Arash Modaresi Rad

Soil inorganic carbon (SIC) is critical for carbon sequestration, infiltration, and climate modeling, yet quantifying its precise spatial distribution at continental scales remains challengings. We introduce a high‐resolution (30 m) CONUS SIC map using machine learning (ML) models trained on the ISRIC World Soil Information Service (WoSIS) database. Integrating point‐based soil data with high resolution climate and land surface variables enhances predictions and reduces uncertainty in unsampled areas over previous polygon‐based inventories. Our SIC Random Forest Regression (RFR) yielded an Root Mean Square Error (RMSE) of 15 kg/m ² , and the multi‐class classifier had an accuracy of 0.56. Our model estimates that CONUS soils store 77 ± 1.8 Pg of SIC in the top 1 m, a significant increase in SIC storage over prior inventories. We distinguish two SIC storage types: lithogenic carbonates persisting in humid regions and pedogenic carbonates, which characteristically form in arid soils.

Soil inorganic carbon (SIC) is critical for carbon sequestration, infiltration, and climate modeling, yet its formation and distribution remain poorly understood. This study introduces a high-resolution (30m) SIC map of CONUS in a Google Earth Engine app. Using climate and land surface data and leveraging machine learning (ML) models, we enhance predictions and reduce uncertainty in unsampled areas. Our SIC RFR achieved an RMSE of 15 kg/m2, and the multi-class classifier had an accuracy of 0.56. Our model estimates that CONUS soils store 77 ± 1.8 Pg of SIC in the top 1 m, a significant increase over prior inventories. Our findings distinguish between two key occurrences: lithogenic carbonates persisting in humid regions and pedogenic carbonates, which characteristically form in arid soils. Soil pH is the strongest predictor of calcareous soil formation, while precipitation, with a threshold of 1700 mm/year, emerges as the primary driver.

With increasing climate variability, natural disaster frequency, and populations there is an emerging need for digital awareness of the built environment. Knowing who and what is where is critical to better allocating resources in emergencies, planning for the future, and translating the outcomes of different environmental models into tangible impacts that help bridge the policy-science gap. To date, a range of open source projects have sought to represent the built environment, but each has focused on either spatial completeness (e.g. AI methods), accuracy (hand digitization in OpenStreetMap), or attribute richness (OpenAddresses). The reality is that all of these are needed for different uses, and integrating these datasets offer a unique path towards a spatial and attribute rich data product. This work describes a method for developing a best available Built Environment Asset Registry (BEAR) from a suite of open - yet disparate - data sources. To date, BEAR conflates commonly utilized, continental-scale footprint and address databases covering the contiguous US. The resulting product is developed as a cloud-native database that conforms to an extensible schema. Using this core representation, we highlight efficient access patterns for retrieving subsets of BEAR from the Lynker Spatial data portal, and examples of applying BEAR to hazard modeling and social vulnerability assessment. We show how this product can increase the skill and utility of forecasts related to flooding and fires.

Knowledge of bankfull hydraulic geometry represents an essential requirement for various applications, including accurate flood prediction, hydrological routing, river behavior analysis, river management and engineering practices, water resource management, and beyond. Our work builds upon an extensive body of literature about estimating bankfull top‐width and depth at ungauged locations to enhance the understanding of observable factors that affect these parameters. Using more than 200,000 USGS Acoustic Doppler Current Profiler (ADCP) records, we developed a method employing machine learning (ML) using discharge estimates and landscape characteristics from sources, including the National Water Model (NWM), the National Hydrologic Geospatial Fabric network (NHGF), the EPA stream characteristic data set (StreamCat), and an array of satellite and reanalysis data products. Our method achieved log‐transformed R² = 0.8 predicting bankfull depth (R² = 0.77 for in‐channel conditions) and R² = 0.76 predicting bankfull top‐width (R² = 0.66 for in‐channel conditions) in the testing data set. The depth and width predictions showed lowest skill in mountainous and plateau regions. Our analysis demonstrates the benefit of data‐driven modeling in contrast to other global scaling‐based or regional statistical methods. In summary, our study illustrates how top‐width and depth can be better predicted using ML, reanalysis streamflow simulations, hydrographic networks, and summarized geospatial data.

