Tradeoffs among multi-source remote sensing images, spatial resolution, and accuracy for the classification of wetland plant species and surface objects based on the MRS_DeepLabV3+ model

Classification of wetland plant species (PlatSpe) and surface objects (SurfObj) in remote sensing images faces significant challenges due to the high diversity of PlatSpe and the fragmented nature of SurfObj. Unmanned aerial vehicle (UAV) images and satellite images are the primary data sources for the classification of wetland PlatSpe and SurfObj. However, there is still insufficient research on the effect of various data sources and spatial resolutions on the classification results. This study essentially focuses on Huixian Wetland in Guilin, Guangxi, China through utilizing UAV images and satellite images with varying spatial resolutions as data sources. To this end, the MRS_DeepLabV3+ model is constructed based on multi-resolution segmentation and DeepLabV3+, and the wetland PlatSpe and SurfObj are appropriately classified based on this model. The obtained results reveal that: (1) MRS_DeepLabV3+ model with optimal scale parameter (SP) is capable of achieving higher classification accuracy compared to DeepLabV3+. The optimal SPs for both UAV images and satellite images gradually lessen with decreasing the spatial resolution, and satellite images require larger SPs compared to UAV images. (2) In both the UAV and satellite image models, both OA and kappa exhibit a decreasing trend with the reduction of the spatial resolution. (3) The overall classification accuracies of the satellite image models are superior to the UAV image models in the spatial resolution intervals of 2 to 16 m. This investigation can be regarded as a valuable reference for selecting data sources and spatial resolutions in the wetland PlatSpe and SurfObj classification.