Mazharul Hossain

Hyperspectral (HS) sensing provides detailed high-dimensional spectral data to identify subtle anomalies and material variations in complex scenes. HS anomaly detection (HS-AD) aims to identify small, spectrally distinct objects or materials in HS imagery without prior target information. However, most HS-AD methods assume that the training and test data are drawn from the same distribution (single-domain). This assumption is often violated in operational sensing because of differences in sensor characteristics, scene composition, atmospheric conditions, and acquisition geometry. To address this challenge, we propose Generalized MoE-AD (GMoE-AD), a neural Mixture-of-Experts (MoE) architecture for robust cross-domain hyperspectral anomaly detection. The framework fuses the outputs of six unsupervised base detectors with representations from a pretrained HS foundation model, combines four neural experts through learned top-2 routing, and applies gradient-reversal-based domain-adversarial training to improve robustness under distribution shifts. A drift-aware test-time adaptation (DTA) variant is evaluated separately. We evaluate GMoE-AD on six public real-world HS benchmark datasets—San-Diego, Salinas, HYDICE-Urban, ABU-Airport, ABU-Beach, and ABU-Urban—and one private Arizona dataset comprising 22 images acquired by four different sensors. Under all-domain training, the model uses the training portions of all seven datasets and is evaluated without access to domain identity or dataset-specific information. GMoE-AD achieves an average ROC-AUC of 0.943, PR-AUC of 0.623, and F1-macro of 0.821 in this setting. In a leave-one-dataset-out (LODO) evaluation, where the target dataset is completely unseen during training, the model maintains an average ROC-AUC of 0.910 and F1-macro of 0.773. The optional DTA variant increases mean ROC-AUC from 0.938 to 0.953 but reduces F1-macro from 0.822 to 0.815, indicating a metric- and dataset-dependent adaptation trade-off. These results suggest that combining neural expert routing, transfer learning, and domain-adversarial representation learning improves robustness across heterogeneous HS datasets and provides a unified approach to anomaly detection in high-dimensional sensor data, supporting deployment in real-world hyperspectral applications where training and deployment conditions differ.

Accurate crop mapping—identifying both the location and types of crops—is crucial for effective agricultural planning and informed decision-making. Advances in remote sensing, notably hyperspectral imagery from unmanned aerial vehicles (UAV), greatly enhance the efficiency and accuracy of crop mapping, reducing the reliance on traditional, labor-intensive field surveys. However, applying deep classifiers directly to hyperspectral data can lead to overfitting. Conversely, deep semantic segmentation models may struggle due to limited annotated hyperspectral imagery. To address this problem, we propose enhancing a U-Net-style model—originally trained on RGB imagery—by incorporating task adaptation, a custom loss function, and a spectral attention mechanism to better optimize it for hyperspectral data and improve crop mapping performance. Our proposed segmentation network achieved 76.6% accuracy and a 74.9% Dice score on a UAV-acquired hyperspectral agricultural dataset, which is competitive and well-rounded compared to other state-of-the-art methods while offering significantly lower computational complexity.

Hyperspectral images include information from a wide range of spectral bands deemed valuable for computer vision applications in various domains such as agriculture, surveillance, and reconnaissance. Anomaly detection in hyperspectral images has proven to be a crucial component of change and abnormality identification, enabling improved decision-making across various applications. These abnormalities/anomalies can be detected using background estimation techniques that do not require the prior knowledge of outliers. However, each hyperspectral anomaly detection (HS-AD) algorithm models the background differently. These different assumptions may fail to consider all the background constraints in various scenarios. We have developed a new approach called Greedy Ensemble Anomaly Detection (GE-AD) to address this shortcoming. It includes a greedy search algorithm to systematically determine the suitable base models from HS-AD algorithms and hyperspectral unmixing for the first stage of a stacking ensemble and employs a supervised classifier in the second stage of a stacking ensemble. It helps researchers with limited knowledge of the suitability of the HS-AD algorithms for the application scenarios to select the best methods automatically. Our evaluation shows that the proposed method achieves a higher average F1-macro score with statistical significance compared to the other individual methods used in the ensemble. This is validated on multiple datasets, including the Airport–Beach–Urban (ABU) dataset, the San Diego dataset, the Salinas dataset, the Hydice Urban dataset, and the Arizona dataset. The evaluation using the airport scenes from the ABU dataset shows that GE-AD achieves a 14.97% higher average F1-macro score than our previous method (HUE-AD), at least 17.19% higher than the individual methods used in the ensemble, and at least 28.53% higher than the other state-of-the-art ensemble anomaly detection algorithms. As using the combination of greedy algorithm and stacking ensemble to automatically select suitable base models and associated weights have not been widely explored in hyperspectral anomaly detection, we believe that our work will expand the knowledge in this research area and contribute to the wider application of this approach.

