Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Accelerating and Improving the Reliability of Low Failure Probability Computations to Support the Efficient Safety Evaluation and Deployment of Advanced Reactor Technologies

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Brookhaven National Laboratory

Problem Scoping

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Office of Electricity

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Idaho National Laboratory

Prioritizing the potential risk posed to human health by chemicals requires tools that can estimate exposure from limited information. In this study, chemical structure and physicochemical properties were used to predict the probability that a chemical might be associated with any of four exposure pathways leading from sources-consumer (near-field), dietary, far-field industrial, and far-field pesticide-to the general population. The balanced accuracies of these source-based exposure pathway models range from 73 to 81%, with the error rate for identifying positive chemicals ranging from 17 to 36%. We then used exposure pathways to organize predictions from 13 different exposure models as well as other predictors of human intake rates. We created a consensus, meta-model using the Systematic Empirical Evaluation of Models framework in which the predictors of exposure were combined by pathway and weighted according to predictive ability for chemical intake rates inferred from human biomonitoring data for 114 chemicals. The consensus model yields an R2 of ∼0.8. We extrapolate to predict relevant pathway(s), median intake rate, and credible interval for 479 926 chemicals, mostly with minimal exposure information. This approach identifies 1880 chemicals for which the median population intake rates may exceed 0.1 mg/kg bodyweight/day, while there is 95% confidence that the median intake rate is below 1 μg/kg BW/day for 474572 compounds

SSA -- Office of Analytics, Review, and Oversight; Office of Hearing Operations, Office of Disability Information Systems

Planned (not in production)

In production: less than six months

In production: less than six months

In production: less than six months

In production: less than six months

In production: less than six months

In production: less than one year

In production: less than one year

In order to get a full accounting of delay, automated voice detection of ATC and aircraft interaction is required. Many delay events, such as vectoring, are not currently reported/detected/accounted for and voice detection would enable automated detection.

In production: less than 1 year

Initial development

Natural Language Processing;

Agency Generated

We use ChatGPT in code writing assistance.

In production: less than 1 year

SCM classifies surface incident reports by event type, such as Runway Incursion, Runway Excursion, Taxiway Incursion/Excursion and categorizes runway incursions further by severity type (Category A, B, C, D, E)

In production: more than 1 year

Refinments planned for future release

Support Vector Machines, Gradient boosting, neural networks, natural language processing

Agency Generated

AIT/EIM Platform

The AVS International office is required to identify means of compliance to ICAO Standards and Recommended Practices (SARPs). Both SARPs and means of compliance evidence are text paragraphs scattered across thousands of pages of documents. AOV identified a need to find each SARP, evaluate the text of many FAA Orders, and suggest evidence of compliance based upon the evaluation of the text. The base dataset used by RCMT is the documents’ texts deconstructed into paragraphs. RCMT processes all the documents’ paragraphs run through Natural Language Processing (NLP) (this process has an AI aspect) to extract the meaning (semantics) of the text. RCMT then employs a recommender system (also using some AI technology) to take the texts augmented by the texts’ meaning to establish candidate matches between the ICAO SARPs and FAA text that provides means of compliance.

Planned (not in production)

User Acceptance Testing to begin early spring '22

ML (Recommender Algorithim) , NLP,

Agency Generated

FAA Orders and ICAO Annexes

AVS identified a need to derive the joint aircraft system codes (JASC) chapter codes from the narrative description within service difficulty reports (SDR), a form of safety event reporting from aircraft operators. A team of graduate students at George Mason University collaborated with AVS employees to apply Natural Language Processing (NLP) and Machine Learning to predict JASC codes. This method can be used to check SDR entries to ensure the correct codes were provided or to assign a code when one was not.

Planned (not in production)

NLP, ML Classification

Agency Generated

Service Difficult Reporting System (SDRS)

NextGen (ANG)

OPC leverages data from several sources such as weather radar, lightning networks, satellite and numerical models to produce a radar-like depiction of precipitation. The algorithm then applies machine learning techniques based on years of satellite and model data to improve the accuracy of the location and intensity of the precipitation areas.

