Statistical algorithms for land/forest cover change detection using remote sensing data

Land cover changes significantly affect climate, hydrology, bio-diversity, socio-economic stability and food security. Some of these changes being studied in remote sensing discipline include, but are not limited to, anthropogenic changes e.g. clear-cutting of forests for human settlements, and beetle/insect infestations in the forests. Beetle/insect infestations cause considerable damage to the forests resulting in tree mortality on large scale which provides fuel for fires and wastes valuable wood. Therefore, early detection of such changes is often desired by the authorities in order to carry out timely actions to mitigate them. However, manual monitoring using high resolution photography or field surveys can become very difficult and time consuming or even infeasible because such changes cover very large areas. This necessitates development of automated remote sensing algorithms which can monitor large areas with minimal human intervention. Several land cover change detection algorithms exist in literature which utilize remotely sensed imagery captured by different satellites. However, there are a very few studies which detect such changes in near-real time manner. Furthermore, there is still a room for improvement in the detection accuracy, detection delays and computational complexity of such algorithms. This thesis utilizes coarse (500 m) and moderate (30 m) spatial resolution satellite imagery (MODIS and Landsat 7 ETM+, respectively) and proposes four statistical algorithms for detection of land cover changes with significant improvements. The first algorithm (published in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing) is a supervised technique developed for near-real time detection of beetle infestations in pine forests of North America (British Columbia and Colorado). It models the hyper-temporal multi-spectral MODIS Vegetation Index (VI) time series with a triply modulated cosine function using a sliding window non-linear least squares and applies a change metric based on log-likelihood ratios to the trend parameter time-series of the fitted model, instead of the raw vegetation index. Significant improvement, in the detection accuracy with reduced detection delays, was achieved with this first published algorithm. The second algorithm (published in IEEE Geoscience and Remote Sensing Letters) is unsupervised and makes use of properties of Martingale Central Limit Theorem (MCLT) in the change metric derived from the parameter time series, in order to avoid threshold tuning while detecting beetle infestation in MODIS vegetation index time series. The third algorithm (published in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing) avoids the Gaussian distribution based change metrics, one of the limitations of the existing methods, and improves the change detection accuracy and detection delays significantly by using assumption free, directly estimated relative density-ratio based Repeated Sequential Probability Ratio Test (RSPRT) as test statistic on the parameter time series. The fourth algorithm (submitted to ISPRS Journal of Photogrammetry and Remote Sensing) is a supervised technique that utilizes bi-temporal multi-spectral Landsat 7 ETM+ imagery and models class posterior probabilities of the change and no change classes non-parametrically, using a linear combination of a large number of Gaussian Kernels through Least Squares Probabilistic Classifier (LSPC) formulation. This helps in avoiding the Gaussian assumption about the data, which is a major drawback of the traditional Bayesian Classifiers e.g. Maximum Likelhood Classifier (MLC) and Naive Bayes (NB). Another popular classifier which avoids this limitation is Support Vector Machine (SVM) but its drawback is that it assigns only class labels to the test samples and does not provide probabilistic class-memberships. Moreover, it is a binary classifier in its original formulation and needs different strategies to be adopted in order to use it for multi-class problems and also to make it probabilistic. The LSPC based framework, on the other hand, is non-parametric as well as capable of assigning degree of class-membership (probabilistic) to test samples and handling multi-class problems in its original formulation. Its application to leaf beetle infestation problem in north-eastern Tasmanian Eucalyptus plantations suggested improvement in detection accuracy. These methods have been compared with their respective counterparts in the literature in order to demonstrate their effectiveness. The contribution of this thesis is four folds: (i) advancing our understanding of land/forest cover change detection using quantitative/statistical approaches, (ii) proposition of reliable statistical approaches for near-real time land/forest cover change detection in hypertemporal coarse spatial resolution MODIS data, with especial emphasis on changes due to beetle infestations in pine forests, (iii) efficient threshold selection in complex scenario of near-real time change detection, when an optimal trade off between more than 2 performance indices is needed, and (iv) proposition of a non-parametric approach for detecting land/forest cover changes (bark beetle problem in north-eastern Tasmanian Eucalyptus plantations) in bi-temporal Landsat 7 EMT+ data.