Spatio-Temporal Algorithms For Anomaly Detection Using Remote Sensing Image Series:Methodology Research And Application

There are numerous abnormal events occuring at any given moment in the world.They are caused by various factors including extreme weather,human behavior and crustal movement and so on.Mostly,these abnormal events can bring about serious ecological damage and economic loss in a rather short time,especially the events of natural disaster like flood,earthquake and fire,which have an extensive and meaningful efffection on the environment.At its worst,the influence is even destructive.Therefore by this field,wide attention has been attracted from many governments and scholars.Although the emergencies are usually inevitable,and even unpredictable as far as our current science and technology development concerns,works on the early warning and real-time monitoring of the emergencies can definitely make a siginificant difference,not only for the evacuation and remission during the emergency,but also for the prevention and reconstruction after the emergency.As a rising detection tool in the past decades,satellite remote sensing technology contains great potential in mapping emergencies,owing to its capacity of multi-spectral,wide-range,all-time,all-weather and easy availability.However,due to the seasonal changes and vegetation growing,the land is always in a dynamic procedure of constant transformation,even there is no emergency occurring.The accumulation of remote sensing data in time series can help handle this situation,and discriminate the changes caused by the abnormal events among all the changes of various time scales.At present,the mostly adopted time series remote sensing datasets are AVHRR,MODIS and Landsat,which have made some achievements in anomaly detection.But as the spatial resolution and the temporal resolution are two conflicting parameters,these datasets can not produce an anomaly detection result with good timeliness and rich details simultaneously.With the recent improvement of sensor technology,some moderate resolution remote sensing time series datasets are emerging,for instances,the HJ-1A/B CCD and the GF-4 PMS data.These new datasets enable a rapid and subtle detection,and propose new requirements for novel algorithms which can make good use of the spatial and temporal information synthetically.In this dissertation,the author studies a set of algorithms combining spatial and temporal information using GF-4 PMS and HJ-1A/B CCD datasets,aiming at obtaining accurate and automatically acquired anomaly mapping results.Taking the 2013 Heilongjiang River flood and 2016 Dongtinghu Lake flood as examples,the main research and contribution realized by this dissertation can be summarized as follows:(1)In remote sensing data,same cover types can have different radiance values,and different cover types can have same radian values.Concerning this issue,a spatio-temporal context model based anomaly detection algorithm is proposed.Firstly,the relationship between pixels with their spatio-temporal neighborhood is statistically formulated and updated along the time.If the context model changes dramatically,some anomaly will be considered existing in this test scene.(2)Apply the spatio-temporal context method to flood detection.In order to allievate the false positive rate brought by the first algorithm,a flood detection algorithm combing spatio-temporal model and weak supervised learning is proposed.Permanent pixels with constant context are automatically selected for training a Modest AdaBoost classifier.With the strong classifier obtained,the flood extent can be precisely and rapidly delineated.(3)Genenrally,the anomaly detection methods can be divided into two categories: the direct anomaly detection methods and the prediction based anomaly detection methods.Besides the direct anomly detection methods utilizing spatio-temporal context information,this dissertation also proposes a prediction based anomaly detection method utilizing a spatio-temporal LSTM model.This method generates the reconstructed time serial using both the spatial and temporal values.With it,the LSTM network get trained.Through integrating the spatio-temporal LSTM prediction and the Gaussian distribution based anomaly judegement strategy,the proposed algorithm achieves certain effects in the experiments.(4)Apply the spatiotemporal LSTM method to flood detection.In order to lift the efficiency of neural network based methods in large-scale flood detection,a flood detection algorithm combining spatio-temporal LSTM and unsupervised clustering is proposed.In an object based fashion instead of a pixel-by-pixel one,this algorithm provides a thermal dynamic diagram indicating the possibility of anomaly occurrence in the test image.Experiments show that,the spatio-temporal LSTM and clustering based method can achieve a wide applicability for detecting different anomaly types with different datasets.In conclusion,with full expoloration in spatio-temporal information,proper construction in modeling and wise selection in parameters,the moderate-resolution remote sensing time series data can perform well in the emergency response applications.