A Research Of Remote Sensing Image Change Detection Algorithm Based On Neural Network Feature

Change detection(CD)based on remote sensing(RS)field plays an important role in many areas.It is related to the sustainable development of regions and has been successfully applied in fields about environmental monitoring,resource management,disaster assessment,etc.However,universal method does not exist yet.Traditional change detection methods always focuse on the spectral data of imagery.With the increasing of spatial resolution and temporal resolution,the accuracy of change detection based on remote sensing images can not meet the requirements.Deep neural networks perform perfect in the tasks with image.It can learn automatically to extract useful features of the images in data set.With thoes features,models have an ability to get more acurrate results of the various specific tasks.Applicating neural network in RS field helps to solve the problem that traditional RS image field needs artificially designed features.Supervised neural network methods require a lot of data with corresponding labels.However,change detection data set in RS field is time-consuming,expensive and low-accuracy.In this thesis,a change detection approach based on neural network features is proposed,which can perform well without pixel-wise data sets about change detection task.The main contents of this thesis can be described as follows.(1)The training data is usually much less compared with data in application.To improve the accuracy of RS image classfication network in the condition of small-scale training dataset,we designed and constructed a flat network model called MICBAM.In MICBAM,we use Inception module to extract multi-scale features and attention mechanism to effectively extract spatial and spatial information.The model was proven effective in different partition of training set and test set.(2)Post-classification detection is the most commonly used change detection method,whose accuracy is greatly affected by classification precision.Usually,the category of ground image is much more than that in training data,and this situation has a huge impact on CD accuracy.In order to improve the classification accuracy We transferred Mic BAM model which pre-training in Euro SAT,and then added category correlation factors into label smoothing.Finnaly,we accuquired CD map by classification confidence.The proposed method was validated effective in Wuhan multi-temporal scene change dataset.(3)The cost of pixel-level CD dataset is quite high,and the it is always not very precise.limiting the scope of supervised methods application.However,the feature extraction ability is limited in unsupervised neural network method and traditional method.Therefore,we proposed a method combined weakly supervised method and traditional method for the CD task.A fully connected classification network is constructed and then pre-trained on object-level dataset.With the pretrained model,we can extract feature change map,filter it with percentile threshold,and anlysis it with PCA and Kmeans methods,finnally we can get the change map.The proposed method was experimented effective on Hong Kong,Quickbird and Ziyuan3 datasets.