Vegetation Index (VI) curves, derived from multi-temporal satellite images, are being widely employed to model the crop-specific phenological events. The current study analyzed a novel approach to mitigate the effect of violating the Independent and identically distributed (i.i.d) assumption in classifying the VI curves. Even though deep learning-based classification methods have produced cutting-edge outcomes, the correlation of spatially adjacent samples is not generally considered. The proposed approach dynamically transformed the VI curves to a graph representation, where the nodes correspond to the curves. Graph convolutional operations along with Kolmogorov–Arnold Network (KAN) were then used to learn the embedded representations, based on the labeled samples in the proximity. The collaborative learning of graph-formulation and classification facilitated the consideration of non-i.i.d nature of the VI curve samples. The proposed and benchmark methods were analyzed using the VI curves collected over three farms, covering multiple crops including wheat, barley, and potato crops. The use of similarity computation based on Dynamic Time Warping and interpolated convolution, in addition to the consideration of sample correlation, resulted in significant accuracy improvement as compared to the baseline approaches.

Hyperspectral video tracking poses unique challenges due to the high dimensionality of spectral data and the limited capacity to capture discriminative texture information. To address this, we propose a novel tracking framework that integrates spectral-spatial similarity reduction with reorganized texture encoding for robust hyperspectral target tracking. Specifically, we introduce a dimensionality compression strategy that converts the multi-band hyperspectral input into a representative grayscale image, preserving key spectral-spatial cues. To enhance discriminative texture modeling, a 3D Gabor filter is applied to the search region, and the extracted responses are adaptively fused based on their local variance. The resulting texture representations are selectively masked to suppress background noise and are then passed into a correlation filter module for precise target localization. Furthermore, we design a template update mechanism that mitigates model drift and cumulative errors during tracking. Extensive experiments on public hyperspectral video benchmarks demonstrate that our method achieves competitive performance against state-of-the-art hyperspectral trackers, especially in scenarios with background clutter.

Monitoring and assessment of critical infrastructure, such as urban roadways, are essential for the overall economy. Roads and bridges are prone to subsurface deformations, leading to significant economic losses and casualties. Ground-penetrating radar (GPR) is widely used for its nondestructive testing capabilities to detect subsurface anomalies. However, acquiring sufficient training data poses a challenge for advancing deep learning applications in subsurface target detection. To address this, a combined image-guided generative adversarial network (CIGGAN) is proposed for generating GPR data with voids by combining various voids and backgrounds. CIGGAN enhances feature diversity by extracting features from combined images and fusing features, creating GPR data significantly different from the original. An evaluation criterion is also proposed for assessing the quality of generated images. This study employs two real GPR datasets from a GPR vehicle-mounted system to evaluate the performance of CIGGAN in GPR data generation. Additionally, two state-of-the-art (SOTA) detection models (Faster-RCNN and Retinanet) are used to test the effectiveness of CIGGAN-generated data for void detection. Results show that CIGGAN has robust generalization capabilities, adapting well to generating GPR data with a small sample size (approximately 100–200 images). Using CIGGAN-generated data, in addition to the original dataset, improved the${F}1$scores on the first dataset by 5.82% and 9.22% for the first and second models, respectively. Similarly, the approach improved the${F}1$score on the second dataset by 3.62% and 2.48% for the first and second models, respectively. Experiments indicate that CIGGAN is a powerful tool for supporting deep learning in the GPR domain.

Hyperspectral data from the Compact Reconnaissance Imaging Spectrometer for Mars (CRISM) have proven instrumental in analysing the mineralogy of the Martian surface and advanced our understanding of the geological history and habitability of Mars. Recently, machine learning-based methods have been used to analyse CRISM data. However, these methods exhibit limitations such as training difficulties and the need for manual feature selection. Therefore, in the contribution we propose a novel algorithm which combines the Random Forest algorithm with Extreme Value Analysis to classify CRISM spectra under an open-set regime. The algorithm’s effectiveness is demonstrated using ∼ 470 000 labelled spectra from the CRISM machine learning toolkit’s mineral dataset where it returns an accuracy of 86.92 % and a Kappa(κ) of 0.85

Although hyperspectral sensors are highly effective in mapping the minerals, the intimate nonlinear mixing and resolution tradeoff affect their effectiveness. In this regard, this study proposes a graph-based spectral unmixing strategy. The proposed approach leverages the advantages of both graph-based and deep learning based approaches. Additionally, the current study is a pioneer approach of using the graph-based approach for spectral unmixing. The spectral and spatial latent manifolds of the input patches are learned, and this information along with the endmember prior is used to formulate a graph-based representation. Further graph convolution approach is used to soft classify the spectra yielding fractional abundances. The results of the proposed approach on standard, synthetic and real-world data indicates that the proposed approach performs better than the state-of-the-art unmixing approaches. Moreover, the graph-based representations make the approach interpretable and facilitate the consideration of the spatial autocorrelation.

