📄 agriculture

Satellite-derived and regional weather data as scalable alternatives to on-site stations for forecasting corn tar spot

This study demonstrates that freely available NASA POWER satellite data and regional Mesonet networks serve as viable, scalable alternatives to costly on-site weather stations for forecasting corn tar spot severity, achieving comparable prediction accuracy across multiple modeling frameworks.

Original authors: Malleswari Kachireddy, Carlos C. Gongora-Canul, Mariela Fernandez-Campos, Brenden Lane, Mohammad R Jahanshahi, C. D. Cruz

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Malleswari Kachireddy, Carlos C. Gongora-Canul, Mariela Fernandez-Campos, Brenden Lane, Mohammad R Jahanshahi, C. D. Cruz

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a corn detective trying to solve a mystery: when will the "tar spot" fungus show up to ruin a cornfield? This nasty fungus, which looks like little black specks of tar on the leaves, has been causing billions of dollars in damage across North America. To catch it early, farmers usually need to know exactly what the weather is doing right where their corn is growing.

For years, the only way to get this super-precise weather info was to install a fancy, expensive weather station right in the middle of the field. It's like hiring a private security guard for every single house in a neighborhood just to check the temperature. But these guards are pricey, they can break, and they don't cover every single house.

So, the researchers at Purdue University asked a big question: Can we use the "free" weather data from space (satellites) or from a regional network of weather stations instead of hiring a private guard for every field?

Here is what they found, broken down simply:

The Great Weather Swap Test

The team set up a massive experiment across nine different corn fields in Indiana over four years (2021–2024). They compared three different ways of getting weather data:

The Private Guard: An on-site sensor (ATMOS 41) sitting right in the corn.
The Neighborhood Watch: A regional network called the Purdue Mesonet (stations about 15–30 km away).
The Space Spy: Satellite data from NASA's POWER project, which covers the whole globe.

They fed all this weather data into three different computer "brains" (math models) to see which one could best predict how bad the tar spot would get.

The Results: Space is the New Hero

The paper suggests that satellite data is a viable substitute for the expensive on-site sensors. Here is the breakdown of how the "Space Spy" compared to the "Private Guard":

Temperature: The satellite data was a great match. It agreed with the on-site sensor about 81% of the time (a correlation of 0.81). It was a little bit warmer than the ground sensor (by about 2.09°C), but the computer models could handle that small difference easily.
Humidity: This was the tricky part. The satellite data was only 61% correlated with the ground sensor. It consistently guessed the humidity was about 17.69% lower than what the ground sensor felt. It's like the satellite thinking the air is drier than it really is.
Rain: This was the big failure. The satellite data and the ground sensors barely agreed at all (correlation of only 0.25). Rain is too patchy and local for a giant satellite to see perfectly. It's like trying to guess if a specific person in a crowd is holding an umbrella by looking at a photo of the whole city.

The Twist: Even though the satellite data was "wrong" about the humidity and rain, the computer models still worked just as well as the ones using the perfect ground data. The models were smart enough to use the pattern of the weather (like how the temperature changes over 21 days) rather than needing the exact number.

The Best Computer Brain

The researchers tested three types of math models:

MLR (Multiple Linear Regression): A standard way of connecting weather to disease.
Bayesian Estimation: A method that gives you a "confidence range" (like saying "it's probably 30% sick, but could be between 20% and 40%").
SARIMA: A time-traveling model that looks at how the disease moves over time, like predicting the next step in a dance based on the last few steps.

The paper found that SARIMA was the best at capturing the "flow" of the disease. It produced smoother, more realistic predictions of how the tar spot spreads day by day. However, all three models performed similarly well in terms of error rates. The difference in accuracy between using satellite data and ground data was tiny—usually less than 0.25 units of error.

What the Paper Rules Out

The paper explicitly argues against the idea that you must have an on-site weather station to get a good forecast. It shows that the high cost and sparse coverage of these stations are not necessary barriers anymore.

However, it also rules out the idea that satellite data is perfect. It clearly states that precipitation (rain) data from satellites is too unreliable to be trusted on its own for this specific job. If you are relying only on rain data from space, the forecast might be shaky.

How Sure Are They?

The authors are very confident about the temperature and humidity results because they measured them directly across nine different "site-years" (different fields and years). They used a strict testing method called "leave-one-site-year-out," which means they trained the model on some fields and tested it on a completely different field it had never seen before. This proves the models aren't just memorizing the data; they are actually learning the rules.

They suggest that this method could work across the whole U.S. Corn Belt, but they admit they haven't tested it everywhere yet. They also note that while the models work well for the middle and top of the corn plant, they are a little less accurate for the bottom leaves (where the disease often starts), which is a problem they hope to fix in the future.

The Bottom Line

You don't need to spend thousands of dollars on a weather station in every cornfield to predict tar spot. The "Space Spy" (NASA POWER) can do the job almost as well as the "Private Guard," especially for temperature. While the satellite isn't great at guessing rain, the computer models are smart enough to compensate. This means farmers and experts can now use free, global data to keep an eye on this dangerous fungus, making it easier to protect crops without breaking the bank.

