Boosted Random Forests Approach for Spectral-Based Supervised Land Cover Mapping in Remote Sense Data

Abstract

In recent decades, numerous supervised machine learning algorithms have emerged to address remote sensing (RS) imagery mapping challenges. Popular choices include Random Forest (RF) and Support Vector Machines (SVMs), which have proven effective in RS applications. This study introduces the Boosted Random Forests Approach for Spectral-Based Supervised Land Cover Mapping in RS Data (BRFA) algorithm for RS image classification, combining feature selection and boosting to enhance RF classifier generalization. Utilizing a Forward Greedy approach with Fuzzy Preference Rough Set (FPRS) for feature selection and integrating boosting into RF, BRFA employs an out-of-bag (OOB) error estimate to adjust sample weights and counter overfitting. Applied to two labeled RS datasets, BRFA demonstrates superior performance in mapping land cover types, surpassing RF, C4.5, SVMs, and k-NN. BRFA yields kappa indices of 0.89 and 0.93, with accuracy measurements of 92.27% and 94.94% for SPOT and IRS datasets, respectively. These results highlight BRFA's consistent superiority, showcasing its potential for accurate RS image identification.