Machine learning-based monitoring of land cover and reclamation plantations on coal-mined landscape using Sentinel 2 data

Abstract

The rehabilitation of degraded coal-mined landscapes has recieved significant global attention due to its critical impact on ecological integrity, economic prosperity, and social development, aiming for zero net land degradation. This study examines the reclamation of coal mine overburdens through reforestation, using high-resolution Sentinel 2 satellite data classified by various Machine Learning (ML) algorithms. Support Vector Machine has been identified as a more accurate and effective ML algorithm compared to Random Forest and Maximum Likelihood Classifier in delineating land use and vegetation classes, particularly forests, and in distinguishing reclamation plantations into three age classes: young (4 ± 3 years), middle-aged (10 ± 2 years), and mature (15 ± 2 years). Significant areas of forests and agricultural land have been lost to coal mining, while a large portion of the overburden has been regenerated with plantations, leaving a small area barren for future mine expansion. The total standing biomass and carbon stock varied significantly (p ≤ 0.05) and increased with the age of reclamation plantations, ranging from 10.5 to 23.7 Mg ha-1 and 4.7–10.9 Mg ha-1, respectively. However, the biomass and carbon stocks in mature stands of mined sites were nearly three times lower than those in natural forests. The recovery rates of soil nutrients under plantations of these sites have surpassed halfway and may take a decade or two to reach levels equivalent to those of natural forests. By integrating crucial eco-technological and geospatial approaches employing ML algorithms, we effectively navigate interventions to reinvigorate the restoration process and reverse land degradation.