Building intelligent systems at the seam of research and product.

Hyperspectral Image Classification is an important topic in remote sensing with numerous applications in precision agriculture, environmental monitoring, defence, and more. It helps classify material composition by mapping spectral signatures from each pixel. Despite its potential, it faces persistent challenges due to the scarcity of labelled data and high dimensionality. This makes it difficult to achieve robust performance that can generalize well. To address this problem, we propose SRSSF (Stability-Reliability Spectral-Spatial Fusion), a hybrid mixture-of-experts framework for hyperspectral image (HSI) classification. It fuses a compact, stability-selected spectral representation with full-spectrum spatial and joint spectral-spatial descriptors (extended morphological, attribute-like, and Gabor texture profiles), combined through validation-derived reliability weights rather than fixed or heuristic fusion rules. SRSSF is evaluated under a strict, footprint-level spatially disjoint protocol in which entire patch receptive fields are guaranteed non-overlapping across training, validation, and test partitions, not mere center pixels. It has an automated, geometry-only (label- and accuracy-independent) seed-acceptance procedure and requires a minimum dataset-specific test-set size for training. The framework is benchmarked against fourteen classical and CNN spectral/spatial baselines and a dedicated ablation suite, across four benchmark datasets (Indian Pines, Pavia University, Salinas, and the Kennedy Space Center (KSC)). Our framework achieves the best OA, AA, and κ on every dataset, with OA scores of 56.43,91.94,90.50, and 75.96 at Indian-Pines, Pavia University, Salinas, and KSC, respectively.

Images are one of the key sources of visual information and communication. It plays a crucial role in defense, AI, and forensic science, among others. However, it is prone to corruption from various types of noise from varying sources, mainly during acquisition and transmission. It creates artifacts or signal distortion due to statistical variance in pixel-value measurements that affect contrast, color, and other aspects. Several denoising techniques exist, and many have been proposed to address it, but their performance is debatable. Noise leads to the loss of critical information, primarily edges and corners, which negatively affects performance. And there is no single, universal perfect solution to this problem. This paper reviews the existing techniques and analyzes the performance of three main techniques, namely Gaussian, linear, and non-linear isotropic smoothing. After careful examination, it is found that both Linear and Non-Linear smoothing can be an effective solution.

Geoscientific models are simplified representations of complex earth and environmental systems (EESs). Compared with physics-based numerical models, data-driven modeling has gained popularity due mainly to data proliferation in EESs and the ability to perform prediction without requiring explicit mathematical representation of complex biophysical processes. However, because of the black-box nature of data-driven models, their performance cannot be guaranteed. To address this issue, we developed a generalizable framework for improving the efficiency and effectiveness of model training and the reduction of model overfitting. This framework consists of two parts: hyperparameter selection based on Sobol global sensitivity analysis and hyperparameter tuning using a Bayesian optimization approach. We demonstrated the framework efficacy through a case study of daily edge-of-field (EOF) runoff predictions by a tree-based data-driven model using the extreme gradient boosting (XGBoost) algorithm in the Maumee domain, USA. This framework contributes towards improving the performance of a variety of data-driven models and can thus help promote their applications in EESs.

As surface water resources in the U.S. continue to be pressured by excess nutrients carried by agricultural runoff, the need to assess runoff risk at the field scale continues to grow in importance. Most landscape hydrologic models developed at regional scales have limited applicability at finer spatial scales. Hybrid models can be used to address the scale mismatch between model simulation and applicability, but could be limited by their ability to generalize over a large domain with heterogeneous hydrologic characteristics. To assist the generalization, we develop a regionalization approach based on the principal component analysis and K-means clustering to identify the clusters with similar runoff potential over the Great Lakes region. For each cluster, hybrid models are developed by combining National Oceanic and Atmospheric Administration's National Water Model and a data-driven model, eXtreme gradient boosting with field-scale measurements, enabling prediction of daily runoff risk level at the field scale over the entire region.