| Summary: | High resolution satellite imagery is more readily available and in larger quantities than ever before. It has become increasingly di cult for researchers in the planetary science community to manually analyse all of these images quickly and accurately. There have been several attempts over the last decade to address this by using image classi cation techniques which structure and organise planetary image datasets. However, these attempts have focused on the creation of the method rather than the analysis of the results. This thesis expands upon previous work by exploring methods to structure planetary image datasets that will allow the structure to be compared with other information about these images. The aim is to search for, and explore, trends within the resultant structure. A Context-Based Image Retrieval (CBIR) system is created using textons, a method to model texture as a histogram. These textural histograms are compared using the Earth Movers Distance (EMD) algorithm and the relationships between them are organised using Multi-Dimensional Scaling (MDS). A method to analyse the structure is presented using Pearson's correlation coe cient. These methods were applied to images of lunar craters, highlighting the e ect that illumination has on the appearance of texture in these images. This is important as it demonstrates the necessity for careful analysis of results produced by CBIR systems, as the work conducted by previous researchers has failed to take this into account. This thesis shows that if illumination is removed, or the data has little or no variation, texture analysis can be used to provide potentially important information about planetary features. In this study, once illumination was removed, it was found that the resultant structure was highly correlated with the depth/diameter (d=D) ratio of the crater, thus providing a method of predicting the d=D ratio of craters from the texture model alone. To explore the strengths and limitations of the proposed system it was applied to the di erent, but challenging, task of estimating grassland poaching damage from digital camera images. This demonstrates its potential use beyond the planetary science community. The work undertaken in this thesis demonstrates the use of CBIR systems within the planetary science community and highlights that the results obtained must be examined carefully. CBIR systems which report good visual similarity may appear to be useful for scienti c research but it is not until the results are analysed that this can be established.
|
|---|