In production: more than 1 year

OPC runs in a pseudo-operational capacity via a webpage maintained by the Massachusetts Institute of Technology - Lincoln Lab, as well as in a test and evaluation capacity in a research mode.

AI, ML via a Convolutional Neural Network

Agency Generated

Aviation Safety (AVS)

The Multiple Airport Route Separation (MARS) program is developing a safety case for reduced separation standards between Performance Based Navigation (PBN) routes in terminal airspace. These new standards may enable deconfliction of airports in high-demand metropolitan areas, including the Northeast Corridor (NEC), North Texas, and Southern California. To build necessary collision risk models for the safety case, several models are needed, including one that describes the behavior of aircraft that fail to navigate the procedure correctly. These events are very rare and difficult to identify with standard data sources. Prior work has used Machine Learning to filter incident data to identify similar events on departure procedures.

Planned (not in production)

Python in Jupyter Labs, ML, NLP

Agency Generated

CEDAR MORs/EORs, Pilot Deviation Reports

Sensitive communications.

Successfully demonstrated use of an AI capability to analyze camera images of a wind sock to produce highly accurate surface wind speed and direction information in remote areas that don’t have a weather observing sensor.

Planned (not in production)

Successfully tested but not in production.

Agency Generated

ROMIO is an operational demonstration to evaluate the feasibility to uplink convective weather information to aircraft operating over the ocean and remote regions. Capability converted weather satellite data, lightning and weather prediction model data into areas of thunderstorm activity and cloud top heights. AI is used to improve the accuracy of the output based on previous activity compared to ground truth data.

Planned (not in production)

Technical transfer of capability to industry planned this summer.

AI, ML via a Convolutional Neural Network

Agency Generated

NSR Human Injury Research Division

Description: Utilize deep learning models for predicting head kinematics directly from crash videos. The utilization of deep learning techniques enables the extraction of 3D kinematics from 2D views, offering a viable alternative for calculating head kinematics in the absence of sensors or when sensor availability is inadequate, and when high-quality sensor data is absent
Input: Vehicle crash videos
Output: Angular velocity - injury prediction

Planned (not in production)

Proof of Concept completed and published

Deep learning models - Convolutional Neural Networks, Long-Short Term Memory based Recurrent Neural Networks

Agency Generated

NSR Human Injury Research Division

Description: Utilize deep learning for predicting crash parameters, Delta-V (change in velocity) and PDOF (principal direction of force), directly from real-world crash images. Delta-V and PDOF are two most important parameters affecting injury outcome. Deep learning models can help predict both Delta-V and PDOF, without the need to run WinSmash software for Delta-V computation, and without requiring estimations by crash examiners. Moreover, with deep learning models, the Delta-V and PDOF can be obtained within milliseconds, providing rapid results for improved efficiency"
Input: Real world crash images
Output: Delta-V & PDOF

Planned (not in production)

Currently under development

Deep learning models - Convolutional Neural Networks

Agency Generated

Data is taken from NASS-CDS, CISS and CIREN databases, which are publically available

PHMSA Office of Chief Counsel (PHC)

Artificial Intelligence Support for Rulemaking - Using ChatGPT to support the rulemanking processes to provide significant efficiencies, reduction of effort, or the ability to scale efforts for unusual levels of public scrutiny or interest (e.g. comments on a rulemaking). ChatGPT will be used to provide:
1. Sentiment Analysis – Is the comment positive / negative or neutral towards the proposed rule.
2. Relevance Analysis – Whether the particular comment posted is relevant to the proposed rule
3. Synopsis of the posted comment.
4. Cataloging of comments.
5. Identification of duplicate comments.

Planned (not in production)

This is a pilot initiative

Agency Generated

Office Automation for Administrative Systems Support

Office of Research, Development and Technology

Description: Deep learning computer vision algorithms aimed at analyzing aggregate particle size grading.
Input: Images of ballast cross sections
Output: Ballast fouling index

In production: less than 1 year

Office of Research, Development and Technology

Description: DeepCNet-based neural network to identify and classify track-related features (e.g., track components, such as fasteners and ties) for "change detection" applications.
Input: Line-scan images from rail-bound inspection systems
Output: Notification of changes from status quo or between different inspections based on geolocation.

