16 Jun 2025

Entropy-Based Guidance of Deep Neural Networks for Accelerated Convergence and Improved Performance

Abstract: Neural networks have dramatically increased our capacity to learn from large, high-dimensional datasets across innumerable disciplines. However, their decisions are not easily interpretable, their computational costs are high, and building and training them are not straightforward processes. To add structure to these efforts, we derive new mathematical results to efficiently measure the changes in entropy as fully-connected and convolutional neural networks process data. By measuring the change in entropy as networks process data effectively, patterns critical to a well-performing network can be visualized and identified. Entropy-based loss terms are developed to improve dense and convolutional model accuracy and efficiency by promoting the ideal entropy patterns. Experiments in image compression, image classification, and image segmentation on benchmark datasets demonstrate these losses guide neural networks to learn rich latent data representations in fewer dimensions, converge in fewer training epochs, and achieve higher accuracy.

25 Feb 2025

The Design of Multimedia Object Detection Pipelines within the HPC Environment

Abstract: Computer vision multimedia pipelines have become both more sophisticated and robust over the years. The pipelines can accept multiple inputs, perform frame analysis, and produce outputs on a variety of platforms with near-real-time performance. Vendors such as Nvidia have significantly grown their framework and library offerings while providing tutorials and documentation via online training and tutorials. Despite the prolific growth, many of the libraries, frameworks, and tutorials come with noticeable limitations. The limitations are especially apparent within the high-performance computing (HPC) environment where graphic processing units may be older, user-level rights more restricted, and access to the graphical user interface not always available. This work describes the process of building multimedia object detection and segmentation pipelines within the HPC environment, its challenges, and ways to overcome the shortcomings. The project describes an iterative design process, which can be used as a blueprint for future development of similar computer vision pipelines within the HPC hosting environment.

12 Feb 2025

Cloud-Based Workflow to Process Regional Topobathymetric Lidar Datasets for Integrated USACE Shoaling Analyses

Purpose: This Coastal and Hydraulics Engineering Technical Note (CHETN) details a methodology to process and format regional topobathymetric datasets for use in the US Army Corps of Engineers (USACE) Corps Shoaling Analysis Tool (CSAT).

13 Sep 2024

Optimization Strategies for Geospatial Data on End-User Devices

Abstract: The ability to quickly disseminate geospatial data across all echelons, particularly those at the tactical edge, is critical to meeting threats described by the Multi-Domain Operations doctrine. The US Army Engineer Research and Development Center, Geospatial Research Laboratory (ERDC-GRL), is researching the optimization of the formats, data models, file sizes, and quality of geospatial products to be exploited by end-user devices (EUDs). This report describes a processing methodology comprising custom software and open-source tools to optimize Army Geospatial Enterprise Standard Sharable Geospatial Foundation and industry-accepted products for exploitation on EUDs. The Integrated Visual Augmentation System (IVAS) was emphasized, but other devices, including the Nett Warrior and Program Executive Office—Soldier targeting systems, were also studied. Additionally, we developed a compression methodology that reduced the size of three-dimensional model data by a factor of 9 without a loss in data quality. A summary of the results describes steps to address remaining technical issues and considers future efforts to further optimize geospatial data for additional EUDs and tactical applications.

26 Sep 2023

During Nearshore Event Vegetation Gradation (DUNEVEG): Geospatial Tools for Automating Remote Vegetation Extraction

Abstract: Monitoring and modeling of coastal vegetation and ecosystems are major challenges, especially when considering environmental response to hazards, disturbances, and management activities. Remote sensing applications can provide alternatives and complementary approaches to the often costly and laborious field-based collection methods traditionally used for coastal ecosystem monitoring. New and improved sensors and data analysis techniques have become available, making remote sensing applications attractive for evaluation and potential use in monitoring coastal vegetation properties and ecosystem conditions and changes. This study involves the extraction of vegetation metrics from airborne lidar and hyperspectral imagery (HSI) collected by the US Army Corps of Engineers (USACE) National Coastal Mapping Program (NCMP) to quantify coastal dune vegetation characteristics. A custom geoprocessing toolbox and associated suite of tools were developed to allow inputs of common NCMP lidar and imagery products to help automate the workflow for extracting prioritized dune vegetation metrics in an efficient and repeatable way. This study advances existing coastal ecosystem knowledge and remote sensing techniques by developing new methodologies to classify, quantify, and estimate critical coastal vegetation metrics which will ultimately improve future estimates and predictions of nearshore dynamics and impacts from disturbance events.

