Search USGS⌕ Search

SEARCH · Search USGS

Results for “Machine Learning with Applications”

Search indexed USGS publications on groundwater, aquifers, geologic maps, mineral resources and earthquakes. Explore source records by subject and place.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 37 records · Page 2Linked to original sources
USGS2025

A real-time fish detection system for partially dewatered fish to support selective fish passage

Recent advances in fish transportation technologies and deep machine learning-based fish classification have created an opportunity for real-time, autonomous fish sorting through a selective passage mechanism. This research presents a case study of a novel application that utilizes deep machine learning to detect partially dewatered fish exiting an Archimedes Screw Fish Lift (ASFL). A MobileNet SSD model was trained on images of partially dewatered fish volitionally passing through an ASFL. Then, this model was integrated with a network video recorder to monitor video from the ASFL. Additional models were also trained using images from a similar fish scanning device to test the feasibility of this approach for fish classification. Open source software and edge computing design principles were employed to ensure that the system is capable of fast data processing. The findings from this research demonstrate that such a system integrated with an ASFL can support real-time fish detection. This research contributes to the goal of automated data collection in a selective fish passage system and presents a viable path towards realizing optical fish sorting.

Michigan↗
USGS2022

Machine learning for understanding inland water quantity, quality, and ecology

This chapter provides an overview of machine learning models and their applications to the science of inland waters. Such models serve a wide range of purposes for science and management: predicting water quality, quantity, or ecological dynamics across space, time, or hypothetical scenarios; vetting and distilling raw data for further modeling or analysis; generating and exploring hypotheses; estimating physically or biologically meaningful parameters for use in further modeling; and revealing patterns in complex, multidimensional data or model outputs. An important research frontier is the injection of limnological knowledge into machine-learning models, which has shown great promise for increasing such models’ accuracy, trustworthiness, and interpretability. Here we describe a few of the most powerful machine learning tools, describe best practices for employing these tools and injecting knowledge guidance, and give examples of their applications to advance understanding of inland waters.

Book chapter↗
USGS2024

Evaluation of streamflow predictions from LSTM models in water- and energy-limited regions in the United States

The application of Long Short-Term Memory (LSTM) models for streamflow predictions has been an area of rapid development, supported by advancements in computing technology, increasing availability of spatiotemporal data, and availability of historical data that allows for training data-driven LSTM models. Several studies have focused on improving the performance of LSTM models; however, few studies have assessed the applicability of these LSTM models across different hydroclimate regions. This study investigated the single-basin trained local (one model for each basin), multi-basin trained regional (one model for one region), and grand (one model for several regions) models for predicting daily streamflow in water-limited Great Basin (18 basins) and energy-limited New England (27 basins) regions in the United States using the CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) data set. The results show a general pattern of higher accuracy in daily streamflow predictions from the regional model when compared to local or grand models for most basins in the New England region. For the Great Basin region, local models provided smaller errors for most basins and substantially lower for those basins with relatively larger errors from the regional and grand models. The evaluation of one-layer and three-layer LSTM network architectures trained with 1-day lag information indicates that the addition of model complexity by increasing the number of layers may not necessarily increase the model skill for improving streamflow predictions. Findings from our study highlight the strengths and limitations of LSTM models across contrasting hydroclimate regions in the United States, which could be useful for local and regional scale decisions using standalone or potential integration of data-driven LSTM models with physics-based hydrological models.

Machine Learning with Applications↗
USGS2019

Artificial intelligence and avian influenza: Using machine learning to enhance active surveillance for avian influenza viruses

Influenza A viruses are one of the most significant viral groups globally with substantial impacts on human, domestic animal and wildlife health. Wild birds are the natural reservoirs for these viruses, and active surveillance within wild bird populations provides critical information about viral evolution forming the basis of risk assessments and countermeasure development. Unfortunately, active surveillance programs are often resource‐intensive, and thus, enhancing programs for increased efficiency is paramount. Machine learning, a branch of artificial intelligence applications, provides statistical learning procedures that can be used to gain novel insights into disease surveillance systems. We use a form of machine learning, gradient boosted trees, to estimate the probability of isolating avian influenza viruses (AIV) from wild bird samples collected during surveillance for AIVs from 2006 to 2011 in the United States. We examined several predictive features including age, sex, bird type, geographic location and matrix gene rRT‐PCR results. Our final model had high predictive power and only included geographic location and rRT‐PCR results as important predictors. The highest predicted viral isolation probability was for samples collected from the north‐central states and the south‐eastern region of Alaska. Lower rRT‐PCR Ct‐values are associated with increased likelihood of AIV isolation, and the model estimated 16% probability of isolating AIV from samples declared negative (i.e., ≥35 Ct‐value) using the rRT‐PCR screening test and standard protocols. Our model can be used to prioritize previously collected samples for isolation and rapidly evaluate AIV surveillance designs to maximize the probability of viral isolation given limited resources and laboratory capacity.