Modeling spectral mixing for geological mixtures is challenging. The endmembers in these mixtures interact at a microscopic scale resulting in nonlinear mixing. Prior to applying inversion techniques for spectral unmixing, it is essential to identify the nature of nonlinear mixing that would facilitate a faster analysis of hyperspectral images for geological mixtures. This paper attempts pixel-wise nonlinearity detection of a hyperspectral image of a geological mixture (rock sample) collected in a controlled environment in the laboratory. The identified nonlinearly mixed regions were mapped and further validated through their spectral features in the principal component space. The insights obtained in this study would further support identifying the nature of mixing in geological mixtures.

In order to effectively utilize spectral and object spatial information to improve tracking performance, we design an Hyperspectral Video (HSV) tracker, namely DSP-Net, to integrate the various prior information. The gradient difference between spectral vectors is explored to develop a clustering technique. The approach generates a binary mask containing target spectral information and appearance clues. A feature cache is introduced to store historical information. Additionally, the channel shift operation is used on the timing to capture the trajectory clues of the target. With the help of the non-local mechanism, the trajectory clues, appearance clues and spectral information of the target are finally integrated, using the designed spectral-spatial joint perception module to enhance the expression of the target. Experimental results show that DSP-Net outperforms state-of-the-art HSV trackers on existing dataset.

Single-image super-resolution (SISR) techniques attempt to reconstruct the finer resolution version of a given image from its coarser version. In the SISR of hyperspectral data sets, the simultaneous consideration of spectral bands is crucial for ensuring the spectral fidelity. However, the high spectral resolution of these data sets affects the performance of conventional approaches. This research proposes the design of 3-D convolutional neural network (CNN)-based SISR architectures that can map the spatial-spectral characteristics of hypercubes to a finer spatial resolution. The proposed approaches facilitate the simultaneous optimization of sparse codes and dictionaries with regard to the super-resolution objective. Our main hypothesis is that the consideration of spectral aspects is essential for the spatial enhancement of hyperspectral images. Also, we propose that the regularized deconvolution of a coarser-scale hypercube, using learned 3-D filters, yields the required high-resolution version. Based on these hypotheses, a convolution-deconvolution framework is proposed to super-resolve the hypercubes in parallel with the reconstruction of a set of regularizing features. Novel sparse code optimization sub-networks proposed in this article give better performance than the existing strategies. The endmember similarities and hyperspectral image prior are considered while designing the proposed loss functions. In order to improve the generalizability, a collaborative spectral unmixing strategy is employed to refine the spectral base of the super-resolved result. The spatial-spectral accuracy of the super-resolved hypercubes, in terms of the validity of regularizing features and endmembers, is explored to devise an optimal ensemble strategy. The experiments, over different data sets, confirm better accuracy of the proposed frameworks compared to the prominent approaches.

In this paper, we propose the use of network inversion for modeling the variation of class distributions with scale. Unlike the state of the art methods that predict the mapping between coarser and finer scale patches without considering the distributions at coarser scale, our approach uses coarser scale features for effective reconstruction. This is the pioneer work of using network inversion for the purpose. Analysis over the proposed framework reveals that both the computational performance and accuracy varies with the depth of the network as well as the size and number of filters in each layer. Also the performance of the approach has been found to improve with the increase in the number of input feature maps. Investigations over standard datasets indicate that the proposed approach performs much better than the recent sub-pixel classification as well as super resolution techniques.

This letter investigates the use of coarse-image features for predicting class labels at a given finer spatial scale. In this regard, two unsupervised subpixel mapping approaches, a semivariogram method, and a pixel-affinity based method are proposed. Furthermore, segmentation-based spectral unmixing is explored so as to address the spectral variability and nonconvexity of classes. In addition, the gradient information is employed to resolve uncertainties in the unmixing process. The proposed modifications based on pixel-affinity and semivariogram have produced an accuracy improvement of 5% or more over the state-of-the-art approaches.

Spectral unmixing techniques decompose the pixels into constituent fractions in order to extract the subpixel information. This study reviews spectral unmixing techniques from a perspective different from earlier approaches in that the problem is studied from a classification as well as clustering perspective. In this research, we focus on addressing some core issues of spectral unmixing such as endmember variability, requirement of pure endmember values, and initialization sensitivity modelling. We propose a Support Vector Machine (SVM) based unmixing technique that incorporates endmember spectral variability. The method uses endmember extraction techniques to give optimal performance even in the absence of training samples. Further, our study presents an alternation of FCM based method for incorporating spectral variability, and the approach is found to be resilient to the brightness variation. An automatic approach for fuzziness parameter selection is also introduced. The sensitivity of FCM towards endmember initialization has been considerably reduced by optimizing the initial seed selection. The proposed approaches have been analyzed over various standard datasets.

Super-resolution techniques use subpixel information to predict high resolution classification maps from coarse images. This study investigates for an unsupervised super-resolution approach which considers the image features to predict target spatial dependencies. Novelty of the approach is that the convolution neural networks and deep autoencoders are explored in this context. Evaluation over standard datasets revealed that the proposed method is more effective than the state of art unsupervised approaches. The method is also found to be preferable over variogram based approaches for complex scenes. This study also compares the effectiveness of shallow and deep networks and investigates the possible assessment of the optimal depth for the learning network. This technique can be further extended to a supervised framework.