|M.Sc Student||Cohen Lior|
|Subject||Improving the Spectral Separability between Objects|
through Wavelet Transformation
|Department||Department of Civil and Environmental Engineering||Supervisors||Professor Maxim Shoshany|
|Dr. Ophir Regev Almog|
|Full Thesis text - in Hebrew|
The main purpose of the mapping process is to characterize spatial objects features, with high precision and accuracy. Mapping products based on multi-spectral remotely-sensed data serve wide variety of environmental applications. The remote sensing image pixel values (spectral reflectance) are highly affected by: a) atmospheric and irradiation conditions; b) sensor's flight and viewing properties; c) spatial resolution which cause mixing of pure spectral signatures. These issues cause the following difficulties in land cover classification when using common statistical methods: a) surface cover types are rarely homogenous and thus represented by high variability and noisy spectral data; b) the use of simple statistics such as Mean and Variance, to characterize them is too simplistic. Our main hypothesis is that by increasing the spectral uniqueness of each surface cover type, based on mathematical transformation of the original spectral data we may improve their identification and mapping.
In this study we developed for this purpose the Spectral Significant Features (SSF) classification method, based on the following principles: a) Increasing inner dimensionality of each signature by representing it by the ratios between all combinations of spectral bands; 2) Usage of 'Sum Of Differences' separability index to detect the bands contributing most to the increase of the spectral uniqueness of each surface cover type. We hypothesized that implementing the SSF method within an iterative process allowing increase in the number of unique spectral features (band ratios) used to resolve confusion between cover types will increase the precision and accuracy of the classification process.
Assessment of the method was initially conducted through the implementation of synthetic experiments based on library spectras. We ‘contaminated’ these spectras by adding artificial noise, topographic effects and by creating mixtures. It was shown then that by using the new spectral feature space of band ratios the differentiation between the ‘contaminated’ signatures was better than by using the original spectral data.
Next, we examined the classification products of well known Support vector machine (SVM) and Maximum likelihood (ML) methods against the SSF classification method. The results indicated that the SSF method achieved better classification results than the SVM and ML methods.