DEVELOPMENT OF SPECTRAL-SPATIAL STRATEGIES FOR DETECTION OF ENGINEERED OBJECTS USING HYPERSPECTRAL DATA

Abstract

Humans all across the globe migrate to cities/urban areas in search of better livelihood. In India alone, the migrant population moving to cities is likely to rise to 40% by 2030. Urbanization takes a heavy toll of the scarce resources. Besides, there are many adverse environmental effects of rapid urbanization. Urban planners, therefore, have to continuously control and monitor the urban expansion, plan amenities, make judicious allocation of lands for industries, residences and agriculture, ensure low environmental pollution and simultaneously also address several other challenges of urban planning. Remote sensing in general has been a very important supporting tool in the hands of urban planners in assessment of existing urban growth particularly in extraction of different levels of urban engineered surfaces such as roads and roofs etc. and its interpolation to assess future urban growth. The development in the field of remote sensing has therefore always been of interest to urban planners. The development of Hyperspectral Remote Sensing has further enabled urban planners in better assessment of urban expanse. However, though hyperspectral data is significantly more useful to the urban planners, it comes with its own set of challenges such as spectral variability, mixed pixel problems, accuracy requirements, requirement of recovery shape for correct identification of urban engineered surfaces (roads and roofs), selection of an appropriate approach such as target detection/classification/machine learning approach for information extraction providing better accuracy etc. The present Thesis explores one of the relevant problems useful for urban planners i.e development of spectral-spatial strategies for detection of engineered objects using hyperspectral data. This problem has been explored under three iv objectives. The first objective deals with an exhaustive comparative assessment of standard spectral target detection algorithms for engineered objects using hyperspectral data, under four categories. Various algorithms reported in literature have been considered for comparison. The second objective involves, development of different strategies for detection of engineered objects. It has been performed under two sub-objectives. In the first part, spectral - spatial urban target detection using Artificial Neural Network (ANN) has been explored. The second part explores, detection of the engineered surfaces (roads and roofs) using deep learning approach. The last objective expounds mixed pixel analysis and shape identification of engineered objects using hyperspectral data. This is also done in two parts, the first part deals with extraction of urban targets using fusion of spectral and shape features, and the second part deals with urban target detection using super-resolution mapping approach by recovery of shape. The data for the research is Airborne Visible and Infrared Imaging Spectrometer – Next Generation (AVIRIS-NG) hyperspectral data collected during a joint ISRO-NASA campaign held during 2016-2017. For the present study, urban hyperspectral data for Udaipur, Rajasthan captured during February 2016 has been used. For comparative assessment, different categories of target detection algorithms have been considered for extraction of roads and roofs. This has been implemented using reference spectra using both the in-scene (derived from image) and in-field (collected while ground data campaign) spectra. One of the findings of the results suggests that, Mahanalobis angle measure may be one of the robust angle measures for detection of roads and roofs. Besides in general, it is found that machine learning based methods such as ANN and ELM perform better amongst all the measures. Associating spatial information such as morphological attribute profiles along with spectral signatures of labelled pixels of targets have yielded higher accuracy as v compared to standard target detection methods. The approach seems to perform better when targets of interests are composed of similar materials. For instance, roads and roofs are often made up of concrete, asphalt etc. and therefore purely spectra-based delineation of these two surfaces is challenging. Further, CNN based measures appear to provide higher accuracy in automated feature extraction of complex urban targets with minimal human intervention. Spectral similarity of urban targets and coarse resolution of the sensor poses multiple challenges in their detection. Extraction of shape of urban engineered surfaces is an important part in urban planning. Therefore, two shape based features, exploiting the spatial aspect of hyperspectral data are proposed. Additionally, the shape of urban engineered surfaces (roads and roofs) is enhanced using unmixing based super resolution approach by taking the neighbouring pixels into account. The study, however, restricts itself in terms of extraction of different levels of roads and roofs. Besides, the study does not link up with extraction of road and roof surfaces with different urban applications such as determination of road and roof conditions and aging. These and several other relevant issues may be explored in future.