Research On The Effect Of Topographic Correction On Classification Accuracy Of Deep Neural Network Classifier

The classification technologies based on remote sensing images have become an extremely important monitoring means in land resources monitoring,forest resources investigation and other related fields,but its development is still restricted by many factors,the topographic effect caused by the relief of terrain in remote sensing images is one of the impacts that cannot be ignored.The shady and sunny slope caused by the topographic factors bring difficulties and challenges to traditional classifiers with weak fitting ability.In the classification and extraction of remote sensing images by using traditional classifier,due to the limitation of classifiers’ fitting ability,the influence of terrain factors often needs to be eliminated by appropriate topographic correction model,and the topographic correction can play a positive role in the improvement of image classification accuracy.Compared with traditional classifiers,deep neural network classifiers based on deep learning theory have the advantages of deep feature learning and self-fitting,which have emerged in the field of image classification and achieved with good accuracy improvement.However,when deep neural network classifiers are widely used in remote sensing classification tasks,whether the treatment of terrain factor still needs to be realized through topographic correction? Can the effects of terrain be fitted by the classifier? These questions have not yet been answered.This paper carries out a series of studies based on Landsat8 OLI 30 m satellite imagery by taking advantages of the U-Net and SegNet deep neural network classifier in image classification.By analyzing and comparing the accuracy of classification results before and after topographic correction obtained by the two classifiers,the necessity of topographic correction for deep neural network classification task is answered.The main achievements and innovations of the paper are as follows:Firstly,with the mixed sampling method,we have completed the classification of Landsat8 OLI remote sensing images before and after topographic correction with deep neural network classifier U-Net and SegNet,and the generalization verification of these two classifiers.By several comparative experiments and the strict precision evaluation,we have proven that topographic correction has no obvious influence on the classification accuracy of U-Net and SegNet deep neural network classifiers.Secondly,we proposed a sampling method of shady-sunny slope samples,and then on this basis,we completed the classification of Landsat8 OLI shady-slope and sunny-slope images before and after topographic correction with U-Net and the generalization verification of this classifier.It is further confirmed that in the classification application of deep neural network illustrated by the examples of the U-Net and SegNet,the topographic correction processing of satellite images before classification is unnecessary.Finally,taking the image classification extraction before and after topographic correction in typical regions as an example,the problem that whether topographic correction is needed in deep neural network classification is solved according to the systematic classification experiments and the changes of classification accuracy,which has practical value for the application of deep learning technology in remote sensing classification.