Complex machine learning (ML) algorithms consistently outperform traditional ML models, yielding significantly improved results. Furthermore, these complex system uses large amounts of data with many features to train, increasing their predictive power but reducing our ability to explain them. Thus, complex ML models primarily operate like black boxes. Explaining their inference processes and predictions poses a challenge due to the necessity to account for millions of interacting weights in a complicated way. Still, we need to clarify their inner workings and procedures, as understanding a model can improve transparency and promote trustworthiness. Therefore, ‘Explainability’ assumes pivotal importance, as it enhances accuracy and broadens the scope of applicability for complex ML methods within critical domains. Thus, in this review paper, we have investigated the complete ML system development lifecycle, compiled recommendations to cover explainability gaps from diverse literature sources, listed different methods of interpretability and explainability, and improved our explainability-augmented ML system development life cycle for all stakeholders to enhance trustworthiness.

Machine Learning (ML) systems are currently used widely. While the requirements elicitation and management processes are well understood and standardized for software systems, they have not been adopted for ML systems. These processes pose additional complexity due to the multitude of stakeholders, their understanding of ML concepts, and their varied requirement orientations. Communication among the stakeholders is a must to generate an appropriate requirements document to guide the development of the system. Another aspect of ML systems is their black box mode of operation. It is essential to know the reasoning behind the predictions of these systems, especially for critical applications. These explainability requirements also must be taken into consideration. This paper reviews state of the art in these two areas to form a basis for developing a framework for stakeholder interaction to facilitate the requirement management of ML systems.

In medical and microscopy imaging applications where the object is not directly visible, images are never identical to the ground truth. In three-dimensional structured illumination microscopy (3D-SIM), acquired images taken from the object have limited resolution due to the the point spread function (PSF) of the imaging system. Additionally, due to the data acquisition process, images taken under low light and in the presence of electrooptical noise can have a low signal-to-noise ratio as well as suffer from other undesirable aberrations. To obtain a high-resolution restored image, the data must be digitally processed. The inverse imaging problem in 3D-SIM has been solved using various computational imaging techniques. Traditional model-based computational approaches can result in image artifacts due to required, yet not accurately known system parameters. Furthermore, some iterative computational imaging methods can be computationally intensive. Deep learning (DL) approaches, as opposed to traditional image restoration methods, can tackle the issue without access to the analytical model. Although some are effective, they are biased since they do not use the 3D-SIM model. This research aims to provide an unrolled physics-informed (UPI) generative adversarial network (UPIGAN) for reconstructing 3D-SIM images utilizing data samples of mitochondria from a 3D-SIM system. This design uses the benefits of physics knowledge in the unrolling step. Moreover, the GAN employs a Residual Channel Attention super-resolution deep neural network (DNN) in its generator architecture. The results from both a qualitative and quantitative comparison, present a positive impact on the reconstruction when the UPI term is used in the GAN versus using the GAN architecture without it.

Research supporting improved anomaly detection performance benefits a wide range of technical applications, and thus, the definition of what anomalies are and the subsequent means to detect them are wide ranging. In this treatment, an overview of the development of an anomaly detection approach based on spectral signatures obtained with hyperspectral unmixing is presented. The algorithm is designed to address some of the shortcomings of current techniques whose functionality is dependent upon normalized differences between discrete frequencies or spectral components, or those based on estimated distances between background spectra and pixels under test. Details about the extracted endmembers and their use for effective anomaly detection will be presented as well as, some thoughts on the expected requirements for future machine learning based implementations.

We introduce a method for enhancing the resolution of Structured Illumination Microscopy images. It uses a deep prior network, unrolled with a model-based analytical process to improve performance, maintain structural details, and guarantee the theory.