Engineer Research and Development Center News Releases

Multisource Knowledge Graph Architecture for Air-Gapped AI Systems: Design Patterns and a Geospatial Reference Implementation

Press Operations — Wed, 10 Jun 2026 12:53:00 GMT

Purpose: This technical note presents a reference architecture for constructing multisource knowledge graphs in air-gapped, domain-specific AI systems. The architecture addresses recurring problems in military, intelligence, and secure-enterprise environments: integrating heterogeneous authoritative data sources while preserving source schema fidelity, enabling deterministic semantic resolution, and operating without external network dependencies. While this is a geospatial implementation, the architectural principles in the core design are transferable. The implementation is organized around four separable layers—schema registry, domain ontology, reasoning patterns, and relationship vocabulary—and adopts a canonical-with-aliasing integration strategy that supports cross-schema reasoning without forcing premature schema con-vergence. These patterns are validated through a geospatial intelligence implementation supporting US Army operations and demonstrate how abstract design principles translate into an operationally relevant system. This knowledge graph is designed to serve as the semantic substrate for router-based AI systems. A companion technical note (Drouillard and Lewis 2026) describes the geospatial AI (GeoAI) agent stack, which is a router-based orchestration architecture that coordinates multiple retrieval backends and reasoning tools. Within that architecture, the knowledge graph functions as a specialized retrieval backend that runs alongside document retrieval and vector search, serving queries that require structured entity-relationship reasoning, provenance tracking, or deterministic semantic resolution. The router directs spatial relationship queries (e.g., “which roads cross this river”), multihop dependency queries (e.g., “what infrastructure depends on this power station”), and schema-resolution queries (e.g., “find all transportation features in Multinational Geospatial Co-production Program [MGCP] format”) to the knowledge graph while routing conceptual or analytical questions to document retrieval. This technical note focuses exclusively on the knowledge graph architecture; the broader orchestration patterns and routing logic are detailed in the companion paper. While the examples presented are geospatial, the architectural principles, validation strategies, and design tradeoffs documented here have broader applicability where deterministic semantic integration is required under air-gapped constraints. The geospatial instantiation serves as a concrete demonstration of abstract patterns that may inform future knowledge graph efforts in other US Army Engineer Research and Development Center (ERDC) research domains.

Processing and Optimization of Global Land Ice Measurements from Space (GLIMS) Glacier Polygon Shapefiles for Army Geospatial Data Model Integration

Press Operations — Wed, 13 May 2026 14:43:00 GMT

Abstract: This technical note documents the methodology used to prepare glacier polygon datasets from the Global Land Ice Measurements from Space (GLIMS) database for integration into Army geospatial workflows. The Army Geospatial Data Model contains a feature class within the GGDM (Ground-Warfighter Geospatial Data Model) for permanent snow, defined operationally as snow persisting on the ground for more than two years. However, in cryospheric science, snow that persists across multiple accumulation seasons transitions into firn and ultimately becomes glacial ice. Thus, most “permanent snow” surfaces are more accurately classified as permanent ice, and GGDM does not currently contain a dedicated feature class representing this land-surface category. The GLIMS database provides authoritative, globally maintained glacier and perennial ice extents, making it ideally suited to fill this structural gap in the GGDM schema. The purpose of this work is to (1) transform raw GLIMS glacier polygons into a clean, nonoverlapping, attribute-free dataset; (2) standardize the geometry for compatibility with GGDM; and (3) establish a US Army Engineer Research and Development Center (ERDC)–compliant workflow for maintaining a credible representation of global permanent ice surfaces.

Geospatial AI (GeoAI) Agent Stack: Router-Based Orchestration and Design Rationale

Press Operations — Wed, 13 May 2026 14:41:00 GMT

Abstract: This report summarizes the current state of a router-based, multiagent Geospatial AI (GeoAI) system designed to reliably execute geospatial workflows while retaining the flexibility of large language model (LLM) reasoning. The architecture is intentionally both language-model agnostic and orchestration-framework agnostic to support organizational controls and mandates, and it is designed to operate in air-gapped environments. It uses a domain router to scope tools before the model is invoked; a microrouter to decide whether the system should execute tools, retrieve knowledge, or produce a direct response; and a bounded cycle of execution and validation that supports multistep tool use. The design emphasizes determinism after the model makes decisions, strict boundaries around what the model can “see,” and modularity that keeps core business logic largely independent from orchestration and tool-protocol frameworks. The remainder of this report describes the architecture as implemented today, explains the design rationale, and outlines anticipated future work.