In production: more than 1 year

Office of Research, Development and Technology

Description: Leveraging large volumes of these recursive track geometry measurements to develop and implement automated machine-learning-based processes for analyzing, predicting, and reporting track locations of concern, including those with significant rates of degradation.
Input: Track geometry measurements and exceptions
Output: Inspection report that includes the trending of track geometry measures and time to failure (i.e., maintenance and safety limits).

In production: more than 1 year

jatracey@usgs.gov

Janice Gordon, janice_gordon@usgs.gov

This software system takes wildlife camera trap images as inputs and outputs the probability of the image belonging to user-specified taxonomic classes based on wildlife species present in each image. (Wildlife camera traps are motion-triggered, time lapse, and other camera systems placed in the field to capture images of wildlife at the location and times where and when the cameras are placed.) The process of humans reviewing, labeling, and QA/QCing labels is labor intensive, time consuming, and costly. Developing AI systems that can perform these tasks within an acceptable level of accuracy can reduce the costs in extracting tabular data from camera-based datasets and increase the volume of data for analysis. The system supports training experiments where model hyperparameters and training dataset characteristics can be varied to find those that are more optimal for training. Training, validation, and testing datasets have human-assigned labels and are used to train and evaluate the models. Once trained, the models can be used to predict classes on unlabeled images. We use a convolutional neural network (CNN) approach based on TensorFlow and training is run on the USGS Tallgrass supercomputer designed for AI/ML workflows.

Planned (not in production)

in development, testing and training experimentation

convolutional neural networks

Agency Generated

CSS/Ecosystems

jatracey@usgs.gov

Janice Gordon, janice_gordon@usgs.gov

This project extends the application of codes developed for wildlife underpass camera trap image classification. Similarly, the system takes walrus haulout camera trap images as inputs and outputs the probability of the image containing walruses and various human disturbances (boats, aircraft, etc.). We will use and further develop the previous system's capability of supporting training experiments. Training, validation, and testing datasets have human-assigned labels and are used to train and evaluate the models. Once trained, the models can be used to predict classes on unlabeled images from ongoing camera monitoring efforts. We use a convolutional neural network (CNN) approach based on TensorFlow and training is run on the USGS Tallgrass supercomputer designed for AI/ML workflows.

Planned (not in production)

in development

convolutional neural networks

Agency Generated

CSS/Ecosystems

jatracey@usgs.gov

Janice Gordon, janice_gordon@usgs.gov

The mission of the USGS Amphibian and Reptile Minitoring Initiative (ARMI) is to provide essential scientific information to managers to help arrest or reverse amphibian population declines. Acoustic monitoring of amphibian (anuran) vocalizations are a core technique used by ARMI researchers. Reviewing audio recordings and identifying species vocalizations captured therein is time consuming and labor intensive. For these reasons, many recordings remain unprocessed, preventing valuable data from being available for analysis. Our goal is to train convlutional neural networks (CNNs) that take audio clips that have been converted to sonograms (images) and classify the species generating the vocalizations in the recordings. Our initial prototype project will attempt to develop models that can identify audio clips containing bullfrog (*Lithobates catesbeianus*, which are native in some parts of the US and a destructive invasive species in others) vocalizations. The software will be build using the TensorFlow Python API and training will be performed on the USGS Tallgrass supercomputer.

Planned (not in production)

in development

convolutional neural networks

Agency Generated

CSS/Ecosystems

jatracey@usgs.gov

Janice Gordon, janice_gordon@usgs.gov

This system will to take pairs of mountain lion (*Puma concolor*) facial images and output the probability that the images come from the same individual mountain lion. This will allow researches to passively "mark" individuals and support population estimation analyses. We will use a "Siamese" convolutional neural network architecture that has been used in other facial recognition and motion tracking applications.

Planned (not in production)

in development

convolutional neural networks

Agency Generated

CSS/Ecosystems

jatracey@usgs.gov

Janice Gordon, janice_gordon@usgs.gov

This system, once developed, will input drone imagery and output bounding boxes for individual walruses. If successful, this will allow researchers with Alaska Science Center to count the numbers of walruses in drone imagery to support population reseaarch. The system will use TensorFlow-based convolutional neural networks for object detection trained on the USGS Tallgrass supercomputer.

