| The mangrove forest is one of the most productive ecosystems on Earth,and the fine classification of mangrove communities is crucial for their conservation.In this study,the coastal mangrove wetlands of Beibu Gulf in Guangxi,China,were selected as the study area.A multidimensional image dataset was constructed based on UAV multispectral images to evaluate the performance of the MCCUNet algorithm compared to the Deep Lab V3+ and HRNet algorithms in classifying mangrove communities.Additionally,three transfer learning strategies were proposed,and the changes in classification accuracy of different mangrove community classes under these strategies were statistically analyzed.Furthermore,three multidimensional image datasets were constructed using UAV-RGB,UAV-Li DAR,and GF-3 dual-polarization SAR images to assess the differences in classification performance of UHRVi T,HRVi T,and HRNet V2 algorithms in mangrove community classification.The superiority of the proposed UHRVi T algorithm was verified.The UAV-RGB images were combined with UAV-Li DAR and GF-3 dual-polarization SAR images separately to create different combinations of passive-active image datasets.The changes in classification accuracy of mangrove communities under these combination datasets were statistically analyzed to evaluate the differences in the recognition capabilities of different image combination methods for mangrove communities.Finally,based on the passive-active image combination dataset,the performance of the UHRVi T algorithm was evaluated and compared with the HRNet V2 and HRVi T algorithms in mangrove community classification,confirming the superiority of the UHRVi T algorithm.The results of the study showed that:(1)The MCCUNet algorithm outperformed the Deep Lab V3+ algorithm in terms of mangrove community classification performance,achieving the highest overall accuracy(OA)of 97.24%,using the multidimensional image dataset based on UAV multispectral images;(2)Compared to the F-TL strategy,the Ft-TL strategy yielded better performance in terms of classification accuracy and stability.Under the Ft-TL strategy,the MCCUNet algorithm showed the highest improvement in F1-score for the SA,reaching a gain of 19.56%.The Sa P-TL strategy achieved better classification accuracy for mangrove communities across different time periods and sensors.In particular,under the Sa P-TL strategy,the MCCUNet algorithm demonstrated the highest improvement in F1-score for AC,reaching a gain of 19.85%;(3)The UHRVi T algorithm exhibited superior classification performance for mangrove communities compared to the HRVi T and HRNet V2 algorithms,achieving the highest OA and Kappa coefficient of 96.81% and 0.9638,respectively,using the dataset constructed based on UAV-RGB and UAV-Li DAR images.(4)Compared to the S&G combination method,the S&L combination method demonstrated better recognition capabilities for mangrove communities.The S&L combination method showed the highest improvement in average F1-score for the specific attribute KC,one of the mangrove vegetation types,with an increase of 1.8%;(5)The UHRVi T algorithm achieved the highest OA and Kappa coefficient,reaching 97.54% and 0.9722,respectively,using the passive-active image combination dataset. |
