SYNERGETIC USE OF MICROWAVE AND OPTICAL SATELLITE DATA FOR LAND COVER CLASSIFICATION AND SURFACE PARAMETER RETRIEVAL

Abstract

In the recent years, with the availability of different optical and microwave satellite sensors, there is a possibility to detect and identify various natural and man-made targets efficiently. Therefore, it is necessary to utilize the available sensor data effectively for various challenging applications, like target identification, mixed pixel classification, soil moisture retrieval over vegetation-covered regions, and crop yield estimation. However, while using these sensor data, one of the important aspects is to know, how to use them with minimum or no a priori information. Therefore, in this thesis, we have attempted to utilize microwave and optical satellite data synergistically for land cover classification and surface parameter retrieval with minimal a priori information. The main emphasis of this thesis is to study and develop methods for target identification using wavelet-based resolution enhancement techniques, segregation of mixed pixels of various land cover classes with the synergism of SAR and optical data, and soil moisture retrieval over vegetation-covered regions using an impedance approach of transmission line theory. The study and experiments have been performed majorly on some regions of Uttarakhand, India, i.e., Roorkee city and its surrounding areas. The general objective is supported by performing the following tasks:  Development of optimal weight algorithm for efficient application of DT-CWT for resolution enhancement of satellite images.  Development of an efficient approach for computation of optimal weights for multisensor data fusion: A special application of land cover classification.  Development of an efficient contextual algorithm for discrimination of tall vegetation and urban for PALSAR data.  Development of EM model based approach for soil moisture retrieval under vegetation-covered areas with SAR data. This thesis includes seven chapters. The introduction of the thesis is portrayed in Chapter 1, which includes motivation, major research gaps, and framework of research objectives. Chapter 2 discusses the brief literature review of related works to give the idea of current state-of-art techniques and challenges persisting in existing approaches of wavelet-based ii resolution enhancement, multisensor data fusion for land cover classification, segregation of mixed pixels of tall vegetation and urban, and soil moisture estimation. In Chapter 3, the role of optimal weights for subbands in dual-tree complex wavelet transform (DT-CWT) for resolution enhancement of satellite images is explored. DT-CWT is used to decompose an input image into different subbands. Every subband possesses a different level of information. Therefore, the optimization of the information on subbands facilitates the possibility of getting better results. Henceforth, an approach is proposed which could effectively select the optimal weights of individual subbands automatically according to the variances of each subband using variance minimization method in real time and is one of the best techniques as it is data independent. It infers that as image statistics (i.e., variance) changes with the image from different date/scene/sensor data, the variance will automatically compute the required weights without the requirement of any a priori parameter. Consequently, in the proposed technique, the effective reconstruction of the image is performed in order to obtain the sharper enhanced image. For performance assessment, PSNR, RMSE and correlation coefficient are evaluated. A good correlation is observed between the resolution enhanced and original images (reference images), with an improvement of 6-7 % in correlation using the proposed technique as compared to the existing DT-CWT method. Using the proposed algorithm, PSNR seems to be improved by 1.5 to 2.3 dB approximately as compared to the other method. Hence, image quality is enhanced using the proposed algorithm. The optimized transform is, consequently, a potentially valuable approach for resolution enhancement, and the developed technique has the potential to be used for different kind of data. Chapter 4 presents an approach for the fusion of multisensor data with optimal weights obtained using artificial neural network (ANN) approach, in order to get the enhanced land cover information. PALSAR is a fully polarimetric data with HH, HV and VV polarizations at good spatial resolution (25m). HH polarization provides double bounce scattering which represents urban regions, HV polarization gives volume scattering which represents agricultural regions, and VV polarization provides surface scattering which represents bare soil regions. On the other hand, some derivatives such as NDVI and NDWI are obtained from freely available Landsat (optical) data at 30m spatial resolution. Henceforth, multispectral reflectance information of Landsat data is fused with the polarimetric information of PALSAR data in order to perform the experiments. The optimal weights of iii multisensor data have been obtained by utilizing the ANN and variance based approach, and observed that weights are quite useful for improvement in information enhancement. Neural network is employed for this study instead of usage of analytical formulae of variance minimization approach, since the trained network could provide a better generalization of the formulae than the direct use of them. The developed fusion approach is able to accomplish improved vegetation and water region accuracy than that obtained by using PALSAR data, and Landsat data individually. Quantitative and visual results show that significantly better classification accuracy is achieved for the vegetation and water region with the developed technique. In Chapter 5, an attempt is made to analyze and observe the effectiveness of different texture features in improving the segregation of tall vegetation and urban land cover classes. A tree canopy/ tall vegetation is considered as an assortment of various randomly distributed discrete scatterers such as leaves, and branches; and the trees in a natural vegetation community are also distributed randomly, both spatially and in height, because of which tall vegetation region becomes more rough. The randomness is the cause of roughness which results in image textural properties that are distinctly different from those associated with urban, as the level of randomness is less in urban as compared to tall vegetation. While dealing with natural targets, like tall vegetation, characteristic of the textural feature, i.e., roughness may be an important parameter, which could identify these targets since both the classes possess a different type of roughness. Henceforth, commonly used texture features, i.e., fractal dimension, lacunarity, Moran's I, entropy, and correlation were critically analyzed and realized that these features are still lacking in the concerned segregation because they are pixel-based in general. Consequently, neighboring pixels are taken into account, and an approach has been developed by considering the randomness response (or manner of distribution of scatterers) based on relative similarity of total backscattering power of neighboring pixels by proposing a similarity entropy feature. An optimized threshold method is also developed by means of contextual thresholding in order to provide a proper decision boundary between the two classes using the proposed measure. The approach is successfully tested and validated on different fully polarimetric PALSAR data with the sensitivity of tall vegetation, and urban as 0.930 and 0.932, respectively. Chapter 6 deals with the task of estimating soil moisture over vegetation-covered areas using dual (VH and VV) polarized Sentinel-1 data. For the effective estimation of soil iv moisture in such areas with the requirement of minimum a priori information, an approach based on separation of impedances of different media (vegetation and soil media), is attempted. Sugarcane is one of the most prominent crops cultivated in the study area around Roorkee, Uttarakhand. Therefore, a transmission line theory based EM model is developed for sugarcane fields. The model is based on physical understanding of scatterers, and it utilizes the concept of transmission line theory. The model consists of one layer of vegetation (sugarcane crop) and another of soil. It helps in obtaining the backscattering coefficient from vegetation and soil regions separately. The model could relate the impedance (hence, backscattering coefficient) of each layer with the dielectric constant (ε =ε’−jε’’) and thickness (t) of that layer. However, some parameters are estimated by employing the propagation two-phase mixing model and optimization technique in order to obtain the volumetric soil moisture for the top layer (first 2cm) of the soil. For performance evaluation, the estimated result is compared with the respective ground truth information collected during the field visits. RMSE value is computed for the data, and the value obtained is 0.119. The observation of the scatter plot shows that the soil moisture obtained is being little under-estimated at some points, but are fairly close to the ground truth data. This may be due to the possible presence of random scattering behavior of sugarcane crop. The approach has also been validated by applying on the data of another date, and the results are obtained with an RMS error of 0.083. It indicates that there is a quite good agreement between the retrieved soil moisture and the ground truth data. Finally, the thesis has been concluded in Chapter 7, summarizing the results and depicting the considerable contributions made in the thesis. The perspective of future exploration utilizing the present results is also deliberated.