Spatio-Temporal Modelling For Remote Sensing Image Change Detection:Methods And Applications

In recent decades,it witnesses the continuous development of remote sensing imaging technology and the increasing in temporal,spatial and spectral resolution.The open access of large-volume,heterogeneous and spatio-temporal remote sensing big data,the significant improvement in cloud computing power and the advance in artificial intelligence such as machine leaning and deep learning,all of these will speed up and enpower the near realtime applications of remote sensing big data in agriculture monitoring,disaster prevention and reduction,environmental and resource prevention,and land-use etc.,especially with the emerging of remote sensing data access and processing platforms that integrate big data,computation power and algorithms,such as Google Earth Engine.With these advances,as one of the main applications of spatio-temporal remote sensing data,change detection problem will be transferred gradually from bi-/multi-temporal to time series,from homogeneous data to multi-source heterogeneous data,from 2-dimension to higher dimensions and from local-scale to large-scale analysis.This thesis conducts five fundamental researches on change detection methods and applications based on spatio-temporal modelling.For change detection problem in remote sensing data with different spatial resolution,Chapter 2 presents a feature mapping network based on the stacked auto-encoder and validates it on both homogeneous and heterogeneous data.Due to the difference of multi-resolution data in spatial details and imaging conditions,it is challenging to compare and analyze them directly.The core of binary change detection is to accurately estimate the difference degree among multi-temporal imageries,and the key lies in how to suppress the difference degree among unchanged samples.In consideration of these,we design a feature mapping network and adopt an unsupervised change detection technique to select reliable unchanged samples for its training,which significantly reduces the difference between unchanged bi-temporal samples in feature space and has the potential to construct a change map with a high contrast among changed and unchanged pixels.To solve the change detection challenge in multi-source heterogeneous remote sensing data,Chapter 3 proposes a multi-source data change detection framework based on coupled dictionary learning.The difference in imaging mechanism of multi-source sensors would make them shape very different descriptions on the same scene or object,which makes it very challenging to conduct direct comparison and analysis.However,multi-source data should be comparable in nature,to some extent at least,though they may have very different perceptions on the same object.Therefore,we transform multi-source data into a comparable high-dimension feature space from the original observation space by establishing a coupled dictionary model,achieving bi-temporal comparison and analysis.Based on the learnt coupled dictionaries,the difference degree can be estimated and the changes can be determined automatically,which can be used to select unchanged samples to fine-tune the current coupled dictionaries again.The procedure can be iteratively executed until a termination condition is met,and the desired coupled dictionaries can be learnt,which will be used to derive the final change map.As for detecting multiple types of changes from remote sensing data,Chapter 4 proposes a deep difference representation learning network.The increasing in spatial,temporal and spectral resolution make it more important to jointly interpret spatio-temporal remote sensing data.And,the objective of multi-type change detection is to detect,classify and interpret the dynamic changes over the earth surface,interfering the relationship between human activities and changes in earth environment.Due to the fact that deep neural network relies too much on labeled data while it is very limited in remote sensing community,this thesis designs the deep difference representation learning model which unifies feature learning,difference estimation and difference representation learning,and the deep networks can be trained to learn discriminative and clustering-friendly representations for characterizing different types of changes.By modelling a unified model like this,both the advantages of deep network and unsupervised clustering are well exploited for detecting multiple types of changes.To meet the urgent demands on the near real-time monitoring on wildfire,both Sentinel-2and Landsat-8 multi-spectral data are combined to form a normalized burn ratio index time series.In Chapter 5,the Season-Trend model is exploited to decompose time series observation into three terms including season term,trend term and residual term.The season term is used to remove the periodic and seasonal normal variations from time series,while the trend term is used to reduce the influence of long-term directional trend on wildfire detection.Then,the mean and standard deviation of the residual term before wildfire event are used to estimate the ratio between the residual term after wildfire and the standard deviation,which accounts for the burn severity.Additionally,the learned Season-Trend model can predict the future observation according to stable historical time series,which gets rid of the dependence on a well-selected,cloud-free and pre-fire image and significantly enhances the detectability of burnt areas,and it will be very helpful in the monitoring and evaluation of the vegetation recovery after wildfire event.As for the application of detecting new built-up areas,Xi'an,the capital of Shaan'xi,is taken as the study area in Chapter 6 and multi-source remote sensing data is exploited to estimate the changes and determine the new built-up areas from 2010 to 2015.Due to the fact that the interferometric coherence has the ability to enhance the stable built-up areas,pixel-based random sampling and superpixel-based sampling are independently investigated to analyze the correlation between PALSAR L-Band HH,ENVISAT/Stentiel-1 C-Band VV,Terra SAR X-Band HH and C-Band VV coherence respectively,which demonstrates that superpixel-based analysis occupies better correlation than pixel-based.Then,we analyze the correlation among multi-source measurements including both optical and radar data based on superpixel-based sampling,and the correlation among bi-temporal variation of multi-source data shows that it is necessary and advantageous to fuse change estimations derived from multi-source data.Finally,we use -percentile to detect the changes from multi-source change estimations and fuse them together,where is automatically determined by minimizing the average sum of mutual information between any two change estimations.The new built-up detection results are compared with high-resolution false color image composed with bi-temporal Terra SAR X-Band HH data for validation.