How storms carry water and carbon out of farmland

November 22, 2026

Storms wash water, soil and carbon from farmland into streams. A model of the watershed can describe this, but it has to be tuned to each stream, and for carbon it does poorly even then. This study lets a neural network learn from an untuned model and from the measurements together. The result matches a tuned model for streamflow and is clearly better for carbon.

At a glance

  • Theme: Streamflow and organic carbon in storms
  • Where: Tuckahoe Creek, a farming watershed of 221 km² by the Chesapeake Bay, United States
  • Data: Daily streamflow and organic carbon at the outlet, 2014 to 2019
  • Approach: A watershed model used untuned, with a network that follows the river and remembers the days before
  • Lead author: Dr. Venkatesh Budamala
  • Status: Published, code open
A map of the Tuckahoe Creek watershed on the eastern side of the Chesapeake Bay, with its streams and the gauge at its outlet, and three small maps of its elevation, soils and land use
The Tuckahoe Creek watershed (pink), the headwaters of the Choptank River, which flows into the Chesapeake Bay. The star marks the gauge at its outlet. Below: elevation, soil groups and land use. Figure 1 of the paper (Budamala and others, Journal of Hydrology, 2026).

The basics

In a farming region most of the water and carbon that leave the land do so in a few storms. The organic carbon travels in two forms. Carbon in particles is washed off with eroded soil and peaks with the flood. Dissolved carbon seeps through the soil and arrives later. Both shape the quality of the water downstream, here in the Chesapeake Bay.

The challenge

A model built on the physics of a watershed has dozens of settings, and it has to be tuned against local measurements before it can be trusted. That takes time and has to be done again for every watershed. A neural network needs no such model, but it needs long records to learn from, and records of carbon in streams are short.

The idea

A diagram in three steps. A model of the watershed, used as it comes and not tuned to the stream, gives the water and carbon in each part of the watershed. A network follows the river from part to part and remembers the days before. It predicts first the flow of the stream and then the carbon in particles and the dissolved carbon
The model supplies the physics, the network learns what the measurements add, and the flow is predicted before the carbon. Drawing by the lab.

The study uses the model as it comes. SWAT-C, the carbon version of a widely used watershed model, is run without tuning. For each of the 19 parts of the watershed it gives the water in the soil, the flow from groundwater, the soil washed off and the carbon in the soil, day by day. These become the inputs of a network with two parts. One follows the river from each part of the watershed to the next. The other looks back over the days before.

The network first predicts the flow of the stream. It then predicts the carbon in particles and the dissolved carbon, with the predicted flow as an extra input, because the flow is what carries them. It learned from the years 2014 to 2017 and was tested on 2019, a year it had not seen.

What was found

A bar chart of how closely four approaches follow the measurements. Streamflow: this work 0.82, the tuned model 0.82, the model with a simpler network 0.53, the untuned model 0.23. Carbon in particles: 0.68, 0.04, 0.34 and minus 2.05. Dissolved carbon: 0.74, 0.43, minus 0.05 and minus 0.01
How closely each approach follows the measurements of 2019, for streamflow and for the two forms of carbon. The measure is the Kling-Gupta efficiency: 1 is a perfect match. Figure by the lab; values from Table 1 of the paper (Budamala and others, Journal of Hydrology, 2026).

For streamflow the new approach reaches an efficiency of 0.82, the same as the model after tuning, where the untuned model alone reaches 0.23. For carbon it is well ahead: 0.68 against 0.04 for carbon in particles, and 0.74 against 0.43 for dissolved carbon. A simpler network combined with the model does worse on all three. The error in the volume of water over the year falls from 40 per cent for the untuned model to 7 per cent.

In the storms

Four time series for the year 2019: rain, streamflow, carbon in particles and dissolved carbon. Circles are the measurements. The untuned model overshoots the peaks of nearly every storm; the new approach stays close to the measurements. Five storms are marked
The test year 2019: rain (a), streamflow (b), carbon in particles (c) and dissolved carbon (d). Circles: measured. Blue: the model alone, untuned. Orange: the new approach. The shaded bands mark five storms of different kinds. Figure 3 of the paper (Budamala and others, Journal of Hydrology, 2026).

The largest storm of 2019 came on 21 to 24 March, with about 40 mm of rain and a peak flow of 26.8 m³ per second. The untuned model overshot that peak by a wide margin, and the peaks of carbon with it. The new approach stayed close to what was measured. The timing of dissolved carbon changes too. In the model alone, dissolved carbon rises and falls with the flood. In the new approach it lags behind, as carbon that travels through the soil is known to do.

