Enhancing spatial resolution of remotely sensed data for mapping freshwater environments
Ashraf, M. S. (2011). Enhancing spatial resolution of remotely sensed data for mapping freshwater environments (Thesis, Doctor of Philosophy (PhD)). University of Waikato, Hamilton, New Zealand. Retrieved from https://hdl.handle.net/10289/5831
Permanent Research Commons link: https://hdl.handle.net/10289/5831
Freshwater environments are important for ecosystem services and biodiversity. These environments are subject to many natural and anthropogenic changes, which influence their quality; therefore, regular monitoring is required for their effective management. High biotic heterogeneity, elongated land/water interaction zones, and logistic difficulties with access make field based monitoring on a large scale expensive, inconsistent and often impractical. Remote sensing (RS) is an established mapping tool that overcomes these barriers. However, complex and heterogeneous vegetation and spectral variability due to water make freshwater environments challenging to map using remote sensing technology. Satellite images available for New Zealand were reviewed, in terms of cost, and spectral and spatial resolution. Particularly promising image data sets for freshwater mapping include the QuickBird and SPOT-5. However, for mapping freshwater environments a combination of images is required to obtain high spatial, spectral, radiometric, and temporal resolution. Data fusion (DF) is a framework of data processing tools and algorithms that combines images to improve spectral and spatial qualities. A range of DF techniques were reviewed and tested for performance using panchromatic and multispectral QB images of a semi-aquatic environment, on the southern shores of Lake Taupo, New Zealand. In order to discuss the mechanics of different DF techniques a classification consisting of three groups was used - (i) spatially-centric (ii) spectrally-centric and (iii) hybrid. Subtract resolution merge (SRM) is a hybrid technique and this research demonstrated that for a semi aquatic QuickBird image it out performed Brovey transformation (BT), principal component substitution (PCS), local mean and variance matching (LMVM), and optimised high pass filter addition (OHPFA). However some limitations were identified with SRM, which included the requirement for predetermined band weights, and the over-representation of the spatial edges in the NIR bands due to their high spectral variance. This research developed three modifications to the SRM technique that addressed these limitations. These were tested on QuickBird (QB), SPOT-5, and Vexcel aerial digital images, as well as a scanned coloured aerial photograph. A visual qualitative assessment and a range of spectral and spatial quantitative metrics were used to evaluate these modifications. These included spectral correlation and root mean squared error (RMSE), Sobel filter based spatial edges RMSE, and unsupervised classification. The first modification addressed the issue of predetermined spectral weights and explored two alternative regression methods (Least Absolute Deviation, and Ordinary Least Squares) to derive image-specific band weights for use in SRM. Both methods were found equally effective; however, OLS was preferred as it was more efficient in processing band weights compared to LAD. The second modification used a pixel block averaging function on high resolution panchromatic images to derive spatial edges for data fusion. This eliminated the need for spectral band weights, minimised spectral infidelity, and enabled the fusion of multi-platform data. The third modification addressed the issue of over-represented spatial edges by introducing a sophisticated contrast and luminance index to develop a new normalising function. This improved the spatial representation of the NIR band, which is particularly important for mapping vegetation. A combination of the second and third modification of SRM was effective in simultaneously minimising the overall spectral infidelity and undesired spatial errors for the NIR band of the fused image. This new method has been labelled Contrast and Luminance Normalised (CLN) data fusion, and has been demonstrated to make a significant contribution in fusing multi-platform, multi-sensor, multi-resolution, and multi-temporal data. This contributes to improvements in the classification and monitoring of fresh water environments using remote sensing.
University of Waikato
All items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.
- Higher Degree Theses