Improved Framework for Monitoring Land-cover Changes by Using Remote Sensing Technology: From unsupervised and supervised, to targeted approaches
MetadataShow full item record
Land-cover change monitoring has remained at the forefront of interest in environmental science and sustainability research for many years. Digital change detection, which mainly relies on remote-sensing technology, has been widely applied to encompass the quantification of dynamic land-cover change. A variety of digital change detection techniques can be categorized into unsupervised and supervised approaches depending on the availability of reference data. Traditional unsupervised and supervised approaches both suffer certain issues which may lead to unreliable results of change discrimination. This thesis seeks to advance current methodology of digital change detection in order to better address the complexity of practical land cover monitoring. Unsupervised methods are firstly investigated in the first manuscript under the condition of the unavailability of reference data. To deal with ambiguous definitions of ‘change’ and ‘no change’ existing in past research, an objective categorization system of ‘interested’ and ‘irrelevant’ change is constructed according to the nature of the considered scene of a given urban region. An advantageous unsupervised change detection procedure is hereby developed to separate these interested and irrelevant changes by fully exploiting the complementary attributes of luminance and saturation. In the second manuscript, a new and more applicable approach called ‘targeted change detection’ is described to improve change-detection techniques when reference data is available. In the proposed targeted approach, only reference data for targeted types are required with the goal of identifying only interested changes and discarding other classes, and thus the drawbacks of uncertainties and tedious sampling in the supervised approach can be addressed. An efficient algorithm for targeted change detection and its operational workflow are presented. Two very different study cases are used to demonstrate the proposed procedure with the results both showing the higher accuracy of our proposed method compared with other potential targeted change-detection methods.