A Scalable Algorithm for Dynamic Vector Model Representation Utilizing Time-Series Reduction

Press Operations — Tue, 12 May 2026 18:29:00 GMT

Abstract: This document follows a technical report published by the US Army Engineer Research and Development Center–Geospatial Research Laboratory (ERDC-GRL), Time-Series Reduction for Dynamic Vector Model Attribute Representation in a Geographic Information System (ERDC/GRL TR-24-2, Drouillard and Lewis 2024). In that publication, we described the theoretical basis for extracting and modeling raster-format spatiotemporal phenomena for inclusion as a vector model attribute and provided a preliminary Python code example that was unsuitable for large-scale application. This report details the algorithm we subsequently developed to enable global-scale application of the time-series reduction method in service of the Intelligent Environmental Battlefield Awareness (IEBA) project.

Quantifying the Role of Vegetation on Urban Heat over Bengaluru, India

Press Operations — Thu, 07 May 2026 16:59:00 GMT

Abstract: The urban heat island (UHI) effect refers to how cities tend to be warmer than their non-urban surroundings, which increases the risk for heat-related illnesses and amplifies energy demands. Therefore, developing UHI mitigation strategies is crucial. Bengaluru, India has been rapidly urbanizing, but has yet to receive attention regarding potential UHI mitigation strategies. This work uses the Weather Research and Forecasting model with the single-layer urban canopy model to determine how UHI intensity changes in Bengaluru with perturbations of −10%, + 10%, + 20%, and + 30% in vegetation amount since recent work has shown that vegetation amount is the leading control of urban heat in Bengaluru. These perturbations illustrate how much the UHI could be amplified by near-depletion of vegetation or mitigated via realistic increases in vegetation. The simulations were investigated diurnally and during the dry and wet seasons. Results show that increases in vegetation were associated with a decrease in urban land surface temperature, an increase in the latent heat flux, and decreases in the sensible heat flux, and vice versa for a decrease in vegetation. Significant changes in UHI intensity usually occurred only when vegetation was increased by 20% or more. However, for the dry season nighttime, which exhibited the highest UHI intensity in the control run (1.70oC), the 10% increase in vegetation produced a significant decrease of − 0.19oC in UHI intensity, likely due to a shallow planetary boundary layer height. These results could have implications for mitigating urban heat, and reducing energy demands and public health risk in Bengaluru.

Multimethod Change-Detection Analysis Using Prithvi-EO-2.0: A Comparative Study of Traditional and Segmentation-Based Approaches for Vector Database Validation

Press Operations — Mon, 20 Apr 2026 19:07:00 GMT

Abstract: This technical note presents an evaluation of the performance of four change-detection methodologies, with a focus on validating and maintaining authoritative vector-feature databases using Earth observation data. In this study, we implemented traditional pixel-to-pixel change detection, feature-data-compliant segmentation, pixel-to-feature segmentation, and feature-to-pixel change detection, leveraging the Prithvi-EO-2.0 Vision Transformer model (Szwarcman et al. 2025), to analyze imagery from California’s Central Valley. The analysis of Sentinel-2 imagery from California’s Central Valley (in 2021–2023) demonstrated that there was a trade-off between sensitivity and reliability in the change-detection approaches: feature-to-feature methods achieved the highest sensitivity (0.637 average), while the feature-to-pixel approach provided the most reliable validation (0.280 average), exceeding the performance of traditional pixel-to-pixel methods (0.256 average).

Utilizing Laser Diffraction for Soil Particle Size Analysis

Press Operations — Mon, 20 Apr 2026 19:04:00 GMT

Abstract: This US Army Engineer Research and Development Center (ERDC) technical note (TN) describes the process and methodology for utilizing laser diffraction to analyze soil samples. The effort fulfills an Intelligent Environmental Battlefield Awareness (IEBA) project’s need to validate the performance of a global soil boundary mapping methodology that was developed as part of the Integration task. To validate the methodology, soil samples were classified by grain size into a texture class and compared against soil maps created for a given study area. The goal of this effort was to develop a repeatable standard operating procedure for the Horiba Partica LA-960V2, a laser diffraction particle size analyzer, that would allow rapid soil analysis to be conducted by individuals without a soil science background. The Horiba Partica has been used for soil particle size analysis, but it is not common in the field. Therefore, only limited documentation details the analysis protocol for the system. This TN will discuss the methodology used to analyze soil samples and the challenges encountered with the Horiba Partica.