Transboundary and Emerging Diseases↗
USGS2022

Machine learned daily life history classification using low frequency tracking data and automated modelling pipelines: Application to North American waterfowl

Background Identifying animal behaviors, life history states, and movement patterns is a prerequisite for many animal behavior analyses and effective management of wildlife and habitats. Most approaches classify short-term movement patterns with high frequency location or accelerometry data. However, patterns reflecting life history across longer time scales can have greater relevance to species biology or management needs, especially when available in near real-time. Given limitations in collecting and using such data to accurately classify complex behaviors in the long-term, we used hourly GPS data from 5 waterfowl species to produce daily activity classifications with machine-learned models using “automated modelling pipelines”. Methods Automated pipelines are computer-generated code that complete many tasks including feature engineering, multi-framework model development, training, validation, and hyperparameter tuning to produce daily classifications from eight activity patterns reflecting waterfowl life history or movement states. We developed several input features for modeling grouped into three broad categories, hereafter “feature sets”: GPS locations, habitat information, and movement history. Each feature set used different data sources or data collected across different time intervals to develop the “features” (independent variables) used in models. Results Automated modelling pipelines rapidly developed easily reproducible data preprocessing and analysis steps, identification and optimization of the best performing model and provided outputs for interpreting feature importance. Unequal expression of life history states caused unbalanced classes, so we evaluated feature set importance using a weighted F1-score to balance model recall and precision among individual classes. Although the best model using the least restrictive feature set (only 24 hourly relocations in a day) produced effective classifications (weighted F1 = 0.887), models using all feature sets performed substantially better (weighted F1 = 0.95), particularly for rarer but demographically more impactful life history states (i.e., nesting). Conclusions Automated pipelines generated models producing highly accurate classifications of complex daily activity patterns using relatively low frequency GPS and incorporating more classes than previous GPS studies. Near real-time classification is possible which is ideal for time-sensitive needs such as identifying reproduction. Including habitat and longer sequences of spatial information produced more accurate classifications but incurred slight delays in processing.

Movement Ecology↗
USGS2025

Mapping bedrock outcrops in the Sierra Nevada Mountains (California, USA) using machine learning

Accurate, high-resolution maps of bedrock outcrops can be valuable for applications such as models of land–atmosphere interactions, mineral assessments, ecosystem mapping, and hazard mapping. The increasing availability of high-resolution imagery can be coupled with machine learning techniques to improve regional bedrock outcrop maps. In the United States, the existing 30 m U.S. Geological Survey (USGS) National Land Cover Database (NLCD) tends to misestimate extents of barren land, which includes bedrock outcrops. This impacts many calculations beyond bedrock mapping, including soil carbon storage, hydrologic modeling, and erosion susceptibility. Here, we tested if a machine learning (ML) model could more accurately map exposed bedrock than NLCD across the entire Sierra Nevada Mountains (California, USA). The ML model was trained to identify pixels that are likely bedrock from 0.6 m imagery from the National Agriculture Imagery Program (NAIP). First, we labeled exposed bedrock at twenty sites covering more than 83 km 2 (0.13%) of the Sierra Nevada region. These labels were then used to train and test the model, which gave 83% precision and 78% recall, with a 90% overall accuracy of correctly predicting bedrock. We used the trained model to map bedrock outcrops across the entire Sierra Nevada region and compared the ML map with the NLCD map. At the twenty labeled sites, we found the NLCD barren land class, even though it includes more than just bedrock outcrops, accounted for only 41% and 40% of mapped bedrock from our labels and ML predictions, respectively. This substantial difference illustrates that ML bedrock models can have a role in improving land-cover maps, like NLCD, for a range of science applications.

