inland Water Body Mapping with Self-Supervised Machine Learning

Abstract

Traditional methods of mapping inland water bodies involve labor-intensive and manual data labeling, limiting their scalability to a larger extent. This study introduces a novel approach: self-supervised machine learning (SSML) for mapping inland water bodies. SSML is a training method where a model learns from data without the need for explicit human-labeled annotations. Using this technique, the study mapped inland water bodies in Pudukkottai, India, using LANDSAT-8 imagery from 2021. The training data for SSML were derived from two spectral indices: the Normalized Difference Vegetation Index and the Modified Normalized Difference Water Index. These indices were used to establish a threshold for automatically generating pseudo labels for two categories: water and non-water. This pseudo-labeled dataset was then utilized to train various machine learning models, including random forest, support vector machine, classification and regression tree, and gradient boosting. The accuracy of the final classified map was assessed using a spatial agreement test, which measures the degree of agreement of the classified map in relation to a reference dataset. The spatial agreement test used the Joint Research Commission (JRC) water map of 2021 as the reference dataset. The final inland water body map, derived from the SSML approach, demonstrated a high spatial agreement of 91% with the JRC water map. Among the SSML models, the random forest model outperformed others due to its ensemble nature. Compared to traditionally supervised classifiers (trained with 137 water points and 74 non-water points), the SSML models exhibited superior performance with a spatial agreement of 91%, significantly higher than the 67% achieved by the supervised model. This study is the first to demonstrate the application of SSML for mapping inland water bodies, offering an efficient and cost-effective alternative to traditional manual labeling. This approach holds significant potential for advancing remote sensing applications, particularly in regions where obtaining ground truth labeling is costly or impractical.