Research Of Cloud And Snow Discrimination From Multispectral High-Resolution Satellite Images

Cloud and snow discrimination is a challenging task in cloud extraction,which is important for satellite image processing and applications.For most high-spatial-resolution satellites,their sensor bands only include visible and near-infrared bands,which makes traditional methods of cloud and snow detection unpractical.However,there are also shortages and deficiencies in recent research on high-spatial-resolution satellite images,most of which only concentrate on panchromatic images.Some of these research have special requirements for satellite capabilities.In this study,a new method for cloud and snow discrimination,which combines spectral-based methods with machine learning methods,is proposed.This method is based on spectral,texture,and shape features and support vector machine multi-classification strategy is applied.A new feature called curvature histogram is designed to describe edge shape.The method begins with the region of interest test to extract cloud and snow area together from the images.Then,the extracted area is combined with the segmentation results using mean shift algorithm to obtain the object of interest while calculating the feature values of the obtained superpixels.The complexity of surfaces in the cloud and snow area is classified into four types,namely,thick cloud,thin cloud,snow,and snow land,such that six kinds of classifiers are obtained by designing a classifier between every two categories.Through these six classifiers and calculating the sum of confidence coefficients for each category,every superpixel is classified into the class with the highest confidence coefficient and a rough cloud and snow mask is obtained Finally,the GrabCut algorithm is applied to optimize the classification results at the pixel level.The experiments were conducted to test the overall framework and every single stage.The test results,which are based on the multispectral images of China's GF-1 satellite,indicate that the proposed method is effective for cloud and snow discrimination on multispectral images of high-resolution satellites.Future work aims at utilizing contextual information and applying some probabilistic models in the classification to reduce commission errors.