The accurate detection of clouds in images is prerequisite for remote sensing image processing and applications. Traditional cloud detection methods rely on particular sensors, and the artificial neural network method only uses spectral or spatial information. In this paper, a novel method combining multiple features based on deep learning (MDL) for cloud detection is proposed. Deep neural network (DNN) and fully convolutional neural (FCN) network are applied to extract the spectral and spatial features of the images respectively, and the features are used for the input of another DNN for re-learning while the image data also serves as an input to the DNN. Finally, joint feature obtained by relearning is classified by Support Vector Machine (SVM). The method makes full use of the spectral-spatial information of the images to detect cloud comprehensively. A comparative experiment was carried on Landsat 8 images containing different types of clouds over various underlying surfaces. The results show that the MDL method performs favorably, which is significantly improved compared to the single neural network algorithm and the function of mask (FMask) algorithm.

A dust storm detection algorithm for NPP VIIRS data is proposed in this paper. The pixel dataset includes a variety of typical feature types, such as dust over different surface type, thick and thin clouds, vegetation, Gobi, ice/snow, etc. were collected and the distribution of the reflectance and brightness temperature were analyzed, based on which, a dust detection algorithm was generated. Multi-temporal NPP VIIRS images with dust storm happened were collected and applied to the experiments of dust storm detection with the proposed method. OMI AI products which can well describe the distribution of dust storm were selected for validation, and the results shows that this algorithm can detect the dust storm from NPP VIIRS over different land types in high precision.

A self-adaptive hybrid differential evolution with simulated annealing algorithm, termed SaDESA, is proposed. In the novel SaDESA, the choice of learning strategy and several critical control parameters are not required to be pre-specified. During evolution, the suitable learning strategy and parameters setting are gradually self-adapted according to the learning experience. The performance of the SaDESA is evaluated on the set of 25 benchmark functions provided by CEC2005 special session on real parameter optimization. Comparative study exposes the SaDESA algorithm as a competitive algorithm for a global optimization.