Underwater hyperspectral imaging systems are essential tools for observing marine geology, seabed minerals, and marine reptiles. However, the complex underwater imaging environment and the high cost of equipment present unpredictable challenges for underwater hyperspectral image acquisition. Consequently, this limits the widespread application of hyperspectral images in underwater survey missions. To address these challenges, this study first explores spectral reconstruction technology to indirectly assist in acquiring underwater hyperspectral images through data-driven deep learning methods. Specifically, a Multi-Scale Large Kernel Spectral Reconstruction Network (MSL-SRN) was designed, which leverages available public RGB and hyperspectral image pairs to train the network to learn hyperspectral information from a single underwater RGB image. To validate the feasibility and effectiveness of this approach, spectral reconstruction experiments were conducted on a real underwater hyperspectral dataset. Experimental results indicate that the proposed method closely approximates underwater hyperspectral images. The concept presented in this study provides a novel approach to overcoming the cost and accessibility limitations of underwater hyperspectral imaging.

Spectral super-resolution technology, which reconstructs 31-band hyper-spectral images from RGB natural scene images within the 400-700nm bands, has seen rapid growth. However, its fixed spectral resolution and spectral coverage limit its application in remote sensing imaging, particularly for aerial images with multi-band information. The lack of corresponding high-spectral image pairs has hindered research progress, leaving the potential spectral information of these remote sensing images untapped. In this study, we explore a hybrid transformer architecture for multispectral images that carry visible light and near-infrared informations to achieve spectral super-resolution. This network integrates both intra-row and intra-column attention mechanisms, along with a cross inter-row and inter-column attention mechanism, to precisely capture and process the spatial and spectral features in spectral images. In the case of two simulated datasets, the experimental results demonstrate favorable outcomes. In classification experiments using multimodal Pavia University datasets, the reconstructed hyper-spectral images exhibit superior performance with higher average accuracy (95.30%), overall accuracy (95.70%), and Kappa coefficient (93.50%).

Cross-domain few-shot learning (FSL) often employs adversarial domain adaptation techniques to address the issue of data distribution discrepancies between the source and target domains. However, forcing the alignment of two distinct domains may lead to distortions in class distribution alignment and result in a decrease in classification performance in hyperspectral image analysis. Moreover, existing cross-domain methods are often applied to satellite/airborne hyperspectral image as both the source and target domain. It is rarely explored whether the same cross-domain methods can be applied for cross applications where the source domain and target domain data could be both satellite/airborne hyperspectral image with lower spatial resolution and unmanned aerial vehicle (UAV) hyperspectral image with higher spatial resolution. To address these issues, this paper proposes a novel domain-adaptive FSL network with dual branches respectively aiming at domain fusion and domain separation. The domain fusion branch uses a conditional adversarial network to align the global distributions of the two domains, while the domain separation branch introduces gate mechanism for discriminative feature learning in each domain to achieve independent category distributions. During the experiment, the proposed method is evaluated by performing cross-transfer learning under the condition that low spatial resolution hyperspectral data and high spatial resolution hyperspectral data are used as source and target data alternately. The experimental results suggest that the proposed method not only mitigates the negative effects of forced alignment in domain fusion but also holds potential for cross-domain transfer learning between low and high spatial resolution hyperspectral images.

The inversion problem of partial differential equation plays a crucial role in cyber–physical systems applications. This article presents a novel deep learning optimization approach to constructing a solver of heat equation inversion. To improve the computational efficiency in large-scale industrial applications, data and model parallelisms are incorporated on a platform of multiple GPUs. The advanced Ring-AllReduce architecture is harnessed to achieve an acceleration ratio of 3.46. Then, a new multi-GPUs distributed optimization method GradReduce is proposed based on Ring-AllReduce architecture. This method optimizes the original data communication mechanism based on mechanical time and frequency by introducing the gradient transmission scheme solved by linear programming. The experimental results show that the proposed method can achieve an acceleration ratio of 3.84 on a heterogeneous system platform with two CPUs and four GPUs.

Monitoring plant growth is essential in modern agriculture to guarantee productivity. Since manual measurement of plant characteristics is laborious and expensive, automatic measures are desirable. This can be accomplished by methods such as vision-based structure from motion (SFM) to obtain the 3-D information of a plant. An SFM method based on binocular vision is here developed to acquire the physical parameters of plants. In this method, image sequences are captured by a binocular camera from multiple views of the target plant to improve the effectiveness and simplify the implementation. The spatial relationships between adjacent images are estimated through image feature extraction and matching. A disparity map is then built and the 3-D coordinate of each image pixel is obtained by applying stereo-vision. The connected coordinates then constitute the 3-D model of the plant. By doing so, plant structure parameters, such as height, canopy size, and trunk diameter, can be derived from the 3-D model. Experimental results show that the measured plant height, the canopy width, and the trunk diameter of the target plant are within an acceptable accuracy at the millimeter level, and the mean errors of the measured sizes are all less than 2%. This demonstrates the potential value of the proposed method for online growth monitoring of agricultural plants.

