25 Apr 2025

Analysis Tools and Techniques for Evaluating Quality in Synthetic Data Generated by the Virtual Autonomous Navigation Environment

Abstract: The capability to produce high-quality labeled synthetic image data is an important tool for building and maintaining machine learning datasets. However, ensuring computer-generated data is of high quality is very challenging. This report describes an effort to evaluate and improve synthetic image data generated by the Virtual Autonomous Navigation Environment’s Environment and Sensor Engine (VANE::ESE), as well as documenting a set of tools developed to process, analyze, and train models from, image datasets generated by VANE::ESE. Additionally, the results of several experiments are presented, including an investigation into using explainable AI techniques, and direct comparisons of various models trained on multiple synthetic datasets.

17 Apr 2025

Assessment of Neural Network Augmented Reynolds Averaged Navier Stokes Turbulence Model in Extrapolation Modes

Abstract: A machine-learned model enhances the accuracy of turbulence transport equations of RANS solver and applied for periodic hill test case. The accuracy is investigated in extrapolation modes. A parametric study is also performed to understand the effect of network hyperparameters on training and model accuracy and to quantify the uncertainty in model accuracy due to the non-deterministic nature of the neural network training. For any network, less than optimal mini-batch size results in overfitting, and larger than optimal reduces accuracy. Data clustering is an efficient approach to prevent the machine-learned model from over-training on more prevalent flow regimes, and results in a model with similar accuracy. Turbulence production is correlated with shear strain in the free-shear region, with shear strain and wall-distance and local velocity-based Reynolds number in the boundary layer regime, and with streamwise velocity gradient in the accelerating flow regime. The flow direction is key in identifying flow separation and reattachment regime. Machine-learned models perform poorly in extrapolation mode. A priori tests reveal model predictability improves as the hill dataset is partially added during training in a partial extrapolation model. These also provide better turbulent kinetic energy and shear stress predictions than RANS in a posteriori tests. Before a machine-learned model is applied for a posteriori tests, a priori tests should be performed.

12 Mar 2025

Reading the Ground: Understanding the Response of Bioelectric Microbes to Anthropogenic Compounds in Soil Based Terrestrial Microbial Fuel Cells

Abstract: Electrogenic bacteria produce power in soil based terrestrial microbial fuel cells (tMFCs) by growing on electrodes and transferring electrons released from the breakdown of substrates. The direction and magnitude of voltage production is hypothesized to be dependent on the available substrates. A sensor technology was developed for compounds indicative of anthropological activity by exposing tMFCs to gasoline, petroleum, 2,4-dinitrotoluene, fertilizer, and urea. A machine learning classifier was trained to identify compounds based on the voltage patterns. After 5 to 10 days, the mean voltage stabilized (+/- 0.5 mV). After the entire incubation, voltage ranged from -59.1 mV to 631.8 mV, with the tMFCs containing urea and gasoline producing the highest (624 mV) and lowest (-9 mV) average voltage, respectively. The machine learning algorithm effectively discerned between gasoline, urea, and fertilizer with greater than 94% accuracy, demonstrating that this technology could be successfully operated as an environmental sensor for change detection.

19 Feb 2025

Autonomous Cyberdefense Introduces Risk; Can We Manage the Risk?

Abstract: We discuss the human role in the design and control of cyberdefenses. We focus on machine learning training and algorithmic feedback and constraints, with the aim of motivating a discussion on achieving trust in autonomous cyberdefenses.

18 Oct 2024

Automated Change Detection in Ground-Penetrating Radar using Machine Learning in R

Abstract: Ground-penetrating radar (GPR) is a useful technique for subsurface change detection but is limited by the need for a subject matter expert to process and interpret coincident profiles. Use of a machine learning model can automate this process to reduce the need for subject matter expert processing and interpretation. Several machine learning models were investigated for the purpose of comparing coincident GPR profiles. Based on our literature review, a Siamese Twin model using a twinned convolutional network was identified as the optimum choice. Two neural networks were tested for the internal twinned model, ResNet50 and MobileNetV2, with the former historically having higher accuracy and the latter historically having faster processing time. When trained and tested on experimentally obtained GPR profiles with synthetically added changes, ResNet50 had a higher accuracy. Thanks to this higher accuracy, less computational processing was needed, leading to ResNet50 needing only 107 s to make a prediction compared to MobileNetV2 needing 223 s. Results imply that twinned models with higher historical accuracies should be investigated further. It is also recommended to test Siamese Twin models further with experimentally produced changes to verify the changed detection model’s accuracy is not merely specific to synthetically produced changes.