30 Sep 2021

Accelerating the Tactical Decision Process with High-Performance Computing (HPC) on the Edge: Motivation, Framework, and Use Cases

Abstract: Managing the ever-growing volume and velocity of data across the battlefield is a critical problem for warfighters. Solving this problem will require a fundamental change in how battlefield analyses are performed. A new approach to making decisions on the battlefield will eliminate data transport delays by moving the analytical capabilities closer to data sources. Decision cycles depend on the speed at which data can be captured and converted to actionable information for decision making. Real-time situational awareness is achieved by locating computational assets at the tactical edge. Accelerating the tactical decision process leverages capabilities in three technology areas: (1) High-Performance Computing (HPC), (2) Machine Learning (ML), and (3) Internet of Things (IoT). Exploiting these areas can reduce network traffic and shorten the time required to transform data into actionable information. Faster decision cycles may revolutionize battlefield operations. Presented is an overview of an artificial intelligence (AI) system design for near-real-time analytics in a tactical operational environment executing on co-located, mobile HPC hardware. The report contains the following sections, (1) an introduction describing motivation, background, and state of technology, (2) descriptions of tactical decision process leveraging HPC problem definition and use case, and (3) HPC tactical data analytics framework design enabling data to decisions.

2 Jun 2020

PUBLICATION NOTICE: Parallel File I/O for Geometric Models: Formats and Methods

Abstract: Processing large amounts of data in a High-Performance Computing (HPC) environment can be throttled quickly without an efficient method for utilizing disk I/O. The Geometry Engine component of the Virtual Environment for Sensor Performance Assessment (VESPA) uses MPI-IO to load the geometric data in parallel and avoid creating a bottleneck on disk I/O interactions. This parallel I/O method requires formatting the data into specific binary file formats so each MPI process of the parallel program can determine where to read or write data without colliding with other MPI processes. Addressing the collision problem resulted in the development of two binary file formats, the Mesh Binary file (.mb) and the Scene Chunk Pack file (.scp). The Mesh Binary file contains the essential data required to recreate the landscape and vegetation geometry used by the Geometry Engine. The Scene Chunk Pack file is used to write the partitioned geometry to disk, so the ray casting engine can reload the distributed geometry without repeating the partitioning process. Both of these files together support reading and writing for the partitioning phase and the ray casting phase of the Geometry Engine. This report discusses these formats in detail and outlines how the Geometry Engine reads and writes these files in parallel on HPC.

7 May 2020

PUBLICATION NOTICE: Using Morton Codes to Partition Faceted Geometry: An Architecture for Terabyte-Scale Geometry Models

Abstract: The Virtual Environment for Sensor Performance Assessment (VESPA) project requires enormous, high-fidelity landscape models to generate synthetic sensor imagery with little to no artificial artifacts. These high-fidelity landscapes require a memory footprint substantially larger than a single High Performance Computer’s (HPC) compute node’s local memory. Processing geometries this size requires distributing the geometry over multiple compute nodes instead of including a full copy in each compute node, the common approach in parallel modeling applications. To process these geometric models in parallel memory on a high-performance computing system, the Geometry Engine component of the VESPA project includes an architecture for partitioning the geometry spatially using Morton codes and MPI (Message Passing Interface) collective communication routines. The methods used for this partitioning process will be addressed in this report. Incorporating this distributed architecture into the Geometry Engine provides the capability to distribute and perform parallel ray casting on landscape geometries over a Terabyte in size. Test case timings demonstrate scalable speedups as the number of processes are increased on an HPC machine.

14 Apr 2020

PUBLICATION NOTIFICATION: Coincidence Processing of Photon-Sensitive Mapping Lidar Data

 Link: http://dx.doi.org/10.21079/11681/35599 Report Number: ERDC/GRL TR-20-1Title: Coincidence