Tree Species Classification of Airborne Hyperspectral Image in Cloud Shadow Area

Abstract

Tree species classification is crucial for forest management. Airborne hyperspectral data provide the image with high spatial and spectral resolution, which has a possibility for forest species classification. The radiance of airborne hyperspectral image was decreased when the cloud exists. The radiance of object in the cloud shadow area is much lower than the non-shaded. And the spectral variance of different objects in the shadow is also lower than that of the non-shaded region. Although the radiance of the image in the cloud shadow is decreased, there is still a certain difference among different objects, which provides the possibility for forest species classification. We used vegetation indices and texture features to recombine a new image. Reflectance images and the recombination images were classified by support vector machine (SVM) classifier. Tree species classification is crucial for forest management. Airborne hyperspectral data provide the image with high spatial and spectral resolution, which bring possibility to forest species classification. The radiance of object in the cloud shadow area of airborne hyperspectral image is much lower than the non-shaded area. And the spectral variance of different objects under the shadow is also lower than that of the non-shaded region. Although the radiance of the image in the cloud shadow is decreased, there is still a certain difference among different objects, which provides the possibility for forest species classification. We used vegetation indices and texture features to recombine a new image. The narrow band vegetation indices include red edge normalized vegetation index (NDVI705), Improved red edge ratio vegetation index (mSR705), Improved red edge normalized vegetation index (mNDVI705), Vogelmann1 (VOG1), Vogelmann2 (VOG2) and Red Edge Position Index. The bands of 31 (0.67 μm), 51 (0.86 μm) and 55 (0.89 μm) were used to calculate the texture information, which were selected using the optimum index factor (OIF). Tree species training samples was selected based on high resolution aerial photographs. The support vector machine (SVM) method was used to classify the reflectivity images and the feature images after recombination. The classification results were verified by filed data, and the overall accuracy and Kappa coefficient were used as the evaluation indices for classification accuracy. Compared with the classification result of reflectance image, the combination of vegetation index and texture information improve classification accuracy significantly. The overall accuracy and Kappa coefficient are 90.4% and 0.88, which increased 18% and 0.2 respectively. The classification accuracy of each tree species is also significantly improved. It can be seen from the confusion matrix that when using the reflectance image for classification, the Pinus koraiensis was divided into Pinus sylvestris by mistake. However, using vegetation index, the pixel number of Pinus koraiensis divided into Pinus sylvestris by mistake is greatly reduced. We concluded that the cloud-shaded forests can be classified based on the narrow band vegetation indices (NDVI705, mSR705, mNDVI705, VOG1, VOG2, REP) and texture information, which is better than the reflectance image only.