Training DCNN For Cloud And Cloud Shadow Detection In Landsat Image Using QA Band

Landsat data is the most widely used remote sensing dataset currently due to its long record,global coverage,and moderate resolution;its application fields include resource exploration,agricultural analysis,surface change monitoring etc.Since the clouds and cloud shadows on Landsat images obscure the ground objects,making remote sensing of those ground objects unrealized.Therefore,accurate cloud and cloud shadow labeling is an important part in the production of Landsat images.With the development of deep learning technology and the improvement of computer performance,the cloud and cloud shadow detection method based on deep convolutional neural network has greatly improved the accuracy.However,the deep convolutional neural network model requires a large number of labeled images(each pixel on the images is manually labeled as cloud,clear or cloud shadow accurately).Manual labeling is costly and time-consuming,which is not conducive when collecting a large remote sensing image sample set,thus it is difficult to train a practical cloud and cloud shadow detection model.In this dissertation,cloud and cloud shadow detection is formulated as a semantic segmentation problem.The investigation and analysis of the latest cloud detection algorithm based on convolutional neural network for this dissertation revealed that it has an inherent disadvantage and constraint of relying on large-scale manually labeled data.Hence,an effort to improve it by proposing a method for model training,and cloud as well as cloud shadow detection of convolutional neural network without the prerequisite of manually labeled data,which is inspired by the idea of weakly supervised learning.Different from the existing detection methods based on deep convolutional neural network,the Landsat Quality Assessment Bands are used as labeled data for model training in the proposed method,and improves the labeled data through iterative training to extract cloud and cloud shadow features more effectively.The sample data covered 8 kinds of biomes,each containing 12 scenes,the selection of which is random but representative,with preprocessing such as band operation and image cropping.The semantic segmentation model based on deep convolutional neural network is applied to detect the cloud and its shadow in Landsat images.Results show that,the proposed method can comparatively completely identify clouds and cloud shadows on Landsat images in different scenes.The overall accuracy of this method is 85.55%,Kappa coefficient is 66.03%,and F1 score reaches 78.14%,43.86% and 90.62% for cloud,cloud shadow and clear categories,respectively,which are better than the results of CFMask,the Landsat official cloud detection algorithm.Hence,the results convincingly demonstrate that it is feasible to train a practical deep network model to detect the clouds and their shadows on Landsat images with QA Band,which lays the foundation for a follow-up study.