Enhanced Spatial Resolution of Landsat Imagery Through Systematic Sensor Offset Exploitation: A Blended Pansharpening Approach

Press Operations — Tue, 03 Mar 2026 21:20:00 GMT

Purpose: This technical note presents a novel blended pansharpening methodology that exploits the systematic 7.5-meter (m) geometric offset between Landsat multispectral (MS) and panchromatic (pan) sensors to achieve selective spatial enhancement beyond conventional 15 m resolution limits. The approach creates a variable resolution product with an effective resolution of approximately 11.25 m and demonstrates superior spatial detail preservation in urban infrastructure while maintaining perfect spectral integrity.

The Use of Nitrocellulose Production Waste for Energy Generation

Press Operations — Tue, 03 Mar 2026 21:18:00 GMT

Abstract: The US Army Engineer Research and Development Center investigated the use of nitrocellulose (NC) fines, an ammunition waste, for energy generation. NC is a natural high polymer obtained from treating cotton or wool with nitric and sulfuric acid. It is widely used in the industry, with military applications being the largest use currently. Since military applications range from bullet propellants to missiles for tube munitions, large quantities must be produced to meet the demand. However, large NC production batches result in large quantities of NC fines waste, generated in the form of insoluble fibers in suspension in wastewater after manufacturing. Hence, a method to reuse this generated waste and convert it into energy was tested. This study evaluated the potential of creating energy from NC waste through hydrothermal liquefaction and gasification of NC, yielding methane (CH4) as the final product. Results demonstrated that the CH4 concentrations increased as the temperature, reaction time, and catalyst addition were increased, yielding a maximum concentration of 2,000 ppm (6,400 peak area of the chromatograph). The homogenous catalyst performed better than the heterogenous catalyst, since it increased the CH4 yield up to 6 times the concentration obtained with no catalyst added.

Relief Well Sustainment Deployable Resilient Installation Water Purification and Treatment System (RWS-DRIPS): Treatment of Relief Wells at Perry Dam, Kansas

Press Operations — Wed, 01 Oct 2025 15:14:00 GMT

Purpose: This report details the treatment process and resulting outcomes for relief wells at Perry Dam (Jefferson County, Kansas) using the Relief Well Sustainment Deployable Resilient Installation Water Purification and Treatment System (RWS-DRIPS) treatment trailer. The RWS-DRIPS is a mobile treatment unit with comprehensive water treatment capabilities designed to disinfect surface and subsurface water with high efficiency. Immediately following treatment with the RWS-DRIPS unit, video monitoring was used to observe the condition of the relief wells. The results of that observation are described in this report.

From Analog to Digital: A Systematic Workflow for Converting Published Landform Maps to Georeferenced Datasets

Press Operations — Thu, 25 Sep 2025 19:29:00 GMT

Abstract: Reference datasets for geomorphological analysis often require the integration of multiple data sources, including legacy maps and published figures that exist only as scanned images or hard copies. This report documents a systematic five-step workflow for converting landform information from these analog sources into georeferenced point datasets suitable for digital analysis. The methodology encompasses acquiring and evaluating imagery, georeferencing using ground control points, manually digitizing landform polygons, converting to centroid points using a systematic grid-based approach, and assigning attributes with quality control measures. In a case study on East Asia, we demonstrate the workflow’s practical application by processing 15 published sources to generate over 2 million labeled landform points representing approximately 1,015 km² of land across China and Mongolia. The dataset encompasses seven landform classes commonly found in arid environments: active washes, alluvial fans, bedrock, pediments, playas, sand dunes, and sand sheets. Quality assessments using analyst confidence ratings revealed reliable classification performance for most landform types. This workflow provides researchers with an efficient approach to leveraging existing published landform data, thus expanding the spatial coverage and temporal depth of reference datasets that are available for geomorphological analysis and machine learning applications.

Expansion of a Landform Reference Dataset in the Chihuahuan Desert for Dust Source Characterization Applications

Press Operations — Thu, 25 Sep 2025 19:26:00 GMT

Abstract: This report details the development of an extensive landform reference dataset for the Chihuahuan Desert region to support validation of a machine-learning-based landform classification model. Building upon previous work by Cook et al. (2022), we expanded both the quantity and spatial coverage of reference points to better represent the study domain’s geomorphic diversity. Analysts integrated information from published literature, government databases, and satellite imagery interpretation to create a dataset of 236,582 points across 12 landform classes, aligned to a 500 m resolution grid. The bedrock/pediment/plateau class was the dominant class (58%), followed by alluvial fans (21%), aeolian sands (11%), and aeolian dunes (5%). Approximately 85% of the reference points received high analyst confidence ratings, and ratings were especially high for classes with distinctive signatures, such as bedrock features, fine-grained lake deposits, urban/developed areas, water, and agricultural lands. Classification challenges consistently emerged in transitional zones between land-forms, areas with anthropogenic modifications, and complex landform assemblages where mapping resolution proved insufficient. The resulting dataset is a valuable resource for model validation and offers insights into arid region geomorphology. Additionally, it has the potential to support multiple applications, including dust hazard forecasting, terrain mobility assessment, soil property inference, and rangeland management.

