| Mangrove is one of the most productive ecosystems flourished in the intertidal zone of tropical and subtropical regions. Hong Kong has ten true mangrove species covering an approximate area of 350 hectares. Mai Po locating in the northwestern part of Hong Kong nourishes the largest mangrove stand and it was listed as a Wetland of Importance under the Ramsar Convention in 1995. Over the years, areas of mangrove have been shrinking globally due to development, pollution, and other unsustainable exploitation and Hong Kong was no exception. In Hong Kong, mangroves are usually sacrificed for urban development and infrastructure construction. Therefore, it is crucial to monitor their growth conditions, change of extent and possible unsustainable practices threatening their existence. Remote sensing being a cost-effective and timely tool for vegetation conservation is most suitable for such purpose.;Taking Mai Po as study area, this study acquired satellite-borne hyperspectral and radar data supplemented with in situ field survey to achieve three purposes. First, features from the remotely-sensed data that are significant to species discrimination were identified through pattern recognition. Second, selected features grouped into different subsets were used to delineate the boundary of mangrove species through supervised classification. In the meantime, classifiers including maximum likelihood (ML), decision tree C5.0 (DT), artificial neural network (ANN) and support vector machines (SVM) were tested for their accuracy performance. The third purpose is to understand the current biophysical condition of mangrove through leaf area index (LAI) modeling by regressing field-measured LAI against vegetation indices, backscatter and textural measures.;Results from feature selection revealed that hyperspectral narrowbands locating in green at 570nm, 580nm, 591nm, 601nm; red at 702nm; red-edge at 713nm; near infrared at 764nm and 774nm and shortwave infrared at 1276nm, 1316nm and 1629nm as well as the multi-temporal filtered backscatter captured in different seasons have high sensitivity to species difference.;Species-based classification using multi-temporal backscatter features alone do not provide a satisfactory accuracy. Comparatively, results from pure spectral bands have better overall accuracy than that from combining spectral and radar features. However, radar backscatter does improve accuracy of some species. Besides, all classifiers had similar variations of training accuracy under the same feature subset. However, the testing accuracy is much lower with the exception of ANN. Performance of ANN was more stable and robust than other classifiers while serious overtraining occurs for the DT classifier. Moreover, most species were mapped accurately as revealed by the producer's and user's accuracy with the exception of A. corniculatum and Sonneratia spp. due to deficiency of training samples.;Simple linear regression model with VIs revealed that triangular vegetation index (TVI) and modified chlorophyll absorption ratio index 1 (MCARI1) had the best relationship with LAI. However, weak relationship was found between field-measured LAI and radar parameters suggesting that radar parameters cannot be used as single predictor for LAI. Results from stepwise multiple regression suggested that TVI combined with GLCM-derived angular second moment (ASM) can reduce the estimation error of LAI. To conclude, the study has demonstrated spectral and radar data are complementarity for accurate species discrimination and LAI mapping. |
