Nature-Inspired Algorithms for Enhancement and Land-Cover Classification of Satellite Images

Abstract

Satellite images are important for various applications in different domains of remote sensing such as geographical information systems, geosciences, land-cover mapping, change detection, etc. Land-cover classification of satellite images is a very predominant area since the last few years. Multispectral (MS) and hyper-spectral (HS) satellite imagery are steadily growing in its popularity as a digital means for remote sensing, terrain analysis and for detecting thermal signature. It is often used as a viable alternative for mapping applications when standard mapping and geodesy products become inadequate or outdated. This thesis investigates the impact of various nature-inspired optimization algorithms in different phases of satellite image processing. Satellite images normally possess relatively narrow brightness value ranges necessitating the requirement for contrast stretching, preserving the relevant details before further image analysis. Most of the traditional enhancement approaches are highly dependent on the image to be processed and hence require manual human intervention. The automation of satellite image enhancement process requires defining an evaluation criterion valid across a wide range of such image datasets. A parameterized transformation function evaluated by a fitness criterion is hence adopted for this process. Nature-inspired optimization algorithms, a subclass of metaheuristic algorithms, are largely exploited for different image processing applications during the last few decades. The potential of such algorithms in following a guided random search path proved to be very useful for solving complex contrast enhancement problems. In this context, two new nature-inspired optimization algorithm based methods are proposed for the purpose of satellite image enhancement. Satellite image denoising is also essential for enhancing the visual quality of images since it often gets affected by noisy signals, thereby corrupting the original image. Two new natureinspired optimization algorithm based adaptive Wiener filtering methods are proposed for denoising multi-spectral satellite images corrupted with Gaussian noise. All the proposed contrast enhancement and denoising methods achieved better qualitative and quantitative results as compared ivwith the state-of-the-art methods. The increase in the internal variability of land-cover units and weak spectral resolution of satellite images, make the pixel level information extraction very difficult. Therefore, a thresholding based image segmentation process prior to the classification of these diverse terrains presents to be an appropriate approach. Satellite image segmentation is challenging due to the presence of weakly correlated and ambiguous multiple regions of interest. In the literature, several algorithms were developed to generate optimum threshold values for segmenting such images efficiently. But their exhaustive search nature makes them computationally expensive when extended to multilevel thresholding. Hence, two new nature-inspired algorithm based segmentation methods are proposed. The proposed algorithms evolved to be most promising, stable and computationally efficient for segmenting satellite images. Satellite images exhibit spatial and/or temporal dependencies in which the conventional machine learning algorithms fail to perform well. Another problem faced by satellite images acquired is the huge number of spectral bands they possess. Hence, a nature-inspired optimization algorithm driven, dimensionality reduction framework for automated land-cover classification is proposed. Experiments are conducted on multi-spectral as well as hyper-spectral satellite image datasets to demonstrate the robustness of the proposed method. It outperforms all the state-of-the-art land-cover classification methods, attaining an overall classification accuracy around 92 % and 99 % for different multi-spectral and hyper-spectral satellite image datasets respectively.