Machine Learning-Derived Channel Width and Depth for the National Hydrologic Geospatial Fabric in CONUS

Background The behavior of a river channel at a given cross-section is often described using power law equations to relate top width (TW), mean depth (Y), and mean velocity (V) to discharge (Q). These equations collectively define the "at a station hydraulic geometry" (AHG), with coefficients (a, c, k) and exponents (b, f, m) for the Q-TW, Q-Y, and Q-V relationships (Leopold & Maddock, 1953). While each relation can be studied independently, together they represent the full hydraulic system.

Rating curves are commonly developed through direct observation, open channel flow models, or mechanical methods, each relying on in-situ measurement. As part of a U.S. effort to provide high resolution, continental scale, flood mapping, synthetic rating curves (SRCs) were developed across the National Hydrography Dataset (NHDPlusV2) to translate flows, like those generated by the NOAA National Water Model, into river depths. This approach uses Digital Elevation Models (DEM) to define the necessary cross-sectional properties for Manning’s equation. A significant limitation, alongside an opportunity for broad improvement, has been assigning suitable roughness without local information. We applied the DEM based methodology to generate SRCs at 7,270 locations with known USGS rating curves, and calibrated roughness to minimize the error between predicted and observed flow. Subsequently, we tested several approaches based on land cover, stream order, and the hydrographic network to estimate the optimized values in a manner that can be extended to ungauged catchments. Among these, a predictive Machine Learning (ML) model based on the NHDPlusV2 network attributes demonstrated superior ability to estimate the optimized roughness with a Spearman correlation of 0.89. Sensitivity analysis showed improving accuracy of DEM and roughness is crucial for accurate estimation of the lower and mid/upper parts of SRC, respectively. Finally, we applied the predictive model over the NHDPlusV2, generating reach-level roughness estimates that can directly support national flood mapping efforts. The method is generalizable to any hydrofabric network that contains topology, however the generated values are dependent on the DEM and hydrofabric used.

Understanding of the vulnerability of populations exposed to wildfires is limited. We used an index from the U.S. Centers for Disease Control and Prevention to assess the social vulnerability of populations exposed to wildfire from 2000–2021 in California, Oregon, and Washington, which accounted for 90% of exposures in the western United States. The number of people exposed to fire from 2000–2010 to 2011–2021 increased substantially, with the largest increase, nearly 250%, for people with high social vulnerability. In Oregon and Washington, a higher percentage of exposed people were highly vulnerable (>40%) than in California (~8%). Increased social vulnerability of populations in burned areas was the primary contributor to increased exposure of the highly vulnerable in California, whereas encroachment of wildfires on vulnerable populations was the primary contributor in Oregon and Washington. Our results emphasize the importance of integrating the vulnerability of at-risk populations in wildfire mitigation and adaptation plans.

With an increasing number of continental-scale hydrologic models, the ability to evaluate performance is key to understanding uncertainty in prediction and making improvements to the model(s). In 2016, the NOAA National Water Model (NWM) was put into operations to improve the spatial and temporal resolution of hydrologic prediction in the U.S. Here, we evaluate the NWM 2.0 historical streamflow record in natural and controlled basins using the Nash Sutcliffe Efficiency metric decomposed into relative error, conditional, and unconditional bias. Each of these is evaluated in the contexts of categorized meteorologic, landscape, and anthropogenic characteristics to assess model performance and diagnose error types. Broadly speaking greater rainfall and snow coverage leads to improved performance while larger potential evapotranspiration (PET), aridity, and phase correlation reduce performance. More rainfall and phase correlation reduce overall bias, while increasing PET, aridity, snow coverage/fraction increase model bias. With respect to landscape traits, more barren and agricultural land yeild improved performance while more forest, shrubland, grassland and imperviousness tend to decrease performance. Lastly, more barren and herbaceous land tend to decrease bias, while greater imperviousness, urban, forest, and shrubland cover increase bias. The insights gained can help identify key hydrological factors in NWM predictions; enforce the need for regionalized physics and modeling; and help develop hybid post-processing methods to improve prediction. Finally, we demonstrate how the NOAA Next Generation Water Resource Modeling Framework can help reduce the structural bias through the application of heterogenous model processes and highlight opportunities for ongoing development and evaluation.