Exploring the effects of different combination ratios of multi-source remote sensing images on mangrove communities classification

Mangroves are one of the most important marine ecosystems globally, their spatial distribution is crucial for promoting mangrove ecosystems conservation, restoration, and sustainable managements. This study proposed a novel Unet-Multi-Scale High-Resolution Vision Transformer (UHRViT) model for classifying mangrove species using unmanned aerial vehicle (UAV-RGB), UAV-LiDAR, and Gaofen-3 Synthetic Aperture Radar (GF-3 SAR) images. The UHRViT utilized a multi-scale high-resolution visual Transformer as its backbone network and was designed to a multi-branch U-shaped network structure to extract features of different scales layer by layer, and to facilitate the interaction of high and low-level semantic information. We further verified the classification performance superiority of UHRViT model by comparing to HRViT and HRNetV2 algorithms. We also systematically investigated the effects of active–passive image combination ratios on mangrove communities mapping. The results revealed that: UAV-RGB images exhibited the better classification accuracy (mean F1-score＞95 %) for mangrove species than UAV-LiDAR and GF-3 SAR images; The classification performances and stability of UHRViT algorithm in the fifteen datasets outperformed the HRViT and HRNetV2 algorithms; Combining UAV-RGB with GF-3 SAR or UAV-LiDAR images respectively, both achieved better classifications than the single data source. Based on the UHRViT algorithm, the combination of UAV-RGB and UAV-LiDAR achieved the highest classification accuracy (Iou = 0.944, MIou = 50.2 %) for Avicennia corniculatum (AC). When the combination ratio of UAV-RGB with GF-3 SAR or UAV-LiDAR was 3:1, Avicennia marina and AC both obtained the optimal classification accuracy with average F1-scores of 98.19 % and 97.3 %, respectively. Our works revealed that the changes in the classification accuracies of mangrove communities under multi-sensor image combination ratios, and demonstrated that our model could effectively improve the classification accuracy of mangrove communities.