As an emerging field, hyperspectral imagery provides interesting opportunities in remote sensing applications, which are mostly based on supervised classification. However, supervised classification has issues related to the unbalance between the high dimensionality and the limited availability of the acquired training samples leading to the Hughes phenomenon. Moreover, noises from the environment and optical sensors can further reduce the classification effectiveness. Thus, to address these problems, intensive work has been devoted to provide accurate pixel-wise classifiers. Although these pixel-wise classifiers can fully use the spectral information in the hyperspectral imagery, the spatial dimension is not taken in consideration. Consequently, the resulting classification maps are corrupted with salt-and-pepper noise.
　　Indeed, the interest of hyperspectral imagery comes from the wide range in the electromagnetic spectral domain and the high resolution. Thus, by integrating spectral and spatial information, there is a great potential to extract more discriminative and comprehensive representation for objects of interest, and making the classification task easier. In recent years, spectral-spatial classification of hyperspectral imagery has attracted a great attention. Thus, the hyperspectral community has arrived at a point where several classification methods provide very high classification accuracy. Nevertheless, the major concurrent aspects between existing methods concern their simplicity and tuning configurations.
　　In this regard, this dissertation aims at conceiving novel simple and effective strategies for the classification of hyperspectral images, placing the focal point on the exploration of the feasibility of applying different transform domain techniques including Discrete Wavelets transform (DWT), Wavelet Packets Transform (WPT), Discrete Cosine Transform (DCT) to fully exploit the spectral and spatial information contained in hyperspectral images. Thus, several approaches based on the idea of performing the spatial filtering on the high frequency of the transform domain of the original spectral signature are proposed to improve the classification of hyperspectral imagery in terms of classification accuracy, the sensitivity regarding different sizes of the training samples, the simplicity of the tuning configuration, and computational time. Indeed, regarding the applied spectral filter, three main strategies are proposed,including spectral multilevel DWT decomposition-based approaches, spectral wavelet Packet-based approaches, and spectral DCT based approaches.
　　In the first strategy, a Spectral multilevel DWT decomposition (SDWT) is performed on the hyperspectral image to separate the approximation coefficients from the detail coefficients. For each level of decomposition, only the detail coefficients are spatially filtered instead of being discarded; as it is often adopted by the wavelet-based approaches. Thus, three different spatial filters are explored, including two-dimensional DWT (2D-DWT), two-dimensional Adaptive Wiener Filter (2D-AWF), and two-dimensional Discrete Cosine Transform (2D-DCT). After the enhancement of the spectral information by performing the spatial filter on the detail coefficients, multilevel DWT reconstruction is carried out on both the approximation and the filtered detail coefficients. Finally, the final preprocessed image is fed into a linear Support Vector Machine (SVM) classifier.
　　In the second strategy, a Spectral Wavelet Packet Transform (SWPT) is performed consecutively on the hyperspectral image to separate the approximation coefficients from the detail coefficients. In the last decomposition level, a spatial filter is performed on all coefficients except the approximation coefficients obtained by the lowpass filtering iteration in all the previous levels. For the spatial filtering step, as our first strategy, three different spatial filters are also explored, including2D-DWT, AWF, and 2D-DCT. Then, Wavelets packets reconstruction is carried out on both the approximation part and the spatially filtered parts, and the final preprocessed image is fed into a linear SVM classifier.
　　In the third strategy, Spectral DCT is performed on the hyperspectral image to obtain a spectral profile representation, where the most significant information in the transform domain is concentrated in a few low-frequency components. The high-frequency components that generally represent noisy data are further processed using a spatial filter to extract the remaining useful information. For the spatial filtering step, 2D-DCT with global default threshold, 2D-AWF, and 2D-DWT are explored. After performing the spatial filter, an inverse spectral DCT is applied on all transformed bands including the filtered bands to obtain the final preprocessed hyperspectral data, which is subsequently fed into a linear SVM classifier.
　　The outcome of this dissertation advances the state-of-the-art performance by proposing novel simple and effective strategies for accurate hyperspectral image classification, where the results obtained by extensive experimentation on three real hyperspectral datasets confirmed their effectiveness in terms of classification accuracy, the sensitivity regarding different sizes of the training samples, and computational time.
　　The main advantage of the proposed strategies is the tuning simplicity along with a reasonable tradeoff between accuracy and computational time. Thus, it will be quite useful for applications such as risk management, which requires a fast response. Moreover, as the spectral filters are performed on each pixel, and the spatial filter are performed on each band, these strategies are easily parallelized, which could further reduce the computational time.