Identifying mangroves through knowledge extracted from trained random forest models: An interpretable mangrove mapping approach (IMMA)

Black-box algorithms are among the dominant mangrove mapping approaches with complex decision-making procedures. Model internals and tacit knowledge were neglected, such as a large number of decision rules provided by random forest (RF) analyses. Explainable artificial intelligence (XAI) has emerged to emphasize the interpretability of an approach. However, current knowledge-based mangrove mapping approaches rely on extensive experiments. Thus, they cannot be easily updated to accommodate new issues, such as prevalent false positives resulting from insufficient consideration of the spectral mixture of vegetation and water in existing studies. To combine the advantages of black-box-based approaches with high update rates and knowledge-based approaches with high interpretability, this study developed a knowledge extraction method by parsing trained RF models, reconstructing decision rules to incorporate the ensemble procedure, and selecting the optimal decision rule as the target. Using this method, an interpretable mangrove mapping approach (IMMA) consisting of five features was constructed, which derived from Sentinel-2 image bands and a digital elevation model: B12 < 0.06 & B8/B2 > 3.50 & elevation < 4.70 & mangrove vegetation index (MVI) > 2.92 & normalized difference index4 (NDI) < 0.07. The study achieved an overall accuracy (OA) of 82.3% along the entire coast of China using test samples. Comparatively, it achieved an OA of 78.8% in south Florida, with no training samples for the RF models. The IMMA approach had a limited number of false positives compared with the black-box-based and knowledge-based approaches. By analyzing, we found B12 < 0.06 & B8/B2 > 3.50 & elevation < 4.70 dominated the IMMA to achieve comparable classification results to the existing studies, and B8/B2 > 3.50 was the key to suppressing the false positives resulting from the spectral mixture. The IMMA provided a bridge between training samples and interpretable decision rules, a tool to discover new knowledge, a key to improving fundamental scientific understanding of mangrove mapping, and an alternative to black-box algorithms in the XAI era expandable to various fields.