Planned (not in production)

convolutional neural networks

Agency Generated

CSS/Ecosystems

tsando@usgs.gov

Kristin Jaeger, kjaeger@usgs.gov

The PROSPER modeling framework was developed to incorporate sparse streamflow observation data representing wet or dry stream conditions and gridded hydroclimatic explanatory data to predict the annual probability of streamflow permanence at 30-m (PROSPER Pacific Northwest) or 10-m (PROSPER Upper Missouri) resolution. The training data are point observations of wet or dry at locations in the Pacific Northwest or Upper Missouri River basin. The PROSPER models were primarily developed using the FCPGTools (Barnhart, Sando, et al., 2020), R, and USGS HPC resources (Yeti).

In production: more than 1 year

Models have been calibrated and outputs published for the Pacific Northwest. Models for other domains (Upper Missouri River basin) are currently in review.

Random Forest Classification

https://doi.org/10.5066/P96F7RX8, https://doi.org/10.5066/F7BV7FSP

Stacey Archfield

sarch@usgs.gov

Roy Sando, tsando@usgs.gov

The goal of this project is to develop a method for predicting daily hydrologic drought using machine learning models calibrated on streamflow data (response) and meteorological forcing data. Models will be built at individual gages across CONUS, then transferred to ungaged basins using a 'donor model' approach that identifies which gages are most similar to the ungaged basin and combines the models from those gages for the final prediction. Models will be developed and run on the USGS HPC systems.

Planned (not in production)

Random forest regression, random forest classification

Agency Generated

John Hammond

jhammond@usgs.gov

Roy Sando, tsando@usgs.gov

The goal of this project is to build and test multiple ML models for predicting and forecasting daily hydrologic drought in the Colorado River Basin (CRB). Similar to the project listed in line 9, we use gridded meteorologic forcing data and daily streamflow data in the CRB to build random forest and neural networks (long-short term memory) to determine the best approach to predicting and forecasting hydrologic drought. The project is being developed on AWS and in cooperation with CHS. We are also using the USGS HPC systems.

Planned (not in production)

Random forest regression, random forest classification, random survival forests, neural networks, long-short term memory, recurrent neural networks

Agency Generated

nhitt@usgs.gov

Ben Letcher, bletcher@usgs.gov

This study focuses on the development of an AI system to recognize individual fish and their disease status from images. Success of this effort could complement or replace traditional mark-recapture methods used for estimating abundance, survival, and movement, and this could greatly reduce costs to fisheries managers. Likewise, disease detection from images could enable new approaches for assessing status and trends in fish health.

In production: less than 1 year

Research on-going to optimize model parameters and data pipeline

Convolutional Neural Networks

Agency Generated

Beta version: https://code.usgs.gov/marami/Fish-AI/-/releases/fish

Lotspeich, Russ

rlotspei@usgs.gov

Frank Engel, fengel@usgs.gov

The River Image Sensing (RISE) project is charged with the development of a reliable camera system for integration into the operational streamgage monitoring network of the USGS Water Mission Area. In addition to capturing images and videos, the RISE system will be capable of producing time-series of surface water levels derived from still camera images using AI/ML modeling techniques.

In production: less than 1 year

Proof-of-concept completed in FY21. Researching now limitations, requirements, and operational implementation approaches.

Convolutional Neural Networks

Agency Generated

bletcher@usgs.gov

Jenn Fair, jfair@usgs.gov

The goals of this project are to 1) measure how much water flows in small, ungaged stream networks using timelapse images captured by inexpensive and off-the-shelf cameras and 2) provide a web-based platform for making the images, associated climate and other related data as well as the model itself easy to access and explore. Data for training come from user-uploaded imagery and flow data (when available). Database is available for uploading and image viewing here: https://www.usgs.gov/apps/ecosheds/fpe/

In production: less than 1 year

Initial models have been trained. Working on estimating relative flow from sites with no gage data based on human-annotated pair-wise comparisons of images. Also working on understanding transferability of models to sites with less data.

