Mangrove taxon mapping from Sentinel-1/2 using a compact one-dimensional network with probability calibration, predictive entropy, and frozen transfer evaluation

Mangrove extent maps lack the taxonomic information needed for carbon assessment, restoration evaluation, and management. Taxon mapping at 10 m remains constrained by spectral similarity, mixed pixels, limited reference coverage, and spatial dependence.

We developed the Inverted-Residual Spectral One-Dimensional Convolutional Neural Network (IR-Spec1D) as a compact classifier for a 14-feature Sentinel-1/2 optical and radar vector. To capture nonlinear optical-radar relationships compactly, inverted-residual and depthwise-separable transformations learn feature interactions, while Efficient Channel Attention emphasizes informative learned channels. Spectral-Domain Dropout reduces dependence on individual inputs through training-time feature masking. The workflow combines this configuration with spatial evaluation, post hoc temperature scaling, predictive-entropy analysis, and frozen transfer assessment. The six-category source taxonomy includes the B. sexangula/B. gymnorhiza complex. Six classifiers were evaluated using April 2024 observations from the selected Dongzhai Harbor study area.

IR-Spec1D achieved the highest mean overall accuracy (0.9720) and macro-F1 (0.9406) under five spatially grouped cross-validation folds. Performance remained comparatively high with a 100-m exclusion buffer. Held-out source-domain diagnostics showed that temperature scaling reduced mean expected calibration error from 0.0346 to 0.0125 without changing categorical predictions. Predictive entropy ranked cross-validated source-domain errors with an AUROC of 0.8808. The frozen ensemble retained an overall accuracy of 0.9450 in both same-site temporal evaluations, whereas four-category closed-set geographic transfer to Gaoqiao yielded 0.7118.

The combined evidence characterizes taxon discrimination, held-out probability behavior, and performance retention under specified transfer conditions within a common mangrove-mapping workflow.