What the network needs the model for

The same networks were also trained without what the model supplies. They failed, most clearly for carbon, where their skill was close to zero or below. A few years of record are too few to learn a watershed from. The untuned model is far from the measurements, but it knows where the water goes, and that is what the network builds on.

What the result does not say

The method was tested in one watershed, flat and dominated by farming, and on daily values, with a single year held back for testing. How well it carries over to other climates, sizes of watershed and kinds of land is still to be shown. It also rests on the model: where the model leaves out the process that matters, the network has less to build on.

Why it matters

Tuning is the slow part of modelling a watershed. A method that reaches the same skill without it, and does better for carbon, can be carried more easily to the many watersheds where measurements are few. Knowing how much carbon storms carry off the land matters for the water quality of estuaries and for the carbon balance of farmland.

The paper

Budamala, V., Choi, E., Kim, M., Lee, S., and Kim, H. (2026). Hierarchical prediction of streamflow and organic carbon at event scale using a physics-guided AI hybrid framework. Journal of Hydrology, 679, 136248. The work was done with colleagues at Korea University. The carbon measurements were provided by the United States Department of Agriculture.

Links

Contact

Read other projects

A longer record of soil moisture, from an older radar

NASA's SMAP gives the most trusted soil moisture from space, but only since 2015. This published study turns the measurements of an older radar into SMAP-like maps of soil moisture and vegetation, back to 2007.

Read this project
Streamflow where the gauges are few

Most of the world's rivers have few gauges or none. A network that learns along the river, guided by an untuned model of the basin, predicts their flow. A published study in three countries.

Read this project
How storms carry water and carbon out of farmland

A model of a watershed and a neural network, used together, predict the streamflow and the organic carbon that storms wash out of farmland, without tuning the model. A published study.

Read this project
Soil moisture from reflected navigation signals

Eight small NASA satellites built to watch tropical cyclones also sense how wet the soil is, several times a day. A published review of what seven years of their data can and cannot do.

Read this project
A slow groundwater response as a sign of flash drought

How long does groundwater take to follow the soil above it? Where that delay is long, flash droughts are frequent. A published study of twenty years of global data.

Read this project
Forecasting flash droughts weeks ahead

A flash drought takes hold within a few weeks, often before a drought has been declared. This project asks how far ahead it can be seen coming, and where that foresight comes from.

Read this project
How well does artificial intelligence forecast rain?

Weather models built with artificial intelligence now match traditional forecasts for wind and temperature. Rain is harder. This project tests how well they forecast it, from the next day to several weeks ahead.

Read this project
Soil moisture under the forest canopy

Satellites read the moisture of the soil almost everywhere on land, except under dense vegetation. This project developed a way of reading it under forests too, with open daily maps of the globe.

Read this project
Reading rainfall from the soil

Rain leaves a mark in the soil: after a storm, the ground is wetter. This project uses that mark, as satellites see it, to estimate how much rain fell.

Read this project
Can satellites give early warning of a slope collapse?

After a mountainside collapses, satellites are often said to have seen it coming. This project looks at what satellites can really observe on a steep, remote slope, and when.

Read this project
The water cycle on land, seen as a whole

Rain falls, part of it evaporates, part runs off, and the rest is stored. Each part is observed by different satellites; this project works on learning them together.

Read this project
How flood water changes the air above it

When a river floods, land turns into open water for days or weeks. This project asks how the lower atmosphere responds.

Read this project
Seeing rain from space to forecast the next hours

Where is it raining now, and where will it rain in the next hours? This project brings fresh satellite views of clouds and rain into artificial-intelligence weather forecasts.

Read this project
Flash droughts that reach groundwater

Some droughts arrive within weeks rather than seasons. This project asks whether such flash droughts stay in the soil or reach the groundwater below.

Read this project
How far ahead can soil moisture be predicted?

Next week's weather is hard to predict; the wetness of the soil is easier, because soil remembers. This project studies how far ahead soil moisture can be forecast around the world.

Read this project
Filling the gaps in satellite soil moisture

Satellite maps of soil moisture are full of holes. This project works on filling them, so that the maps are complete in space and time.

Read this project
How wet soil turns rain into floods

How wet the ground is before a storm decides how much of the rain runs off. This project asks when knowing soil moisture from satellites makes a flood forecast better.