Convolutional Neural Networks

bletcher@usgs.gov

Jenn Fair, jfair@usgs.gov

The goal is to link existing hydrological flow data (e.g., USGS stream gages) and models (e.g., USGS Process-Guided Deep Learning Models for flow and temparture) with trout population dynamic models, changes to fish catch, and the economic benefits of recreational fishing. These trout population dynamic models will be developed based on observational data, existing literature estimates, and existing models.

In production: less than 6 months

Model training will begin in the 2nd quarter of FY22

Convolutional Neural Networks

Cashman, Matthew

mcashman@usgs.gov

The project objective is to take a large dataset of rapid habitat assessment data collected by multiple jurisdictions in the Chesapeake Bay Watershed, train a predictive model using those data, and use that model to predict stream habitat conditions for all unmeasured stream reaches in the region. The model is able to generate predictions for multiple aspects of physical habitat condition. The model directly connects to EPA's database containing the training data, enabling it to be rapidly updated when new data updates occur.

In production: less than 1 year

End-to-end workflows have all been developed. Preliminary models have been trained and models continued to be iterated for improvements using additional feature datasets. Models will be finalized end of Q3 FY22.

Random Forest Regression, XGBoost

From Another Agency

klandolt@usgs.gov

Jennifer Dieck, jdieck@usgs.gov

Our project includes two stages of AI, the first is a binary detector to automate the detection of wildlife in aerial imagery and the second is a robust classification algorithm to automate the taxonomic classification of wildlife from the binary detector. The input of the first and second stage are manually annotated polygons around targets of interest and their taxonomic classification values from family to species, respectively. We use Tallgrass to develop and train our algorithms, BlackPearl/Caldera to store our large image datasets, a hosted instance of a customized version of the Computer Vision Annotation Tool to gather manually annotated data, and a separate PostgreSQL database to store annotations and image metadata.

In production: less than 1 year

Model training for wildlife detection is complete, further training, tuning, and optimization for detection and classification set for FY22, large production-level tests set for FY22

Convolutional neural networks

Goodling, Phillip

pgoodling@usgs.gov

"Fleming, Brandon, bjflemin@usgs.gov Stackelberg, Paul, pestack@usgs.gov, Belitz, Ken, kbelitz@usgs.gov"

This project uses observations of the depth to bedrock reported by private well drillers in the Delaware River Basin to train a Random Forest model to map the thickness of the regolith layer. This data product will support groundwater and hydrologic modeling efforts in the basin.

In production: less than 1 year

Prelminary models developed, optimization is underway.

Random Forest Classification and Regression

Buscombe, Daniel

dbuscombe@contractor.usgs.gov

jwarrick@usgs.gov

A course in application of deep learning image segmentation, image classification, and object-in-image detection. Course includes software written in Python using Keras and Tensorflow ML libraries, software documentation, data, website, and slides. See course website https://dbuscombe-usgs.github.io/MLMONDAYS/ for more details

In production: more than 1 year

Course is completed and has been taken by more than 100 particpants

Deep convolutional neural networks; ResNet, MobileNet, UNet, RetinaNet

Natural Hazards

Buscombe, Daniel

dbuscombe@contractor.usgs.gov

pwernette@usgs.gov; jwarrick@usgs.gov

Coast Train is a multi-labeler ML-ready dataset of orthomosaic and satellite images of coastal, estuarine, and wetland environments and corresponding thematic label masks. The data consist of spatial and time-series, and contains 1.2 billion labelled pixels, representing over 3.6 million hectares.

