Mapping mangrove functional traits from Sentinel-2 imagery based on hybrid models coupled with active learning strategies

Abstract

Accurately quantifying functional traits across large scales is considered fundamental for the management and conservation of existing mangrove ecosystems. In recent years, hybrid models, which combine radiative transfer model simulations with machine learning regression algorithms (MLRA), have been effectively employed in satellite-based estimations of plant functional traits across diverse ecosystems. Nevertheless, the inevitable data redundancy stemming from heavy-parameterization radiative transfer models restricts the application of the hybrid model. Previous studies have indicated that active learning (AL) strategies can mitigate this redundancy through smart sampling selection criteria. While many studies have attempted to investigate mangrove functional traits using various models, there is limited understanding of the performance of hybrid models coupled with active learning strategies in retrieving the traits. In recent years, Sentinel-2 has become mainstream for retrieving detailed and reliable information across diverse ecosystems. The aim of this study is to utilize a retrieval scheme to extract four mangrove functional traits from Sentinel-2 imagery: leaf area index (LAI), leaf chlorophyll content (Cab), leaf dry matter content (Cm), and leaf equivalent water thickness (Cw). In order to achieve this goal, we systematically evaluated 36 different MLRA-AL models, which were combinations of six MLRAs and six AL strategies. Retrieval results showed that GPR (Gaussian processes regression)-ABD (angle-based diversity) and GPR-PAL (variance-based pool of regressors) yielded the highest accuracies for LAI (R2 = 0.68, NRMSE = 10.488 %) and Cw (R2 = 0.47, NRMSE = 13.868 %), respectively. GPR-EBD (Euclidean distance-based diversity) had the highest accuracies of Cm (R2 = 0.54, NRMSE = 11.695 %) and Cab (R2 = 0.71, NRMSE = 13.764 %). The retrieval models were subsequently applied to produce distribution pattern maps of four mangrove functional traits within a Ramsar site. This study represents the first attempt to utilize AL strategies to enhance the efficiency of traditional hybrid models and map multiple functional traits of mangrove forests. The retrieval scheme and mapping results could significantly contribute to the management of mangrove ecosystems and provide a fundamental data source for future research on the ecological services of mangroves.

Keywords

• Mangrove functional traits
• Active learning
• Machine learning regression algorithm
• Radiative transfer model
• Sentinel-2