Read this project
Evapotranspiration at the scale of single fields

How much water a field gives back to the air decides how much irrigation it needs. This project works on maps that tell one field from the next.

Read this project
Bringing satellites and models together to follow soil moisture

A satellite sees soil moisture only where it passes, and a model can only estimate it. This project works on combining the two, and on finding out which future observations would help most.

Read this project
Root-zone soil moisture: below what satellites see

Satellites sense soil moisture near the surface, while plants draw water from the root zone below. This project estimates what lies below from what satellites see.

Read this project
From Iran to North Korea: Investigating Flood Impacts with AI and Satellite Data Sets

This project focuses on improving weather prediction by integrating multiple satellite-based soil moisture observations into a coupled land-atmosphere modeling system. By combining radar and radiometer data, it aims to better capture land surface conditions and their influence on atmospheric processes. The project seeks to enhance predictions of temperature, humidity, and precipitation through more accurate and physically consistent representation of soil moisture dynamics.

Read this project
Multi-Sensor Satellite Data Assimilation for Numerical Weather Prediction

This project focuses on improving weather prediction by integrating multiple satellite-based soil moisture observations into a coupled land-atmosphere modeling system. By combining radar and radiometer data, it aims to better capture land surface conditions and their influence on atmospheric processes. The project seeks to enhance predictions of temperature, humidity, and precipitation through more accurate and physically consistent representation of soil moisture dynamics.

Read this project
AI and GNN-Based Streamflow Prediction in Data-Sparse and Ungauged Basins

This project focuses on advancing streamflow prediction in data-sparse and ungauged basins using AI and graph neural networks. By capturing spatial relationships between river networks and hydrological processes, it aims to overcome limitations of traditional models that rely on dense observations. The project seeks to improve prediction accuracy and generalization across regions, enabling more reliable water resource management and flood forecasting.

Read this project
AI-Driven Water Body Detection Using Satellite Constellations and SAR: A GNSS-R and Sentinel Fusion Approach

We are developing an AI based water body detection system that combines CYGNSS satellite constellation data with Sentinel optical and SAR observations. CYGNSS provides all weather and high frequency measurements, while Sentinel offers detailed spatial information. By applying AI based fusion and classification, our approach identifies both permanent and temporary surface water even under clouds, vegetation, or limited ground monitoring. This research supports flood response, drought assessment, and improved water resource management in data denied regions.

Read this project
Building Koreas First NASA Core Validation Site for Advanced Earth Observation Satellites

-

Read this project
WRF-Hydro and AI: Advancing Hydrological Modeling with Next-Generation Tools with AI

-

Read this project
Physics-Informed Large AI Models for Hydrology and the Water Cycle

-

Read this project
Multi-Scale Soil Moisture Monitoring Using ELBARA-III and Drone Radiometer

The Portable L-Band Radiometer, mounted on a drone, enables high-resolution soil moisture retrieval across diverse terrains. Operating at 1.4 GHz (L-band), it collects brightness temperature (TB) data over rice paddies and varied vegetation zones, complementing satellite missions like SMAP and SMOS. By integrating airborne observations with ground-based sensors, the system enhances spatial coverage and improves soil moisture retrieval models for climate and agricultural applications. This drone-based approach significantly increases data accuracy and efficiency compared to traditional methods.

Read this project
Developing the Long-Term Brightness Temperature Measurement Site in South Korea

Soil moisture (SM) is essential for agriculture and hydrometeorology but difficult to measure. To improve validation, Korea’s first Core Validation Site is being built in Hampyeong-gun and Naju-si with TEROS sensors, ESA’s ELBARA-III radiometer, and drone-based PoLRa. The site will provide continuous SM and temperature data, advancing satellite validation, retrieval models, and climate resilience research.

Read this project
AI-Based Water Quality Prediction for Inland Water using Satellite and Land Surface Model Data

Climate change-driven heatwaves and water cycle disruptions threaten inland water quality (WQ), necessitating efficient monitoring. Traditional methods are labor-intensive with limited coverage, prompting us to develop an AI-based model for predicting chlorophyll-a concentrations in lakes and rivers. By integrating high-resolution satellite imagery (Landsat-8/9, Sentinel-2/3) with land surface models (ERA5-Land, GLDAS, MERRA-2), our model tailors predictions to different water bodies. Future plans include incorporating socio-statistical data (e.g., population, livestock) and climate scenarios (RCP, SSP). Using AI, GIS, and high-performance computing, we explore low-concentration chlorophyll-a prediction, precipitation and flow speed impacts on WQ, transfer learning, multi-sensor data fusion, and uncertainty quantification. Beyond chlorophyll-a, we aim to extend predictions to turbidity and dissolved oxygen, providing a comprehensive AI-driven approach to monitoring and mitigating climate change effects on inland water quality.

