Comparison of RFE-DL and stacking ensemble learning algorithms for classifying mangrove species on UAV multispectral images

Mangroves are highly productive wetland ecosystems, located at the interlocking area of tropical and subtropical coastal zones. Accurately mapping the distribution, quality and quantity of species are crucial for mangrove management, protection, and restoration. This study proposed a mangrove species mapping approach by combining recursive feature elimination (RFE) with deep learning (DL) algorithms, and further assess the effectiveness of feature selection for DL (DeeplabV3+ and PSPNet) algorithm to improve classification accuracy under the high dimensional UAV image datasets. We constructed an ensemble learning models (SEL) by stacking five base models (Random Forest, XGBoost, LightGBM, CatBoost, and AdaBoost), and evaluate the classification ability of mangrove species between SEL and RFE-DL algorithms. Comparison of the classifications of mangrove species was to evaluate the accuracy differences between SEL and base models. Results indicated that: (1) RFE algorithm could improve the classification accuracy of DL algorithms. RFE-DL models using the optimal features achieved 94.8% of overall accuracy (OA), which was 0.2%-8.5% higher than only using the original multispectral bands; (2) SEL algorithm produced better classification performance than RFE-DL with a higher 1.6%-12.7% of overall accuracy. Mcnemar's test showed the classifications of mangrove species were significant differences between the three algorithms; (3) the SEL algorithm had a strong and stable ability for classifying mangrove species. The OA of six classification scenarios was from 75.5% to 92.2%, and the highest OA using SEL algorithm was 0.8%-4.2% higher than base models; (4) XGBoost algorithm had the highest importance, while AdaBoost had the lowest importance in the SEL-based classifications.