Simulating Environmental Conditions for a Severe Dust Storm in Southwest Asia Using the Weather Research and Forecasting Model: A Model Configuration Sensitivity Study

Press Operations — Wed, 24 Sep 2025 19:06:00 GMT

Abstract: Dust aerosols create hazardous air quality conditions that affect human health, visibility, and military operations. Numerical weather prediction models are important tools for predicting atmospheric dust by simulating dust emission, transport, and chemical evolution. We assessed the Weather Research and Forecasting (WRF) model’s ability to simulate the atmospheric conditions that drove a major dust event in Southwest Asia during July–August 2018. We evaluated five WRF configurations against satellite observations and Reanalysis Version 5 (ERA5) reanalysis data, focusing on the event’s synoptic evolution, storm progression, vertical structure, and surface wind fields. Results revealed substantial differences between configurations using Noah and Noah Multiparameterization (Noah-MP) land surface models (LSMs), with Noah providing a superior representation of meteorological conditions despite theoretical expectations of similar performance in arid environments. The best-performing configuration (Noah LSM, Mellor–Yamada–Nakanishi–Niino planetary boundary layer scheme, and spectral nudging) of the five considered accurately simulated the progression of a low-level jet streak and the associated surface winds responsible for dust mobilization throughout the event. This study supports the US Army Engineer Research and Development Center’s efforts to improve dust forecasting and establishes a foundation for evaluating dust emission parameterizations by isolating meteorological forcing errors from dust model physics.

Using the Robot Operating System for Uncrewed Surface Vehicle Navigation to Avoid Beaching

Press Operations — Tue, 23 Sep 2025 17:08:00 GMT

Abstract: Our research explores the use of the Robotic Operating System (ROS) to autonomously navigate an uncrewed surface vehicle (USV). As a proof of concept, we set up a simulated world and spawned a virtual Wave Adaptive Modular Vehicle (WAM-V). We used the robot_localization package to localize the WAM-V in the virtual world and used move_base for the navigation of waypoints. The move_base package used both costmaps and path planners to reach its intended goal while simultaneously avoiding sub-merged shallow-water obstacles. Shallow-water obstacles are obstacles at a depth that is less than a user-defined value (1 meter in this case). Finally, we investigated using vizanti as a mission planner. This report provides a detailed explanation of the parameters that were modified to demonstrate a successful proof of concept.

Validating Predicted Soil Boundaries with In Situ Collections

Press Operations — Wed, 17 Sep 2025 20:20:00 GMT

Abstract: This US Army Engineer Research and Development Center (ERDC) technical note describes the process used by the Intelligent Environmental Battlefield Awareness (IEBA) team to validate the spatial distribution and texture class attribution of soil boundary predictions. The predicted global soil boundary polygons will serve as a primary base layer for populating other environmental variables; thus, it is essential to assess their robustness prior to the attribution stage.

Exploring Burnt Area Delineation with Cross-Resolution Mapping: A Case Study of Very High and Medium-Resolution Data

Press Operations — Wed, 03 Sep 2025 16:29:00 GMT

Abstract: Remote sensing is essential for mapping and monitoring burnt areas. Integrating Very High-Resolution (VHR) data with medium-resolution datasets like Landsat and deep learning algorithms can enhance mapping accuracy. This study employs two deep learning algorithms, UNET and Gated Recurrent Unit (GRU), to classify burnt areas in the Bandipur Forest, Karnataka, India. We explore using VHR imagery with limited samples to train models on Landsat imagery for burnt area delineation. Four models were analyzed:(a) custom UNET with Landsat labels, (b) custom UNET with PlanetScope-labeled data on Landsat, (c) custom UNET-GRU with Landsat labels, and (d) custom UNET-GRU with PlanetScope-labeled data on Landsat. Custom UNET with Landsat labels achieved the best performance, excelling in precision (0.89), accuracy (0.98), and segmentation quality (Mean IOU: 0.65, Dice Coefficient: 0.78). Using PlanetScope labels resulted in slightly lower performance, but its high recall (0.87 for UNET-GRU) demonstrating its potential for identifying positive instances. In the study, we highlight the potential and limitations of integrating VHR with medium-resolution satellite data for burnt area delineation using deep learning.