California↗
USGS2020

Well predictive performance of play-wide and Subarea Random Forest models for Bakken productivity

In recent years, geologists and petroleum engineers have struggled to clearly identify the mechanisms that drive productivity in horizontal, hydraulically-fractured oil wells producing from the middle member of the Bakken formation. This paper fills a gap in the literature by showing how this play’s heterogeneity affects factors that drive well productivity. It is important because understanding the relative strength of productivity drivers and how predictors vary spatially facilitates best-practices for well site selection and well completion design. The paper describes an application of the Random Forest (RF) machine learning technique to identify these mechanisms and to evaluate their importance across 9 subareas of the North Dakota portion of the Bakken play. The study examined productivity of 7311 wells initiating production from 2010 through 2017. Well productivity varied considerably across the 9 subareas within the play, so it was not surprising that the dominant predictors, the initial 180-day water cut and the 30-day initial gas production, vary spatially to mirror local conditions that strongly affect well productivity. The relative importance of well completion predictor variables, that is, the numbers of fractures stages per well, volume of injected proppant per stage, volume of injected fluids per stage, and lateral length, varied considerably across the subareas. Statistical permutation tests are presented that generally confirm the importance rankings. Subarea Random Forest models explained from 50 percent to 82 percent of the variation in productivity test samples while the play-wide model explained 73 percent of the test sample well productivity. Weakness in the predictive ability of the Random Forest models are traced to the limited variability in the training data. Implications of the empirical findings regarding the Bakken play for operators and for research and government institutions are discussed in the concluding section.

Montana, North Dakota, South Dakota↗
USGS2024

Wildfire probability estimated from recent climate and fine fuels across the big sagebrush region

Background Wildfire is a major proximate cause of historical and ongoing losses of intact big sagebrush ( Artemisia tridentata Nutt.) plant communities and declines in sagebrush obligate wildlife species. In recent decades, fire return intervals have shortened and area burned has increased in some areas, and habitat degradation is occurring where post-fire re-establishment of sagebrush is hindered by invasive annual grasses. In coming decades, the changing climate may accelerate these wildfire and invasive feedbacks, although projecting future wildfire dynamics requires a better understanding of long-term wildfire drivers across the big sagebrush region. Here, we integrated wildfire observations with climate and vegetation data to derive a statistical model for the entire big sagebrush region that represents how annual wildfire probability is influenced by climate and fine fuel characteristics. Results Wildfire frequency varied significantly across the sagebrush region, and our statistical model represented much of that variation. Biomass of annual and perennial grasses and forbs, which we used as proxies for fine fuels, influenced wildfire probability. Wildfire probability was highest in areas with high annual forb and grass biomass, which is consistent with the well-documented phenomenon of increased wildfire following annual grass invasion. The effects of annuals on wildfire probability were strongest in places with dry summers. Wildfire probability varied with the biomass of perennial grasses and forbs and was highest at intermediate biomass levels. Climate, which varies substantially across the sagebrush region, was also predictive of wildfire probability, and predictions were highest in areas with a low proportion of precipitation received in summer, intermediate precipitation, and high temperature. Conclusions We developed a carefully validated model that contains relatively simple and biologically plausible relationships, with the goal of adequate performance under novel conditions so that useful projections of average annual wildfire probability can be made given general changes in conditions. Previous studies on the impacts of vegetation and climate on wildfire probability in sagebrush ecosystems have generally used more complex machine learning approaches and have usually been applicable to only portions of the sagebrush region. Therefore, our model complements existing work and forms an additional tool for understanding future wildfire and ecological dynamics across the sagebrush region.

Fire Ecology↗
USGS2026

The use of fluorite geochemistry and machine learning to identify critical mineral systems

Fluorite (CaF 2 ) is a potential pathfinder to critical mineral and rare earth element (REE) deposits but its application has been limited to a narrow range of mineralization types. I show that fluorite is a robust recorder of mineralization fertility by applying statistical and machine-learning methods to a new global fluorite geochemical database. Distinct median rare earth and trace element patterns are observed among deposit types and genetic environments. Fluorite associated with carbonatites and REE deposits are relatively enriched in Sr and have minimal Eu anomalies. These characteristics define new bivariate discrimination diagrams that correctly identify 78% of carbonatite-related fluorite and 88% of fluorite from REE deposits. Random forest classifiers were developed for a wide range of mineralization types and genetic settings. Trained solely on rare earth element patterns, these models achieve accuracies of 77–79%. Higher classification accuracies (up to 88–96%) are obtained when including elements such as Sr, highlighting the significance of trace elements for optimal fluorite classification. The recognition of diagnostic fluorite compositional fingerprints, particularly in REE-fertile systems, underscores its potential as a pathfinder and indicator for critical mineral exploration in F-bearing environments.