In production: less than 6 months

Data release and journal article being prepared for publication

Doodler: https://github.com/dbuscombe-usgs/dash_doodler

Natural Hazards

josh_adams@usgs.gov

cahorton@usgs.gov

The Seabird Studies Team at the Western Ecological Research Center (WERC), with support from the Bureau of Ocean Energy Management (BOEM), completed aerial photographic surveys of the ocean off central and southern California between 2018-2021. Over 800,000 high resolution images of the ocean were collected, with the goal of extracting and counting marine birds and mammals contained within. To process this volume of images machine learning offered the best methodology, but publicly available training data did not exist for this specific purpose. Through a collaboration with Conservation Metrics, Inc. we created a labeled training dataset using Faster RCNN models via active learning and transfer learning. We then evaluated a set of candidate models trained on different label aggregation schemes, selected a final model utilizing YOLOv5 architecture, and ran the final model on the complete image dataset. Images output from the final model classified targets into seven categories: bird, dark bird, dark bird flying, light bird, fish, marine mammal, and other. We are currently reviewing the final model output for false positives and negatives to evaluate performance. Next, we will reclassify model labels to the lowest taxonomic group possible. This manual review is occurring in the USGS cloud environment (Amazon Web Services) utilizing the opensource Computer Vision Annotation Tool (CVAT). Once low taxonomic reclassification is complete, we will generate maps of species distribution and abundance to inform BOEM’s planning in advance of potential offshore wind energy development along the California coast.

Review final model output

Full image review is being conducted to determine model performance on real world data. Manual low taxonomic identification of targets identified to grouped classes by model in imagery.

Active learning, transfer learning, deep learing, convolutional neural networks (Fastern-RCNN, YOLOv5)

Collaborator Generated

Katoski, Michelle

mkatoski@usgs.gov

Cashman, Matthew mcashman@usgs.gov

Our product is an end-to-end system that is used to predict and detect sensor (sonde) fouling at USGS stream gages. The system is trained using supervised learning on multiple features derived from archived stream gage data labelled by expert field technicians. The system operated in real-time on Amazon Web Services (AWS), providing predictions every 30 minutes based off of raw data collected from the USGS AQUARIUS database. The system produces values detecting the likelihood that fouling is currently present and likelihood of fouling predicted to occur in the next 24 hours. These values are displayed on a Tableau dashboard that is connected to AWS using Amazon Athena. This dashboard also displays other stream gage network monitoring information from AQUARIUS, like time since a sonde was last visited by a technician.

In production: 1 year

Our models currently provide real time predictions in a production environment. Model tuning, training and testing continued for 6 months after deploying the first iteration in the production environment.

Long short term memory (LSTM) models

Agency Generated

We are using high frequency satellite images from the Planetscope constellation to estimate water depth in river channels. The short time lags between images allows us to average multiple scenes collected on the same day or within a couple of days to improve accuracy. In addition to established depth retrieval methods, we developed a neural network regression approach for this purpose. The training data consist of field measurements of water depth collected as part of other USGS projects on five different rivers. The neural network regression method is implemented in MATLAB using the Deep Learning Toolbox.

Planned (not in production)

Prototype code developed; manuscript describing this approach currently under review

Neural network regression

Agency Generated

USGS ScienceBase data releases: https://doi.org/10.5066/P9HZL7BZ; https://doi.org/10.5066/P9K54WDL; https://doi.org/10.5066/F7Q52NZ1; https://doi.org/10.5066/P92PNWE5; https://doi.org/10.5066/P9S4T8YM

Manuscript currently in review

This study involves using orthophotos acquired from a manned, fixed-wing aircraft and multispectral images from two different satellites to map bottom-attached (benthic) algae along the Buffalo National River in northern Arkansas. The training data for this effort consist of field observations of water depth and percent cover of benthic algae along 8-10 cross sections from two distinct reaches of the Buffalo River. These field data are used to train a bagged trees (aka random forest) classification algorithm to distinguish among four ordinal levels of algal density: none, low, medium, and high.

Planned (not in production)

Prototype code developed; manuscript describing this approach currently under review

Bagged trees (aka random forest) classification

Agency Generated

Will be made available through USGS ScienceBase catalog

Manuscript currently in review

Williamson, Tanja

tnwillia@usgs.gov

Without knowing how tile-drain extent (sub-surface agricultural drainage) has changed with time, it is difficult to differentiate how streamflow and water quality have changed as a result of spatial extent and characteristics of tile-drain networks. Our method delineates tile drains in satellite imagery, providing a way to look at historical imagery and to use satellite data to maintain an up-to-date geospatial layer of tile drain extent in basins of interest. We use panchromatic imagery that is processed using a UNet model that was trained on a library of panchromatic images on which visible tile-drain networks had been traced. Our workflow uses a combination of python scripting that is encapsulated in a Jupyter notebook; the entire process is open source.

We have the final version of the Unet model. We are working on a manscript to document the model and workflow.