25 Sep 2024

Comparing the Thermal Infrared Signatures of Shallow Buried Objects and Disturbed Soil

Abstract: The alteration of physical and thermal properties of native soil during object burial produces a signature that can be detected using thermal infrared (IR) imagery. This study explores the thermal signature of disturbed soil compared to buried objects of different compositions (e.g., metal and plastic) buried 5 cm below ground surface (bgs) to better understand the mechanisms by which soil disturbance can impact the performance of aided target detection and recognition (AiTD/R). IR imagery recorded every five minutes were coupled with meteorological data recorded on 15-minute intervals from 1 July to 31 October 2022 to compare the diurnal and long-term fluctuations in raw radiance within a 25 × 25 pixel area of interest (AOI) above each target. This study examined the diurnal pattern of the thermal signature under several varying environmental conditions. Results showed that surface effects from soil disturbance increased the raw radiance of the AOI, strengthening the contrast between the object and background soil for several weeks after object burial. Enhancement of the thermal signature may lead to expanded windows of object visibility. Target age was identified as an important element in the development of training data sets for machine learning (ML) classification algorithms.

25 Sep 2024

Thermography Conversion for Optimal Noise Reduction

Abstract: Computer vision applications in terms of raw thermal radiance are limited by byte size. Normalizing the raw imagery reduces functional complexities that could otherwise aide a computer processing algorithm. This work explores a method to normalize 16-bit signed integer (I16) into unsigned 8-bit (U8) while maintaining the integrity of the correlation coefficients between the raw data sets and the environmental parameters that affects thermal anomaly detectability.

18 Sep 2024

Automated Mapping of Land Cover Type within International Heterogenous Landscapes Using Sentinel-2 Imagery with Ancillary Geospatial Data

Abstract: A near-global framework for automated training data generation and land cover classification using shallow machine learning with low-density time series imagery does not exist. This study presents a methodology to map nine-class, six-class, and five-class land cover using two dates of a Sentinel-2 granule across seven international sites. The approach uses a series of spectral, textural, and distance decision functions combined with modified ancillary layers to create binary masks from which to generate a balanced set of training data applied to a random forest classifier. For the land cover masks, stepwise threshold adjustments were applied to reflectance, spectral index values, and Euclidean distance layers, with 62 combinations evaluated. Global and regional adaptive thresholds were computed. An annual 95th and 5th percentile NDVI composite was used to provide temporal corrections to the decision functions, and these corrections were compared against the original model. The accuracy assessment found that the regional adaptive thresholds for both the two-date land cover and the temporally corrected land cover could accurately map land cover type within nine-class, six-class, and five-class schemes. Lastly, the five-class and six-class models were compared with a manually labeled deep learning model (Esri), where they performed with similar accuracies. The results highlight performance in line with an intensive deep learning approach, and reasonably accurate models created without a full annual time series of imagery.

9 Sep 2024

Leveraging Artificial Intelligence and Machine Learning (AI/ML) for Levee Culvert Inspections in USACE Flood Control Systems (FCS)

Abstract: Levee inspections are essential in preventing flooding within populated regions. Risk assessments of structures are performed to identify potential failure modes to maintain the safety and health of the structure. The data collection and defect coding parts of the inspection process can be labor-intensive and time-consuming. The integration of machine learning (ML) and artificial intelligence (AI) techniques may increase accuracy of assessments and reduce time and cost. To develop a foundation for a fully autonomous inspection process, this research investigates methods to gather information for levees, structures, and culverts as well as methods to identify indicators of future failures using AI and ML techniques. Robotic plat-form and instrumentation options that can be used in the data collection process are also explored, and a platform-agnostic solution is proposed.

13 Aug 2024

Predicting Soil Moisture Content Using Physics-Informed Neural Networks (PINNs)

Abstract: Environmental conditions such as the near-surface soil moisture content are valuable information in object detection problems. However, such information is generally unobtainable at the necessary scale without active sensing. Richards’ equation is a partial differential equation (PDE) that describes the infiltration process of unsaturated soil. Solving the Richards’ equation yields information about the volumetric soil moisture content, hydraulic conductivity, and capillary pressure head. However, Richards’ equation is difficult to approximate due to its nonlinearity. Numerical solvers such as finite difference method (FDM) and finite element method (FEM) are conventional in approximating solutions to Richards’ equation. But such numerical solvers are time-consuming when used in real-time. Physics-informed neural networks (PINNs) are neural networks relying on physical equations in approximating solutions. Once trained, these networks can output approximations in a speedy manner. Thus, PINNs have attracted massive attention in the numerical PDE community. This project aims to apply PINNs to the Richards’ equation to predict underground soil moisture content under known precipitation data.