Challenges of Retrieving LULC Information in Rural-Forest Mosaic Landscapes Using Random Forest Technique

Abstract

Land use and land cover (LULC) information plays a crucial role in determining the trend of the global carbon cycle in various fields, such as urban land planning, agriculture, rural management, and sustainable development, and serves as an up-to-date indicator of forest changes. Accurate and reliable LULC information is needed to address the detailed changes in conservation-based and development-based classes. This study integrates Sentinel-2 multispectral surface reflectance and vegetation indices, and lidar-based canopy height and slope to generate a random forest model for 3-level LULC classification. The challenges for LULC classification by RF approach are discussed by comparing it with the SVM model. To summarize, the RF model achieved an overall accuracy (OA) of 0.79 and a macro F1-score of 0.72 for the Level-III classification. In contrast, the SVM model outperformed the RF model by 0.04 and 0.09 in OA and macro F1-score, respectively. The accuracy difference increased to 0.89 vs. 0.96 for OA and 0.79 vs. 0.91 for macro F1-score for the Level-I classification. The mapping reliability of the RF model for different classes with nearly identical features was challenging with regard to precision and recall measures which are both inconsistent in the RF model. Therefore, further research is needed to close the knowledge gap associated with reliable and high thematic LULC mapping using the RF classifier.