In this study, we compared four different water extraction methods with GF-1 data according to different water types in Tibet, including Support Vector Machine (SVM), Principal Component Analysis (PCA), Decision Tree Classifier based on False Normalized Difference Water Index (FNDWI-DTC), and PCA-SVM. The results show that all of the four methods can extract large area water body, but only SVM and PCA-SVM can obtain satisfying extraction results for small size water body. The methods were evaluated by both overall accuracy (OAA) and Kappa coefficient (KC). The OAA of PCA-SVM, SVM, FNDWI-DTC, PCA are 96.68%, 94.23%, 93.99%, 93.01%, and the KCs are 0.9308, 0.8995, 0.8962, 0.8842, respectively, in consistent with visual inspection. In summary, SVM is better for narrow rivers extraction and PCA-SVM is suitable for water extraction of various types. As for dark blue lakes, the methods using PCA can extract more quickly and accurately.
Remote sensing inversion of heavy metal in vegetation leaves is generally based on the physiological characteristics of vegetation spectrum under heavy metal stress, and empirical models with vegetation indices are established to inverse the heavy metal content of vegetation leaves. However, the research of inversion of heavy metal content in vegetation-covered soil is still rare. In this study, Pulang is chosen as study area. The regression model of a typical heavy metal element, copper (Cu), is established with vegetation indices. We mainly investigate the inversion accuracies of Cu element in vegetation-covered soil by different vegetation indices according to specific spectral resolutions of ASD (Analytical Spectral Device) and Hyperion data. The inversion results of soil copper content in the vegetation-covered area shows a good accuracy, and the vegetation indices under ASD spectral resolution correspond to better results.
With the deterioration of ecological environment, rare plants on the earth are decreasing rapidly, so there is an urgent
need for the study on sophisticated vegetation classification. Hyperspectral data have great potential in sophisticated
classification. FISS(Field Imaging Spectrometer System) is a newly developed system, and pixels of FISS could be
considered as pure pixels with high spatial and spectral resolution, which makes FISS a perfect option on the study of
methodology. This study aims to evaluate different methods based on FISS data and find out the best one of
sophisticated vegetation classification. The methods are as follows: Maximum Likelihood (ML), Spectral Angle
Mapping (SAM), Artificial Neural Net (ANN), Support Vector Machine (SVM) and Composite Kernel Support Vector
Machine (C-SVM). Firstly, segmented principal components transformation is adopted for spectral dimensionality
reduction, and all bands are divided into 2 subsets according to the correlation matrix. Secondly, 16 principal
components are kept. After that, 5 methods mentioned above are tested. The Overall Accuracy and Kappa coefficient of
C-SVM, SVM and ANN are higher than 90%, and C-SVM obtains the highest accuracy, which is consistent with visual
interpretation. The result shows that C-SVM, SVM and ANN are more suitable for sophisticated vegetation classification
of hyperspectral data, and C-SVM is the best option.