Examination of the performance of ASEL and MPViT algorithms for classifying mangrove species of multiple natural reserves of Beibu Gulf, south China

Abstract

Mangrove is the highest productive ecosystems in the global coastal zone, which has high blue carbon sink function and carbon neutrality potential. Fine species classification is essential for mangrove conservation and sustainable development, and has attracted much attention in recent years using ensemble learning and multi-dimensional data. However, the current mangrove species classification based on traditional stacking ensemble learning still faces challenges due to the correlation between base classifiers, differences in meta-classifier capabilities, the subjectivity of parameter tuning, and data redundancy. To address these issues, this paper utilized unmanned aerial vehicle (UAV) multispectral images of three mangrove nature reserves in Beibu Gulf, south China, to examine the classification and generalization ability of our proposed Adaptive Stacking Ensemble Learning (ASEL) algorithm for different mangrove species. We also aim to verify the feasibility of the Multi-Path Vision Transformer for Dense Prediction (MPViT) algorithm for mangrove species mapping, and compare its performance with the ASEL algorithm for mangrove species classification. Finally, we used the SHapley Additive Explanations (SHAP) method to measure the contribution of feature variables to the model, exploring the sensitivity of different image features to mangrove species mapping. This study highlights that: (1) The two ASEL algorithms achieved high accuracy classification of mangrove species with the overall classification accuracy ranging from 79.8% to 96.2%. The ACE-Stacking and AOM-Stacking algorithms performed better classification ability than the traditional stacking algorithm, with the mean overall classification accuracy increasing from 0.9% to 3.3%. The McNemar test further indicated that the differences in classification results derived from three algorithms were significant at the 95% confidence interval, demonstrating the better classification and generalization ability of the ASEL algorithm. (2) The MPViT algorithm achieved better classification accuracy in the three reserves, with an overall accuracy of 95.5%-97.3%, which was 0.6%-5.4% higher than the two ASEL algorithms. The average accuracy of identifying mangrove species was over 95.3% in all reserves, demonstrating the desirable mangrove classification performance of MPViT algorithm. (3) The identification accuracies (F1 scores) of mangrove species in different reserves were ranged from 0.848 to 0.984. Cyperus malaccensi had the highest identification accuracy. (4) The SHAP method interpreted the great contribution of digital surface model and variable atmospherically resistant index features on mangrove species classification. The red band and ratio vegetation index was sensitive to Aegiceras corniculatum and to Cyperus malaccensi.