Continuous change detection and classification of land cover using all available Landsat data
MetadataShow full item record
Land cover mapping and monitoring has been widely recognized as important for understanding global change and in particular, human contributions. This research emphasizes the use ofthe time domain for mapping land cover and changes in land cover using satellite images. Unlike most prior methods that compare pairs or sets of images for identifying change, this research compares observations with model predictions. Moreover, instead of classifying satellite images directly, it uses coefficients from time series models as inputs for land cover mapping. The methods developed are capable of detecting many kinds of land cover change as they occur and providing land cover maps for any given time at high temporal frequency. One key processing step of the satellite images is the elimination of "noisy" observations due to clouds, cloud shadows, and snow. I developed a new algorithm called Fmask that processes each Landsat scene individually using an object-based method. For a globally distributed set ofreference data, the overall cloud detection accuracy is 96%. A second step further improves cloud detection by using temporal information. The first application ofthe new methods based on time series analysis found change in forests in an area in Georgia and South Carolina. After the difference between observed and predicted reflectance exceeds a threshold three consecutive times a site is identified as forest disturbance. Accuracy assessment reveals that both the producers and users accuracies are higher than 95% in the spatial domain and approximately 94% in the temporal domain. The second application ofthis new approach extends the algorithm to include identification of a wide variety of land cover changes as well as land cover mapping. In this approach, the entire archive of Landsat imagery is analyzed to produce a comprehensive land cover history ofthe Boston region. The results are accurate for detecting change, with producers accuracy of 98% and users accuracies of 86% in the spatial domain and temporal accuracy of 80%. Overall, this research demonstrates the great potential for use of time series analysis of satellite images to monitor land cover change.
Thesis (Ph.D.)--Boston University
RightsThis work is being made available in OpenBU by permission of its author, and is available for research purposes only. All rights are reserved to the author.