Comparison of random forest, gradient tree boosting, and classification and regression trees for mangrove cover change monitoring using Landsat imagery

Abstract

Ineffective land use in coastal areas negatively impacts the environment and destroys mangrove ecosystems, contributing to increasing greenhouse gas emissions and decreasing carbon sequestration. This study aimed to monitor the land use changes in mangrove areas with Landsat data using several machine learning (ML) methods. According to the random forest (RF), gradient tree boosting (GTB), and classification and regression trees algorithms (CART), the mangrove area exhibited significant fluctuations over the study period, with the largest expansion observed from 1999 to 2008 (4,240.57 ha), followed by a slight increase in 2023 (368.36 ha from 2019). Accuracy assessment revealed distinct performance levels across the models. The RF algorithm demonstrated the highest overall accuracy (OA) of 98.8 %, with kappa values ranging from 0.96 to 0.98, indicating high consistency and reliable predictions over time. The CART algorithm, while accurate, showed more variability, especially between 1991 and 1994, with an OA ranging from 85.3 % to 92.5 % and kappa values between 0.92 and 0.96. The GTB algorithm had moderate performance, with OA values between 85.6 % and 95.7 % and kappa values ranging from 0.92 to 0.96, suggesting reliable results but with some inconsistency compared to RF. The RF algorithm’s superior OA and consistency make it the most suitable long-term land cover monitoring method. Future studies can benefit from incorporating RF in assessing ecosystem changes, including carbon sequestration potential in mangrove forests.