Code and publication are still in progress. Questions 7 on are answered for the future.

convolutional neural networks

will be - but not yet there.

Casazza, Michael

mike_casazza@usgs.gov

Cory Overton, coverton@usgs.gov

The model we developed provides a highly accurate daily classification of waterfowl behavior into 8 life history states/movement patterns using hourly GPS relocations and, optionally, remotely sensed habitat data. This will provide waterfowl researchers and managers a tool for real-time capable rapid assessments and notification of important life history events to improve research and management outcomes and reduce project operational costs.

Planned (not in production)

extreme gradient boosted classification, stochastic gradient decent (LinearLearner®), multi-layer perceptron

Agency Generated

Upon manuscript acceptance, and IPDS review

Upon manuscript acceptance, and IPDS review

Arundel, Samantha

sarundel@usgs.gov

The goal of this project is to create a database of summit spot elevations from the HTMC labeled for summits in CONUS.

In production: more than 1 year

convolutional neural networks

Agency Generated

Arundel, Samantha

sarundel@usgs.gov

The objective of this project is to use DL tools to extract terrain features.

In production: more than 1 year

convolutional neural networks

Agency Generated

dewitz@usgs.gov

NLCD uses AI/ML to develop Landcover across all 50 states. The system includes HPC processes, cloud services, and local resources to create thematic and continuous field classifications. These classifications serve as the base for users and federal agencies across the nation to provide wildlife habitat estimates, urban runoff estimates, population growth, etc. etc.

cnns, nns, and decision trees

Bagstad, Kenneth

kjbagstad@usgs.gov

"ARIES is an international research project based at the Basque Centre for Climate Change (Bilbao, Spain), to which USGS has been a long-term collaborator. ARIES uses semantics and machine reasoning to enable AI-assisted multidisciplinary, integrated modeling of coupled human-natural systems. See https://aries.integratedmodelling.org/ and https://docs.integratedmodelling.org/technote/ for full details. ARIES is a full-stack solution for integrated modelling, supporting the production, curation, linking and deployment of scientific artifacts such as datasets, data services, modular model components and distributed computational services. Its purpose is to ensure — by design rather than just intent — that the pool of such artifacts constitutes a seamless knowledge commons, readily actionable (by humans and machines) through a full realization of the linked data paradigm augmented with semantics and powered by artificial intelligence. This design enables automation of a wide range of modeling tasks that would normally require human experts to perform. ARIES’ underlying software stack, called k.LAB, includes client and server components that support the creation, maintenance and use of a distributed semantic web platform where scientific information can be stored, published, curated and connected. The software is licensed through the Affero General Public License (AGPL) v.3.0."

In production: more than 1 year

Project originated in 2007, with increasing public functionality (IDE: 2013, Web Explorer: 2018, first Web Application: 2021)

Semantics, machine reasoning, various ML algorithms (e.g., KNN, Bayesian structural learning)

ajlynch@usgs.gov

Daniel Wieferich, dwieferich@usgs.gov

We applied coupled manual and machine learning methods to an expansive literature set for major global inland fisheries to explore opportunities for improving user efficiency for linking anthropogenic drivers of environmental change to direct impacts. This work informs the relative influence of threats in the development of a global inland fisheries assessment using boosted regression trees to derive a spatially-explicit risk index of stressors.

In production: more than 1 year

Manuscript in prep for publication

Natural language processing; boosted regression trees

Ecosystems / CSS

ajlynch@usgs.gov

Craig Paukert, cpaukert@usgs.gov

The Fish and Climate Change Database (FiCli) is a comprehensive database of peer-reviewed literature compiled through an extensive, systematic primary literature review to identify English-language, peer-reviewed journal publications with projected and documented examples of climate change impacts on inland fishes globally. We are currently exploring options to automate certain portions of the review process to increase our efficiency in maintaining and updating the database.