Bare Ground Classification Using a Spectral Index Ensemble and Machine Learning Models Optimized Across 12 International Study Sites

Press Operations — Mon, 18 Aug 2025 20:04:00 GMT

Abstract: This research investigates a global approach to map bare ground across diverse geographies with an ensemble of spectral indices using optimal thresholds identified in testing to train and evaluate machine learning models to extract bare ground pixels from Sentinel-2 imagery. Twelve locations in four Köppen climate zones with data from two seasons were evaluated. Accuracy assessment showed a mean F1 score of 80% and a mean Overall Accuracy (OA) of 81% for random forest and an F1 score of 78% and OA of 79% for support vector machine. Higher accuracies were observed in climate region-based models with mean F1 = 84% in three of four climate zones. Low accuracies occurred in winter imagery with leaf-off tree cover or building materials similar to bare ground. This framework provides a global approach to map bare ground without need for high-density time-series or deep learning models and moves beyond locally effective methods.

Creating an Augmented Soil Texture Master List Using the Gridded Soil Survey Geographic Database (gSSURGO)

Press Operations — Thu, 14 Aug 2025 14:39:00 GMT

Purpose: This US Army Engineer Research and Development Center (ERDC) technical note (TN) describes the workflow for creating an augmented soil texture master list that describes the surface-most (i.e., uppermost) USDA soil texture class and coarse fragment modifier. In conjunction with a soil similarity search algorithm, the soil texture master list fulfills a need identified by the Intelligent Environmental Battlefield Awareness (IEBA) project to generate detailed global soil boundary polygons. These polygons will serve as the base layer for populating other environmental variables, like soil temperature, soil moisture, depth to permafrost, and vegetation type, in the battlespace. This TN describes the purpose of the augmented soil texture master list, provides an overview of the gridded Soil Survey Geographic Database (gSSURGO), and describes the methodology used to create the soil texture master list.

A Revised Landform Map for Areas Prone to Dust Emission in the Southwestern United States

Press Operations — Wed, 13 Aug 2025 12:47:00 GMT

Abstract: An area’s landform composition can provide insight into its dust emission potential. In 2017, geomorphologists from the Desert Research Institute provided the US Army Engineer Research and Development Center with a 32-class landform map for portions of the Mojave and Sonoran Deserts in the southwest United States (SWUS) to support air quality and dust hazard modeling applications. We collaborated with the University of California to independently assess the map. Our review identified opportunities to improve the dataset, such as using a simpler landform classification system and revising individual geomorphic unit assignments to ensure consistent labeling across the study area. This report describes our approaches for refining the SWUS map and documents the updated 15-class landform map that resulted from our efforts.

KANICE: Kolmogorov-Arnold Networks with Interactive Convolutional Elements

Press Operations — Mon, 09 Jun 2025 20:36:00 GMT

Abstract: We introduce KANICE, a novel neural architecture that com-bines Convolutional Neural Networks (CNNs) with Kolmogorov-Arnold Network (KAN) principles. KANICE integrates Interactive Convolutional Blocks (ICBs) and KAN linear layers into a CNN framework. This leverages KANs’ universal approximation capabilities and ICBs’ adaptive feature learning. KANICE captures complex, non-linear data relationships while enabling dynamic, context-dependent feature extraction based on the Kolmogorov-Arnold representation theorem. We evaluated KANICE on four datasets: MNIST, Fashion-MNIST, EMNIST, and SVHN, comparing it against standard CNNs, CNN-KAN hybrids, and ICB variants. KANICE consistently outperformed baseline models, achieving 99.35% accuracy on MNIST and 90.05% on the SVHN dataset. Furthermore, we introduce KANICE-mini, a compact variant designed for efficiency. A comprehensive ablation study demonstrates that KANICE-mini achieves comparable performance to KANICE with significantly fewer parameters. KANICE-mini reached 90.00% accuracy on SVHN with 2,337,828 parameters, compared to KAN-ICE’s 25,432,000. This study highlights the potential of KAN-based architectures in balancing performance and computational efficiency in image classification tasks. Our work contributes to research in adaptive neural networks, integrates mathematical theorems into deep learning architectures, and explores the trade-offs between model complexity and performance, advancing computer vision and pattern recognition. The source code for this paper is publicly accessible through our GitHub repository (https://github.com/m-ferdaus/kanice).