Abstract:

Remote sensing inversion is one of the effective ways to monitor the distribution of suspended sediment concentration (SSC) in water surface. Nevertheless，the present empirical regression models are based on assumption of large samples, while the samples number are actually small in most cases, so it is necessary to introduce new retrieval models which could be based on small sample number. Support Vector Machine (SVM) is established on the theories of Structural Risk Minimization Principle and VC Dimension, it has better generalization ability and is suitable for small-sample regression model. In this paper, one SVM remote sensing retrieval model for SSC has been established with HJ1B CCD2 satellite images and synchronous in situ SSC data in the middle reaches of Yangtze River, and its parameters was optimized by using Fruitfly Optimization Algorithm (FOA). The results showed that, SVM model has higher prediction accuracy and stability than the traditional empirical regression models, and in the parameter optimization of SVM model, FOA algorithm has satisfactory results and much smaller calculation amount than Grid Search Algorithm. Finally, the already trained SVM model was used to retrieve SSC in the water of Yangtze River and Dongting Lake around Chenglingji, and the spatial distribution features of SSC was analyzed. The results showed that, SSC is significantly smaller in the mainstream of Yangtze River than in Dongting Lake. This is mainly caused by significant decrease in sediment discharge because of Three Gorges Project. As a result, one blending zone can be distinguished in the Yangtze River between Chenglingji and Honghu city when the muddy waters of Dongting Lake flows into the river. In addition, the SSC is significantly higher in the outlet channel than in the other area in Dongting Lake, which may be caused by the strong disturbance of the sand mining activity in the region.