In production: less than 1 year

natural language processing

kkowalski@usgs.gov

Alexandra Bozimowski, abozimowski@usgs.gov

Wetland managers are restoring coastal wetland habitats in the Great Lakes, and often seek more information on when and how fish access restored habitats. Terabytes of hydroacoustic data on fish movement need to be analyzed more efficiently, so a collaboration between USGS, USFWS, and the University of Michigan is developing a machine learning model (MLM) that identifies, tracks, and quantifies fish movement. The completed model will read proprietary sonar image files, convert them to a universal file format (i.e., .mp4), place bounding boxes around individual fish detected by the model, and track them across consecutive image frames to determine bi-directional movement. The model uses training data and TensorFlow-based convolutional neural networks for object detection. Post-processing uses sonar geometry to estimate length of individual fish, and output files include labeled videos showing bounding boxes (.avi format), model metrics (.txt format), an enumeration of bi-directional fish movement (i.e., left or right) in .csv format, and individual fish length estimates (.csv format). Model output will allow USGS researchers to estimate fish habitat use and associated community metrics in restored wetland habitats. This information will support USFWS and other management agencies restoring coastal wetland habitats.

In production: less than 1 year

Currently adding additional parameters and testing model accuracy on different training datasets

convolutional neural networks

Agency Generated

Wieferich, Daniel

dwieferich@usgs.gov

Wiltermuth, Mark, mwiltermuth@usgs.gov

Species distribution models are developed for 271 fluvial fish species in their native ranges of the conterminous United States. Boosted Regression Tree (BRT) models were used to develop presence/absence predictions for each of the National Hydrography Dataset Plus Version 2.1 stream segments within a species' native range. Landscape data that describe the natural variation (e.g., slope, precip) and anthropogenic impacts (e.g., stream fragmentation) were summarized to stream segments and used as predictor variables. Native species ranges were used to geographically constrain distribution modeling efforts. R Version 4.0.2 (or newer) with ‘dismo’ and ‘labdsv’ packages are used for modeling.

In production: less than 6 months

in review for release, additional modeling techniques such as Baysian Additive Regression Tree (BART) and Maxent approaches are also being explored

boosted regression tree (BRT)

Agency Generated

https://doi.org/10.5066/P9V390V2 (doi will be activated after review completed)

jsmith@usgs.gov

Alison Appling, aappling@usgs.gov

Once developed, the system will input watershed characteristics (soils, land cover), land use (road salt application) and meteorological timeseries, and output predictions of specific conductance (SC) for inland stream reaches in the Delaware River Basin (DRB). The model will be trained using SC sample data from within the DRB. The resulting model will allow for predictions in ungaged locations and time periods, and allow for an evaluation of salinity exposure in these stream reaches. The model will be built using pyTorch on the USGS Tallgrass supercomputer.

Planned (not in production)

in development

random forests, recurrent graph convolutional neural networks

Agency Generated

Our processed training and testing data are not yet ready to be shared. The raw data are publicly available at the link below and on NWIS for the listed years. Harmonized Water Quality Portal data: https://www.sciencebase.gov/catalog/item/5e010424e4b0b207aa033d8c

In development here: https://github.com/USGS-R/drb-inland-salinity-ml

ggorski@usgs.gov

Alison Appling, aappling@usgs.gov

We are developing a machine learning model to make predictions of the 250 mg/L isochlor (salt front location) within the Delaware River Estuary. The model will be driven by river discharge into the estuary, tidal forcings, and meterological data from several points throughout the estuary. Model predictions will be compared with a process-based, hydrodynamic model, COAWST. The machine learning model is currently in development, but it will consist of a recurrent neural network architecture built using tools from pyTorch.

Planned (not in production)

Have initial results, plan to have presentation-ready results by 6/2022

recurrent neural networks

From Another Agency

Once data are processed and prepared they will be released in a publicly available data release

In development here: https://github.com/USGS-R/drb-estuary-salinity-ml

Oliver, Samantha

soliver@usgs.gov

Alison Appling, aappling@usgs.gov

We published a machine learning model to make water temperature predictions at 456 reaches in the Delaware River Basin. The recurrent graph convolutional network (RGCN) was pre-trained with predictions from a coupled process-based model that predicts stream flow and temperature (the Precipitation Runoff Modeling System with the coupled Stream Network Temperature Model or PRMS-SNTemp).

In production: more than 1 year

Model published in Jia et al 2021

recurrent graph convolutional neural network

jzwart@usgs.gov