Publication Notices

Notifications of New Publications Released by ERDC

Contact Us






ERDC Library Catalog

Not finding what you are looking for? Search the ERDC Library Catalog

Tag: Machine learning
  • Data-Driven Modeling of Groundwater Level Using Machine Learning

    Purpose: This US Army Engineer Research and Development Center (ERDC), Coastal and Hydraulics Laboratory engineering technical note (CHETN) documents a preliminary study on the use of specialized machine learning (ML) methods to model the variations in groundwater level (GWL) with time. This approach uses historical groundwater observation data at seven gage locations in Wyoming, USA, available from the USGS database and historical data on several relevant meteorological variables obtained from the ERA5 reanalysis dataset produced by the Copernicus Climate Change Service (usually referred to as C3S) at the European Center for Medium-Range Weather Forecasts to predict future GWL values for a desired period of time. The results presented in this report indicate that the ML method has the potential to predict both short-term (4-hourly) as well as daily variations in GWL several days into the future for the chosen study region, thus alleviating the need for employing sophisticated process-based numerical models with complicated model structure configurations.
  • CRREL Environmental Wind Tunnel Upgrades and the Snowstorm Library

    Abstract: Environmental wind tunnels are ideal for basic research and applied physical modeling of atmospheric conditions and turbulent wind flow. The Cold Regions Research and Engineering Laboratory's own Environmental Wind Tunnel (EWT)—an open-circuit suction wind tunnel—has been historically used for snowdrift modeling. Recently the EWT has gone through several upgrades, namely the three-axis chassis motors, variable frequency drive, and probe and data acquisition systems. The upgraded wind tunnel was used to simulate various snowstorm conditions to produce a library of images for training machine learning models. Various objects and backgrounds were tested in snowy test conditions and no-snow control conditions, producing a total of 1.4 million training images. This training library can lead to improved machine learning models for image-cleanup and noise-reduction purposes for Army operations in snowy environments.
  • Scaling and Sensitivity Analysis of Machine Learning Regression on Periodic Functions

    Abstract: In this report we document the scalability and sensitivity of machine learning (ML) regression on a periodic, highly oscillating, and 𝐶∞ function. This work is motivated by the need to use ML regression on periodic problems such as tidal propagation. In this work, TensorFlow is used to investigate the machine scalability of a periodic function from one to three dimensions. Wall clock times for each dimension were calculated for a range of layers, neurons, and learning rates to further investigate the sensitivity of the ML regression to these parameters. Lastly, the stochastic gradient descent and Adam optimizers wall clock timings and sensitivities were compared.
  • Environmentally Informed Buried Object Recognition

    The ability to detect and classify buried objects using thermal infrared imaging is affected by the environmental conditions at the time of imaging, which leads to an inconsistent probability of detection. For example, periods of dense overcast or recent precipitation events result in the suppression of the soil temperature difference between the buried object and soil, thus preventing detection. This work introduces an environmentally informed framework to reduce the false alarm rate in the classification of regions of interest (ROIs) in thermal IR images containing buried objects. Using a dataset that consists of thermal images containing buried objects paired with the corresponding environmental and meteorological conditions, we employ a machine learning approach to determine which environmental conditions are the most impactful on the visibility of the buried objects. We find the key environmental conditions include incoming short-wave solar radiation, soil volumetric water content, and average air temperature. For each image, ROIs are computed using a computer vision approach and these ROIs are coupled with the most important environmental conditions to form the input for the classification algorithm. The environmentally informed classification algorithm produces a decision on whether the ROI contains a buried object by simultaneously learning on the ROIs with a classification neural network and on the environmental data using a tabular neural network. On a given set of ROIs, we have shown that the environmentally informed classification approach improves the detection of buried objects within the ROIs.
  • Snow-Covered Region Improvements to a Support Vector Machine-Based Semi-Automated Land Cover Mapping Decision Support Tool

    Abstract: This work builds on the original semi-automated land cover mapping algorithm and quantifies improvements to class accuracy, analyzes the results, and conducts a more in-depth accuracy assessment in conjunction with test sites and the National Land Cover Database (NLCD). This algorithm uses support vector machines trained on data collected across the continental United States to generate a pre-trained model for inclusion into a decision support tool within ArcGIS Pro. Version 2 includes an additional snow cover class and accounts for snow cover effects within the other land cover classes. Overall accuracy across the continental United States for Version 2 is 75% on snow-covered pixels and 69% on snow-free pixels, versus 16% and 66% for Version 1. However, combining the “crop” and “low vegetation” classes improves these values to 86% for snow and 83% for snow-free, compared to 19% and 83% for Version 1. This merging is justified by their spectral similarity, the difference between crop and low vegetation falling closer to land use than land cover. The Version 2 tool is built into a Python-based ArcGIS toolbox, allowing users to leverage the pre-trained model—along with image splitting and parallel processing techniques—for their land cover type map generation needs.
  • Landform Identification in the Chihuahuan Desert for Dust Source Characterization Applications: Developing a Landform Reference Data Set

