Satellite Image Time Series Classification And Change Detection Based On Recurrent Neural Network Model

As a carrier of multi-temporal images,satellite image time series contains tremendous amounts of information of ground targets in space and time.How to effectively mine the spatiotemporal information plays an important role in the application of remote sensing images.However,the existing data mining approaches often have insufficient utilization of temporal information,leading to various problems.For sequential data modeling and forecasting,Recurrent Neural Networks(RNNs)have natural advantages and have been successfully applied in the fields of speech recognition and computer vision.RNNs keep expanding into new areas,under this trend,we intend to introduce them into the processing and analysis of remote sensing images.We want to solve some key issues of applications of remote sensing with the high performance of RNNs.For this reason,this paper applies RNN models to classification and change detection of satellite image time series.According to the characteristics of RNNs and research objectives of this paper,two kinds of deep RNN models are proposed for MODIS time series classification.This paper realizes land cover classification in Beijing with the two models.At the same time,we apply the sequence-to-sequence(seq2seq)model with attention widely applied in the field of machine translation to the change detection of Landsat-8 image time series.This paper realizes land cover change detection in Changeping District of Beijing with seq2 seq model.In this paper,a series of problems related to satellite image time series classification and change detection are studied based on the RNNs.The concrete contents and contributions are listed as follows.(1)Two kinds of deep RNN models are proposed for MODIS time series classification,one is dSLSTM(deep Stacked Long Short Term Memory)model and the other is dBLSTM(deep Bidirectional Long Short Term Memory)model.The DSdSLSTM model constructs a multi-layer network by stacking LSTMs,and the DB-dBLSTM model realizes ―forward‖ and ―reverse‖ processing of time seies by combining two separate LSTM models.As deep neural networks,each model contains a multi-layer network structure.Experimental results show that dSLSTM model and dBLSTM model have good performance in the classification of satellite image time series.The classification accuracy is better than the SVM(Support Vector Machine)and RF(Random Forest)algorithms.What’s more,selecting the appropriate network structure and parameters of two models can improve the classification accuracy.(2)The seq2 seq model with attentional mechanism used in the field of machine translation is applied in land cover change detection.We use seq2 seq model to achieve the correspondence between remote sensing images and ground targets’ states,and use this "sequence to sequence" corresponding structure to achieve direct extraction of land cover change information.In order to improve the accuracy of the change detection and eliminate some of the pseudo-changes,the land use transition probability matrix is added as a prior knowledge to the auxiliary judgment in the change detection.Experimental results show that the use of seq2 seq model for land cover change detection is effective,and it has better performance than CCDC(Continuous Change Detection and Classification)model in change detection accuracy.(3)In order to solve the problem of lack of training samples,data augmentation methods for satellite image time series were proposed,including truncated augmentation,hybrid augmentation,translation and scaling.The truncated augmentation method augments the training data by truncating and jointing time series of same class.Hybrid augmentation method synthesizes time series of mixed pixels by combining time series of two different type sequences.In order to increase diversity of the training sample set,the translation and scaling method simulates satellite image time series changes due to advance or delay of phenology by translating and scaling time series.Similar to truncated augmentation method,this paper also proposes a method for simulating changed time series.That is using the selected unchanged time series to construct the changed time series.Experimental results show that the data augmentation methods can impore the classification accuracy of deep RNN models.The analog time series can be used for training the change detection model,and can obtain satisfactory change detection results.The data augmentation methods and data generation mehod for satellite image time series proposed in this paper are effective.