Spatial-spectral local discriminant projection for dimensionality reduction of hyperspectral image

Statistical retrieval of atmospheric profiles with deep convolutional neural networks

“…We have shown that there is a high potential for models that incorporate feature extracting abilities as well as capabilities of modeling non-linear phenomena in statistical retrieval. Exploring a greater variety of dimensionality reduction techniques, such as the spatio-spectral supervised technique presented in [42], or incorporating it into the CNN architecture is also an area to further explore. Finding optimal architectures for CNNs remains an open task in the deep learning literature, and due to the non-convexity of the problem, experiments are the only way to find optimal models.…”

Deep Fully Convolutional Embedding Networks for Hyperspectral Images Dimensionality Reduction

“…To prove the effectiveness, DFCEN is compared with several dimensionality reduction algorithms, such as LE [11], LLE [11], SAE, spatial-domain local pixel NPE (LPNPE) [38], spatial and spectral regularized local discriminant embedding (SSRLDE) [38], SSMRPE [39], spatial-spectral local discriminant projection (SSLDP) [40]. The former three methods are spectral-based methods while the latter four approaches make use of both spatial and spectral information for dimensionality reduction of HSIs.…”

“…SSMRPE [39] shares the same DR concept as LLE. SSLDP [40] designs a weighted within neighborhood scatter to reveal the similarity of spatial neighbors. Among them, SSRLDE [38] and SSLDP [40] are supervised and require class labels to implement dimensionality reduction, while others are unsupervised.…”

“…For the fairness of the experimental comparison, the numbers of the nearest neighbor samples k of LE and LLE are the same as that of DFCEN_LE and DFCEN_LLE in the following experiments. We also choose the optimal parameters of their source literature for LPNPE [38], SSRLDE [38], SSMRPE [39], SSLDP [40]. In all the experiments below, all algorithms including DFCEN use raw data (that is not filtered to de-noise and smoothen pixels).…”

“…As we have seen, the spatial-spectral combined algorithm, LPNPE [38], SSRLDE [38], SSMRPE [39], SSLDP [40] and DFCEN, are superior to the spectral-based algorithm, LE, LLE and SAE, which indicates that spatial features are beneficial to the dimensionality reduction of HSIs. Neural network based methods, SAE and DFCEN, are superior to traditional dimensionality reduction algorithms, which testifies that neural network is suitable for dimensionality reduction of HSIs.…”