    Abstract: ERDC-Geo is a surface erodibility parameterization developed to improve dust predictions in weather forecasting models. Geomorphic landform maps used in ERDC-Geo link surface dust emission potential to landform type. Using a previously generated southwest United States landform map as training data, a classification model based on machine learning (ML) was established to generate ERDC-Geo input data. To evaluate the ability of the ML model to accurately classify landforms, an independent reference landform data set was created for areas in the Chihuahuan Desert. The reference landform data set was generated using two separate map-ping methodologies: one based on in situ observations, and another based on the interpretation of satellite imagery. Existing geospatial data layers and recommendations from local rangeland experts guided site selections for both in situ and remote landform identification. A total of 18 landform types were mapped across 128 sites in New Mexico, Texas, and Mexico using the in situ (31 sites) and remote (97 sites) techniques. The final data set is critical for evaluating the ML-classification model and, ultimately, for improving dust forecasting models.
  • Buried-Object-Detection Improvements Incorporating Environmental Phenomenology into Signature Physics

    Abstract: The ability to detect buried objects is critical for the Army. Therefore, this report summarizes the fourth year of an ongoing study to assess environmental phenomenological conditions affecting probability of detection and false alarm rates for buried-object detection using thermal infrared sensors. This study used several different approaches to identify the predominant environmental variables affecting object detection: (1) multilevel statistical modeling, (2) direct image analysis, (3) physics-based thermal modeling, and (4) application of machine learning (ML) techniques. In addition, this study developed an approach using a Canny edge methodology to identify regions of interest potentially harboring a target object. Finally, an ML method was developed to improve automatic target detection and recognition performance by accounting for environmental phenomenological conditions, improving performance by 50% over standard automatic target detection and recognition software.
  • Adversarial Artificial Intelligence: Implications for Military Operations

    Introduction: Artificial intelligence and machine learning algorithms are at the forefront of current research to help military analysts deal with triaging ever larger amounts of data from deployed sensors. These automated approaches will become increasingly embedded into the military decision making process, which makes it crucial to understand how these algorithms generate outputs and how sensitive they are to perturbations during training or classification. In other words, humans must have a ‘theory of mind’ for these sets of approaches in order to begin to trust them enough to make life or death decisions. Research in this area is known as adversarial examples for artificial intelligence / machine learning. Previous works in this domain focused on degrading classification performance with respect to added noise to new data. Some of these works achieved notable results on image data by subtly increasing noise, such that the image appeared unaltered to the human eye, but significantly impacted performance (Athalye et al. 2017). Povolny and Trivedi (2020) achieved similar results, but made a small visually obvious change to induce a degradation in performance. One notable work examined the effects of an increase in physical scale of the sensed environment (such as the large areas recorded for remote sensing platforms) on adversarial perturbations (Czaja et al. 2018). This technical note (TN) describes an initial foray into understanding how physical changes to the appearance of military vehicles resulted in performance degradation for a convolutional neural network (CNN). The military vehicles chosen were the M2 Bradley Infantry Fighting Vehicle and the M1064 Mortar Carrier. As stand-ins for the actual vehicle, plastic scale models were used, each a 1/35 scale replica. The results of this research have yielded a curated training and test data set of images related to the M2 and M1064, trained models based on a combined ResNet / Inception implementation from the Keras project, and adversarial examples mocked up using the scale models with images taken by a smartphone.
  • Assessing the Feasibility of Detecting Epileptic Seizures Using Non-Cerebral Sensor Data

    Abstract: This paper investigates the feasibility of using non-cerebral, time-series data to detect epileptic seizures. Data were recorded from fifteen patients (7 male, 5 female, 3 not noted, mean age 36.17 yrs), five of whom had a total of seven seizures. Patients were monitored in an inpatient setting using standard video-electroencephalography (vEEG), while also wearing sensors monitoring electrocardiography, electrodermal activity, electromyography, accelerometry, and audio signals (vocalizations). A systematic and detailed study was conducted to identify the sensors and the features derived from the non-cerebral sensors that contribute most significantly to separability of data acquired during seizures from non-seizure data. Post-processing of the data using linear discriminant analysis (LDA) shows that seizure data are strongly separable from non-seizure data based on features derived from the signals recorded. The mean area under the receiver operator characteristic (ROC) curve for each individual patient that experienced a seizure during data collection, calculated using LDA, was 0.9682. The features that contribute most significantly to seizure detection differ for each patient. The results show that a multimodal approach to seizure detection using the specified sensor suite is promising in detecting seizures with both sensitivity and specificity. Moreover, the study provides a means to quantify the contribution of each sensor and feature to separability. Development of a non-electroencephalography (EEG) based seizure detection device would give doctors a more accurate seizure count outside of the clinical setting, improving treatment and the quality of life of epilepsy patients.
  • Semi-Automated Land Cover Mapping Using an Ensemble of Support Vector Machines with Moderate Resolution Imagery Integrated into a Custom Decision Support Tool

    Abstract: Land cover type is a fundamental remote sensing-derived variable for terrain analysis and environmental mapping applications. The currently available products are produced only for a single season or a specific year. Some of these products have a coarse resolution and quickly become outdated, as land cover type can undergo significant change over a short time period. In order to enable on-demand generation of timely and accurate land cover type products, we developed a sensor-agnostic framework leveraging pre-trained machine learning models. We also generated land cover models for Sentinel-2 (20m) and Landsat 8 imagery (30m) using either a single date of imagery or two dates of imagery for mapping land cover type. The two-date model includes 11 land cover type classes, whereas the single-date model contains 6 classes. The models’ overall accuracies were 84% (Sentinel-2 single date), 82% (Sentinel-2 two date), and 86% (Landsat 8 two date) across the continental United States. The three different models were built into an ArcGIS Pro Python toolbox to enable a semi-automated workflow for end users to generate their own land cover type maps on demand. The toolboxes were built using parallel processing and image-splitting techniques to enable faster computation and for use on less-powerful machines.