Monitoring crop growth is of great significance to obtain crop growth status information for development of smart agriculture. The traditional way to measure the phenotypic parameters of crops is labor-intensive and encounters inconvenient operations. In this study, we propose to obtain the phenotypic parameters of crops from 3-D reconstruction of plants from single RGB images using a data-driven plant phenotypic parameters estimation network (P3ES-Net) deep neural network, which enables to estimate the depth shift and camera focal length used for depth estimation and reconstruction of the 3-D model of plants. Based on the principles of the monocular ranging and pinhole imaging model, crop phenotypic parameters such as height, canopy size, and trunk diameter can then be calculated from the 3-D model. Experiments with four practical plants present that our method is able to achieve acceptable evaluation of the growth status of plants. Of more significance, it achieves particular superior depth estimation performance over a commercial depth camera, which is a very new on-sale depth camera using stereo vision and deep learning network. This potential performance throws light on the low-cost measurement of crop phenotypic parameters using RGB camera in monitoring crop growth.

Hashing has been widely studied for cross-modal retrieval due to its promising efficiency and effectiveness in massive data analysis. However, most existing supervised hashing has the limitations of inefficiency for very large-scale search and intractable discrete constraint for hash codes learning. In this paper, we propose a new supervised hashing method, namely, Discrete Semantic Matrix Factorization Hashing (DSMFH), for cross-modal retrieval. First, we conduct the matrix factorization via directly utilizing the available label information to obtain a latent representation, so that both the inter-modality and intra-modality similarities are well preserved. Then, we simultaneously learn the discriminative hash codes and corresponding hash functions by deriving the matrix factorization into a discrete optimization. Finally, we adopt an alternatively iterative procedure to efficiently optimize the matrix factorization and discrete learning. Extensive experimental results on three widely used image-tag databases demonstrate the superiority of the DSMFH over state-of-the-art cross-modal hashing methods.

The abundance of a mixed pixel of certain class can be understood as to get the relative score referring to the pure representative of this class, while not be classified with two absolute and discrete value as ”lor 0”. This is in accordance with the Relative Attribute Learning (RAL) problem in computer vision. In RAL, the concept of “relative attribute” is used to describe the belonging level of an obj ect to certain class with a score which is achieved from a learn-to-rank problem using rankSVM framework. To utilize information between data samples and even of mixed pixels, Relative Attribute based Unmixing (RAU) is proposed first time by using relative attribute to describe the abundance of mixed pixel as relative purity of certain class and learn the abundance with rankSVM. The mixed data sample are used to construct training comparisons set in rankSVM with archetypes generated by the reported Kernel Archetypal Analysis (KAA) unmixing method. In addition, spectral variability is also addressed by constructing comparisons set with synonyms spectrum achieved from KAA. Experiments on both synthetic and real hyperspectral mixed image have demonstrated the potential value of proposed method for mixed pixel analysis.

Multilayer learning has shown promising performance in machine learning due to its strong capacity in data representation. The core idea is associated with the generation and the use of high-level features for task learning. Motivated by visual and brain learning mechanism, differential abundance features are proposed in this study as high-level features for further unmixing to refine obtained unmixing results. Kernel archetypal analysis (KAA) shows potential value for spectral unmixing which realizes endmember extraction and abundances estimation simultaneously. In this paper, the abundances after bilateral filtering which incorporates influence of both spatial and spectral information of neighbor pixels are used for further unmixing by cascade KAA to improve abundance estimation. The developed method was tested on both synthetic and real hyperspectral image data which demonstrate that multilayer unmxing with the filtering abundance features outperforms the conventional unmixing approach.

The multilayer network in deep learning provides a promising means for rich data representation. Inspired by this approach, we investigate multilayer unmixing for spectral decomposition with fast kernel archetypal analysis (KAA). KAA is used for endmember extraction and abundance estimation simultaneously. To refine the initial unmixing results, a multilayer process is utilized to provide final unmixing results at the end of the network. Moreover, a fast implementation of KAA is proposed via using the Nyström method to relieve KAA's memory issue and decrease the processing time. The proposed method is tested on both synthetic and real hyperspectral image data sets. The results demonstrate that the multilayer unmixing algorithm outperforms the conventional unmixing techniques.

Relative attributes provide accurate information for image processing to describe which image is more natural, more open, etc. Robustness of relative attribute learning depends on the labeled comparative image pairs. However, manually labeling is a labor intensive and time-consuming task. In this paper, a semi-supervised learning approach based on group sparse is proposed to discover pairwise comparisons automatically. We generate an initial level division of the labeled training images for the basic of new constraints. Then, group sparse representation for the unlabeled images is introduced by embedding the level information into the dictionary. The semi-supervised process is conducted by selecting samples which have minimum reconstruction errors and adding new constraints to the model by comparing the selected ones with the samples in dictionary. Experiments on three public datasets demonstrate the effectiveness of our proposed method.

Conventional methods for mixed pixel analysis have their limitations in performance when the scenario is highly mixed without pure endmembers or only virtual endmembers can be generated. Moreover, theses approaches do not address the endmember variability. In this study, a multiple endmembers extraction algorithm based on Archetypal Analysis (AA) is proposed to solve the above problems. AA aims at finding distinct patterns in the data and thus, is suitable for endmember extraction. It can also generate vitual pure archetypes when no pure samples exist in the data. Kernel version of AA is investigated for multiple endmember extraction. Informative samples which contribute to the generation of each endmember class can be extracted and used as the multiple endmembers of a single ground cover type. Experimental results show that the multiple endmembers unmixing method using Kernal AA achieves more realistic unmixing results than single endmember based unmixing.