Method Of Intelligent Cloud Masking For Remote Sensing Image

Clouds cause distortion or loss of corresponding area information on optical remote sensing images.Cloud masks are a process of marking pixels affected by clouds and are the basis for the correct use of remote sensing data.Due to the diversity and large differences in the spectral and shape characteristics of clouds,and the uneven number distribution,cloud masks have become a difficult problem to be solved in remote sensing image preprocessing.In recent years,booming deep learning methods have provided new technical ideas for solving cloud mask problems.this paper considers the cloud mask as an example segmentation process of extracting the target of interest(cloud)from the complex background(the underlying surface composed of clean ground objects),and develops a cloud mask based on the improved Mask R-CNN Method,showing good application potential.However,Mask R-CNN has the following problems when applied to cloud masks:(1)There are many types of clouds,which not only have large differences in spectral characteristics and shape characteristics,but also have uneven number distribution.When the samples are marked in units of scenes,the problem of uneven sample distribution is caused.As a result,due to the insufficient number of training samples,the classifier cannot effectively capture its features,which leads to missed extraction.(2)During the feature extraction and masking process,the network performs a downsampling operation,plus the cloud and the underlying surface.There is no clear boundary between them,which leads to inaccurate mask boundaries of clouds(especially larger ones).this paper improves the original Mask R-CNN cloud mask algorithm and develops an intelligent cloud mask system.The main work and conclusions are as follows:(1)This paper proposes a sample equalization method to solve the problem of missed extraction of small categories due to uneven sample size.This method divides the samples into different training batches,then uses the first batch of samples to train,and performs a mask test on the second batch of samples,and then retrains the wrong or missed samples to the classifier.Then repeat the process for all batches of samples until stable results are obtained or all categories of samples have been used.(2)This paper proposes a boundary optimization method based on Mask R-CNN.Aiming at the problem of inaccurate boundary of the result when masking large clouds,a sliding window prediction method was proposed to reduce the boundary positioning error caused by RoI Align,realize the accurate extraction of the boundary of large clouds,and improve the accuracy of mask boundaries.Finally,this paper chooses two open source datasets and one self-labeled dataset for accuracy evaluation.The results show that compared with other deep learning-based cloud mask algorithms,the cloud mask obtained by this method has better recall,accuracy,and cross-ratio,F1 Score and overall accuracy have been improved,and can effectively adapt to the detection of different types of clouds,providing a new masking method.Compared with the original Mask R-CNN,the proposed method has greatly improved the miss detection rate and boundary positioning accuracy,which proves the effectiveness of the improved method.In future work,we will further study the method of this paper to expand the data set,enhance the training efficiency and robustness of the network,and build a practical cloud mask method.