Cloud Fraction Of Satellite Based On Convolutional Neural Networks

Satellite imagery interpretation is the precondition for meteorological services when using the satellite imagery. Cloud classification and cloud fraction are the core fields of meteorology satellite imagery interpretation. However, as the utilization rate of satellite optical parameters and cloud image characteristics is not high, so the cloud classification and cloud fraction which in cloud image interpretation are inaccurate. Moreover, the application of cloud fraction is not mature to the point of business. After the basis related home and abroad research, this paper presents a Convolutional Neural Networks method. The proposed method is used for satellite images’ cloud detection.In recent years, Deep learning, especially the Convolutional Neural Network (CNN), has showed strong adaptability and robustness in many application fields. CNN has good tolerance and parallel processing and self-leaning. Therefore, this paper proposes a CNN method for satellite imagery interpretation to solve the problem of satellite cloud fraction. At the same time, as the CNN detection is slow, this paper proposes a method for cloud detection based on ELM for further added. The main work includes the following aspects:Firstly, Original satellite imagery are extracted as the CNN training samples, and then study how to classify thick clouds, thin clouds,clear sky and the overlap which lies at the intersection of thick cloud and thin cloud through variable expressions of different visible channels. Finally, this paper uses the traditional threshold method, the dynamic threshold method and Extreme Learning Machine to do the contrast.This proposal calculates total pixel-level cloudiness by improving "spatial correlation method" based on cloud detection and improves the cloud classification algorithm and cloud amount computing algorithm through comparison to standard database results. This work will lay a solid theoretical foundation for full automatic observation of satellite cloud image.