Quantifying scattering characteristics of mangrove species from Optuna-based optimal machine learning classification using multi-scale feature selection and SAR image time series

Abstract

Mangroves play a significant role in carbon sequestration and storage. Mapping mangrove species and monitoring their conditions have been a crucial issue for achieving sustainable development goals. Currently combing multidimensional optical and SAR images with machine learning have become an important approach for mangrove species classification, but there are still some challenges in feature selection and hyperparameter optimizations. In this study, we proposed a novel classification framework by combing multi-scale variable selection algorithm (MUVR) with state-of-the-art machine learning hyperparameter optimization method (Optuna) for mapping mangrove species in the Beilun Estuary and Maowei Sea nature reserves using optical and dual-polarization SAR images, and further quantified the scattering characteristics of mangrove species using SAR image time series. We found that: (1) The MUVR algorithm could determine the optimal scale features for different scenarios and mangrove species, and improve the classification performance of machine learning with an overall accuracy (OA) improvement of 12.85%; (2) The Optuna-based optimal CatBoost outperforms LightGBM and NGBoost algorithms in mapping mangrove species, which achieved the highest OA (93.18%). This study demonstrated that LightGBM was suitable for identifying Aegiceras corniculatum, while the CatBoost algorithm was suitable for discriminating Avicennia marina, Bruguiera gymnorrhiza, Cyperus malaccensis, Kandelia candel and Sonneratia apetala; (3) SAR images and its derivatives improved identification ability of mangrove species, and collaboration of multispectral images and SAR-derived features produced the better classification; (4) From 2018 to 2020, the backscattering coefficients of mangrove species in VV and VH polarization focused on 0.053–0.327 and 0.015–0.062, respectively. The coherence coefficients of mangroves displayed a seasonal change trend with the large variations in summer and small variations in winter. The range of Entropy and Alpha of mangrove species was from 0.65 to 0.88 and 17–33, which indicated that the main scattering mechanism of mangroves was moderate random surface scattering.

Keywords

• Mangrove species classification
• Machine learning
• MUVR variable selection
• Optuna hyperparameter optimization
• Scattering characteristics evaluation
• Dual-polarization SAR images