Mineralium Deposita↗
USGS2020

Improving the accessibility and transferability of machine learning algorithms for identification of animals in camera trap images: MLWIC2

Motion‐activated wildlife cameras (or “camera traps”) are frequently used to remotely and noninvasively observe animals. The vast number of images collected from camera trap projects has prompted some biologists to employ machine learning algorithms to automatically recognize species in these images, or at least filter‐out images that do not contain animals. These approaches are often limited by model transferability, as a model trained to recognize species from one location might not work as well for the same species in different locations. Furthermore, these methods often require advanced computational skills, making them inaccessible to many biologists. We used 3 million camera trap images from 18 studies in 10 states across the United States of America to train two deep neural networks, one that recognizes 58 species, the “species model,” and one that determines if an image is empty or if it contains an animal, the “empty‐animal model.” Our species model and empty‐animal model had accuracies of 96.8% and 97.3%, respectively. Furthermore, the models performed well on some out‐of‐sample datasets, as the species model had 91% accuracy on species from Canada (accuracy range 36%–91% across all out‐of‐sample datasets) and the empty‐animal model achieved an accuracy of 91%–94% on out‐of‐sample datasets from different continents. Our software addresses some of the limitations of using machine learning to classify images from camera traps. By including many species from several locations, our species model is potentially applicable to many camera trap studies in North America. We also found that our empty‐animal model can facilitate removal of images without animals globally. We provide the trained models in an R package (MLWIC2: Machine Learning for Wildlife Image Classification in R), which contains Shiny Applications that allow scientists with minimal programming experience to use trained models and train new models in six neural network architectures with varying depths.

Ecology and Evolution↗
USGS2025

Phytoplankton biomass dynamics in wet (2019) and dry (2023) years in Lake Pontchartrain estuary, Louisiana from Sentinel 2-MSI and PACE-OCI observations

This study provides a comprehensive assessment of phytoplankton biomass dynamics in Lake Pontchartrain, Louisiana, by combining monthly water quality data with multispectral and hyperspectral satellite observations using a machine learning algorithm. A machine learning model based on Variational Autoencoder (VAE), globally applicable, was used to estimate phytoplankton biomass via chlorophyll- a (Chl- a ) from Sentinel 2-MSI and NASA's new hyperspectral mission, PACE-OCI, enabling the first direct comparison between the two sensors. The model performed well in this complex estuarine system, with higher accuracy from PACE-OCI (MAE: 1.48, RMSE: 10.40, slope: 0.87) than Sentinel 2-MSI (MAE: 1.57, RMSE: 11.08, slope: 0.83). This approach enabled continuous high-resolution monitoring of phytoplankton biomass across space and time. Comparative analysis of 2019, a wet year with Bonnet Carré Spillway (BCS) openings, and 2023, a dry year with extremely low riverine inputs, revealed distinct biomass dynamics. In 2019, BCS discharge initially suppressed Chl- a within turbid waters (<5 mg Chl- a m −3 ) but later acted as a nutrient and hydrodynamic driver, transporting nutrients toward the lake outlet and Mississippi coast, promoting high biomass (25–45 mg Chl- a m −3 ) near the entrance. In contrast, dry conditions in 2023 led to more frequent-than-expected high biomass (>35 mg Chl- a m −3 ), persisting in the lake center. Similar spatial patterns were observed again in 2024, revealed for the first time by PACE-OCI. This study demonstrates the value of satellite-derived observations for capturing transient phytoplankton biomass events and highlights the potential of PACE-OCI's hyperspectral capabilities to better distinguish phytoplankton communities and improve understanding of their responses to freshwater inflows and associated processes driving pulses into estuaries.

Louisiana↗
USGS2020

Leveraging deep learning in global 24/7 real-time earthquake monitoring at the National Earthquake Information Center

Machine‐learning algorithms continue to show promise in their application to seismic processing. The U.S. Geological Survey National Earthquake Information Center (NEIC) is exploring the adoption of these tools to aid in simultaneous local, regional, and global real‐time earthquake monitoring. As a first step, we describe a simple framework to incorporate deep‐learning tools into NEIC operations. Automatic seismic arrival detections made from standard picking methods (e.g., short‐term average/long‐term average [STA/LTA]) are fed to trained neural network models to improve automatic seismic‐arrival (pick) timing and estimate seismic‐arrival phase type and source‐station distances. These additional data are used to improve the capabilities of the NEIC associator. We compile a dataset of 1.3 million seismic‐phase arrivals that represent a globally distributed set of source‐station paths covering a range of phase types, magnitudes, and source distances. We train three separate convolutional neural network models to predict arrival time onset, phase type, and distance. We validate the performance of the trained networks on a subset of our existing dataset and further extend validation by exploring the model performance when applied to NEIC automatic pick data feeds. We show that the information provided by these models can be useful in downstream event processing, specifically in seismic‐phase association, resulting in reduced false associations and improved location estimates.