Read this project
Harnessing Deep Learning to Predict and Decode the Mysteries of Flash Droughts (GAN/SHAP/3D-CNN with Transfer Learning)

The application of deep learning in predicting flash droughts offers a transformative approach to understanding and anticipating these rapid-onset events, significantly enhancing preparedness and response strategies. By unraveling the complex mechanisms behind flash droughts, this project aims to provide precise, timely forecasts, thereby mitigating the severe agricultural, ecological, and socioeconomic impacts associated with these phenomena.

Read this project
Streamflow and Drought Predictions over Ungaged Regions using Deep and Transfer Learning Approaches

Streamflow and flash drought predictions are essential for managing water resources and mitigating potential disasters in ungaged regions. With remotely-sensed data, deep and transfer learning approaches provide powerful tools to analyze complex hydrological data, enabling more accurate predictions and better decision-making in these areas.

Read this project
Applications of Bayesian Machine Learning in Big Data in Earth Science

Bayesian methods help us improve our guesses by using new information. In Earth science, these methods are applied to big data to better understand our planet. This approach is useful for predicting things like natural disaster patterns and climate changes. By continuously updating our knowledge with new data, we can make more accurate predictions and decisions in Earth science.

Read this project
Water Balance Budgeting with Bayesian Machine Learning

The water balance equation in Earth science, P = E + R + etc, describes the relationship between precipitation (P), evaporation (E), runoff (R), and etc (e.g., soil moisture, ground water) in a given area. Bayesian inference can be applied to solve this equation by incorporating prior knowledge and updating the probability distributions of the variables based on new data, ultimately improving water resource management and prediction.

Read this project
Integrating Earth Science and Engineering for Climate Resilience: Innovative Approaches to Infrastructure and Societal Justice

Earth science informs infrastructure development by providing insights into site suitability, resource management, and sustainable design, enhancing the resilience and long-term viability of projects. It also plays a crucial role in addressing societal justice related to climate change by helping identify vulnerable communities and develop mitigation strategies, ensuring equitable access to resources and protection from environmental hazards.

Read this project
Enhancing Earth Science Predictions through Advanced Data Assimilation Techniques

Data assimilation is vital in earth science as it integrates diverse observations and model simulations, improving the accuracy of forecasts and predictions. This process enhances our understanding of complex Earth systems, enabling better decision-making for environmental management and climate adaptation.

Read this project
Floods and Droughts Predictions using Machine Learning Approaches

Satellite data and machine learning transformed Earth science by predicting and monitoring natural disasters. This combination delivers precise and timely predictions, crucial for mitigating the impacts of events like floods and droughts.

Read this project
Data Error Characterizations

Characterizing the error of satellite data and land surface models is vital in Earth science, as it ensures the accuracy and reliability of information used for monitoring and predicting environmental phenomena. By understanding these errors, scientists can refine data interpretation, enhance models, and ultimately make better-informed decisions about the Earth's complex systems.

Read this project
Developing Algorithms to Improve the Temporal Sampling of Satellite Data

Enhancing the temporal repeat of satellite data for obtaining soil moisture information is a vital research area due to its implications for agriculture, water resource management, climate change research, and ecosystem health. It helps in making informed decisions, increasing productivity, and reducing the impact of natural disasters, as well as contributing to our understanding of the global climate system.

Read this project
Exploring the Impact of Human Activities on the Subdaily Global Terrestrial Water Cycle

Humans have been modifying the Earth's surface for thousands of years, with practices like clearing forests for agriculture and creating uniform land covers. But how do these changes impact the subdaily global terrestrial water cycle? That's the question a project aims to answer.

Read this project
Satellite Image Disaggregation with Machine Learning

Microwave soil moisture data is critical for agriculture, weather, and climate modeling, but has low spatial resolution. Disaggregation via machine learning can improve resolution, offering detailed local soil moisture data. Machine learning can handle complex relationships between microwave signals and soil moisture.

Read this project