Seismological Research Letters↗
USGS2024

Treed Gaussian processes for animal movement modeling

Wildlife telemetry data may be used to answer a diverse range of questions relevant to wildlife ecology and management. One challenge to modeling telemetry data is that animal movement often varies greatly in pattern over time, and current continuous-time modeling approaches to handle such nonstationarity require bespoke and often complex models that may pose barriers to practitioner implementation. We demonstrate a novel application of treed Gaussian process (TGP) modeling, a Bayesian machine learning approach that automatically captures the nonstationarity and abrupt transitions present in animal movement. The machine learning formulation of TGPs enables modeling to be nearly automated, while their Bayesian formulation allows for the derivation of movement descriptors with associated uncertainty measures. We demonstrate the use of an existing R package to implement TGPs using the familiar Markov chain Monte Carlo algorithm. We then use estimated movement trajectories to derive movement descriptors that can be compared across individuals and populations. We applied the TGP model to a case study of lesser prairie-chickens ( Tympanuchus pallidicinctus ) to demonstrate the benefits of TGP modeling and compared distance traveled and residence times across lesser prairie-chicken individuals and populations. For broad usability, we outline all steps necessary for practitioners to specify relevant movement descriptors (e.g., turn angles, speed, contact points) and apply TGP modeling and trajectory comparison to their own telemetry datasets. Combining the predictive power of machine learning and the statistical inference of Bayesian methods to model movement trajectories allows for the estimation of statistically comparable movement descriptors from telemetry studies. Our use of an accessible R package allows practitioners to model trajectories and estimate movement descriptors, facilitating the use of telemetry data to answer applied management questions.

Ecology and Evolution↗
USGS2025

Mapping predicted ecological states at landscape scales using remote sensing data and machine learning

Dryland ecosystems, covering 45% of the Earth's land and supporting over one-third of the global population, face significant threats from land degradation and ecological state change. Managing these ecosystems is complex, and science-based frameworks like Ecological Site Descriptions and state-and-transition models are essential tools for guiding decisions to support ecological health while maintaining stakeholder values such as grazing, wildlife, and recreation. However, alignment of these frameworks with smaller scale soil survey maps limits their applicability to broader ecological processes. Here, we extend these frameworks to larger landscapes with a machine learning approach that integrates large-scale, high-resolution vegetation data with identified ecological states from a data-driven state-and-transition model developed for a landscape-scale Ecological Site Group. A “global” model, which used combined inputs from multiple remotely sensed datasets, outperformed individual dataset models based on evaluation with independent data. Ecological state maps generated through this approach broaden the utility of state-and-transition models across Ecological Site Groups, providing a more spatially robust tool for land management at watershed and larger landscape scales. These methods, and the associated ecological state maps, can help meet critical needs for improved land condition assessments that support development of resource management plans and help identify priority areas for restoration and conservation.

Arizona, Colorado, New Mexico, Utah, Wyoming↗
USGS2025

Determination of provenance soil type using inductively coupled plasma mass spectrometry (ICP-MS) analyses of Pinus ponderosa ash

This study demonstrates the feasibility of determining soil provenance from tree ash composition using elemental analysis and chemometric techniques. To date, no published studies have applied chemometric approaches to classify ash for provenance determination following forest fires. In this work, Pinus ponderosa ash was analyzed to distinguish samples based on soil type and geographic location. Pinus ponderosa , a widely distributed pine species in the western United States where wildfires are prevalent, was selected as a model system. Needles were collected from trees grown in five distinct soil types across northern Arizona and Colorado, then dry-ashed under controlled conditions. Classification was performed using three preprocessing techniques and five machine learning algorithms, including hierarchical modeling structures to optimize separation. Partial least squares discriminant analysis (PLS-DA) following a Box-Cox transformation yielded the highest classification accuracy, achieving a prediction kappa value of 0.98 for soil type identification. However, classification performance decreased when distinguishing both soil type and geographic location, indicating that additional variability may influence predictive accuracy in broader applications. These findings highlight the potential of inductively coupled plasma mass spectrometry (ICP-MS) and machine learning for post-wildfire forensic analysis and environmental monitoring.

Applied Spectroscopy↗
USGS2025

Real-time oil spill concentration assessment through fluorescence imaging and deep learning

Oil spills may pose severe ecological and socioeconomic threats, necessitating rapid and accurate environmental assessment. Traditional assessment methods used to determine the extent of a spill including gas chromatography-mass spectrometry, satellite imaging, and visual surveys, are often time-consuming, expensive, and limited by weather conditions or sampling constraints. Furthermore, these methods frequently struggle to provide real-time data crucial for prompt decision-making during spill emergencies. This study addresses these limitations by combining fluorescence imaging, deep learning, a mobile application, and a data management system for automated and real-time oil spill assessment. Our approach leverages a convolutional neural network architecture for feature extraction coupled with a custom regression model, trained and evaluated on a self-curated comprehensive dataset of 1,530 fluorescence images from two distinct oil types, a napthalenic crude oil and an aromatic-napthalenic crude oil, at concentrations ranging from 0 to 500 mg/L. The proposed approach demonstrates superior performance compared to both traditional machine learning models and more complex deep learning architectures, achieving an R² score of 0.9958 and RMSE of 9.28. The application enables rapid, cost-effective field measurements with robust data tracking and analysis capabilities. This research advances oil spill monitoring technology with a scalable solution that balances accuracy, speed, and accessibility for real-time environmental assessment and emergency response.

Journal of Hazardous Materials↗
USGS2020

Modeling false positives

Many of the models we are concerned with included explicit descriptions of false negative errors. However, false positive errors can also be commin in practice, especially in citizen science applications where observer skill is highly variable. In addition, new methods which determine detection based on statistical classification or machine learning methods are also prone to false positive errors which must be accounted for. An early treatment of the false positive detection problem by Royle & Link (2006) recognized that false positive errors can be accommodated by a mixture model for detection probability: one value of detection at occupied sites and another non-zero value at unoccupied sites. This model has been extended greatly in recent years to include more informative data about false positives including validation or confirmation data (Miller et al. 2011) and multiple detection methods, among others. A new frontier for the application of false positives models lies in the use of modern technologies such as bioacoustics for efficient automated monitoring. For these technologies to realize their promise there must be improvements in automated processing of the vast quantities of output produced. Statistical classification methods (machine learning) are fallible and necessarily produce false positive detections. Therefore models which account for this process are necessary (Chambert et al. 2017). It stands to reason that false positives will need to be accounted for in other new technologies that rely on automated digital processing, including eDNA, genetic barcoding, and automated detection in remote camera studies. We devise a new occupancy model that integrates data from bioacoustics sampling with an occupancy model. This integrated model allows occupancy probability to inform species classification of samples and vice versa bioacoustics detection data inform occupancy. We provide a proof of concept for this new model in this chapter. As the core hierarchical model for the false positives models covered in this chapter are just ordinary occupancy models, extension of the ideas to open systems poses no technical challenges. We provide a suite of illustrations of these extensions. Perhaps the most prominent mechanism that leads to false positive errors it he mis-classification of species detections, or the confusion of one species for another. Very little work has been done on developing models based on this mechanistic understanding although Chambert et al. (2018) develop this idea as a 2-species occupancy model with error. We believe one important area of future research is to extend these ideas to truly multi-species systems.

Book chapter↗
USGS2024

Spatio-temporal ecological models via physics-informed neural networks for studying chronic wasting disease

To mitigate the negative effects of emerging wildlife diseases in biodiversity and public health it is critical to accurately forecast pathogen dissemination while incorporating relevant spatio-temporal covariates. Forecasting spatio-temporal processes can often be improved by incorporating scientific knowledge about the dynamics of the process using physical models. Ecological diffusion equations are often used to model epidemiological processes of wildlife diseases where environmental factors play a role in disease spread. Physics-informed neural networks (PINN) are deep learning algorithms that constrain neural network predictions based on physical laws and therefore are powerful forecasting models useful even in cases of limited and imperfect training data. In this paper, we develop a novel ecological modeling tool using PINNs, which fits a feedforward neural network and simultaneously performs parameter identification in a partial differential equation (PDE) with varying coefficients. We demonstrate the applicability of our model by comparing it with the commonly used Bayesian stochastic partial differential equation method and traditional machine learning approaches, showing that our proposed model exhibits superior prediction and forecasting performance when modeling chronic wasting disease in deer in Wisconsin. Furthermore, our model provides the opportunity to obtain scientific insights into spatiotemporal covariates affecting spread and growth of diseases. This work contributes to future machine learning and statistical methodology development by studying spatio-temporal processes enhanced by prior physical knowledge.

Spatial Statistics↗