Technical Summary: Satellite-Derived and Regional Weather Data as Scalable Alternatives to On-Site Stations for Forecasting Corn Tar Spot

Problem Statement
Tar spot, caused by the obligate biotrophic fungus Phyllachora maydis, has emerged as a devastating foliar disease of corn (Zea mays L.) in North America, causing estimated losses of 7.1 million metric tons in the northern U.S. in 2024 alone. Effective management relies heavily on weather-based forecasting to optimize fungicide timing, as infection is driven by specific environmental conditions: moderate temperatures (16–20°C), high relative humidity (>75%), and extended leaf wetness. However, current forecasting models depend on data from on-site weather stations. These stations are costly to install and maintain, prone to damage, and provide spatially sparse coverage, limiting the scalability of precision disease management. While freely available satellite-derived (e.g., NASA POWER) and regional network (e.g., Mesonet) data offer potential alternatives, their reliability for tar spot forecasting has not been rigorously evaluated against on-site measurements, nor has the impact of data source choice on model accuracy across different forecasting frameworks been established.

Methodology
The study developed a structured forecasting pipeline integrating multi-source weather data with georeferenced disease severity observations collected across nine site-years (2021–2024) in Indiana, USA.

Data Sources: Daily weather data were obtained from three sources spanning different spatial scales:
1. On-site: ATMOS 41 sensors deployed within experimental plots.
2. Regional: The Purdue Mesonet network (nearest station, ~15–30 km away).
3. Satellite: NASA Prediction Of Worldwide Energy Resources (POWER) project data (~50 km resolution).
Disease Data: Severity was assessed at three canopy positions (lower, middle, upper) at 7–14 day intervals. Discrete ratings were interpolated into continuous daily severity series using nonlinear logistic population-growth curves.
Modeling Frameworks: Three distinct modeling approaches were implemented and compared:
1. Multiple Linear Regression (MLR): Used forward stepwise selection to relate weather predictors to severity.
2. Bayesian Estimation: Employed Markov chain Monte Carlo (MCMC) sampling to generate probabilistic forecasts with uncertainty quantification.
3. Time-Series Models (ARIMA/SARIMA): Designed to capture temporal dependence and seasonal structures in epidemic progression.
Validation: To ensure the evaluation reflected true forecasting rather than within-season interpolation, a "leave-one-site-year-out" cross-validation scheme was used. Entire site-years were withheld from training and predicted solely by models fitted to remaining years.

Key Contributions

First Comparative Evaluation: This is the first study to rigorously compare NASA POWER satellite data and regional Mesonet data against on-site sensors specifically for tar spot forecasting.
Framework Diversity: The study evaluates three distinct modeling paradigms (MLR, Bayesian, and SARIMA) to determine if the choice of weather data source affects performance differently across frameworks.
Scalability Assessment: It quantifies the feasibility of replacing expensive, site-specific infrastructure with freely available, broad-coverage data sources for operational disease management.

Results

Data Agreement:
- Temperature: NASA POWER showed strong agreement with on-site sensors (Pearson $r = 0.81$ ; Lin's concordance $\rho_c = 0.74$ ), though with a warm bias ( $+2.09^\circ$ C).
- Relative Humidity: Agreement was moderate ( $r = 0.61$ ; $\rho_c = 0.37$ ). NASA POWER systematically underestimated humidity by approximately 18 percentage points compared to on-site sensors.
- Precipitation: Agreement was poor across all comparisons ( $r \le 0.25$ ), reflecting the high spatial variability of rainfall and limitations of both point sensors and coarse satellite grids.
Model Performance:
- Accuracy: All three modeling frameworks (MLR, Bayesian, SARIMA) achieved comparable prediction errors. Models using NASA POWER data performed similarly to those using on-site ATMOS data, with Root Mean Square Error (RMSE) differences typically less than 0.25 units.
- Temporal Structure: SARIMA models best captured the temporal structure of epidemic progression, producing smoother severity trajectories and effectively modeling the weekly seasonal cycles inherent in disease development.
- Predictors: Temperature (21-day moving average), relative humidity, and solar radiation were identified as the most significant predictors. Despite the systematic bias in NASA POWER humidity data, MLR models accommodated these biases effectively, maintaining predictive accuracy.
Canopy Effects: Prediction errors varied by canopy position, with lower canopy models often showing higher coefficients of determination ( $R^2$ up to 0.982), though this layer is also the most operationally critical for early management.

Significance and Claims
The paper claims that freely available gridded weather data (specifically NASA POWER) and regional networks (Purdue Mesonet) are viable, scalable alternatives to on-site sensors for forecasting tar spot severity. The study concludes that:

Substitution is Feasible: Replacing on-site weather inputs with satellite or regional data does not result in a meaningful loss of forecasting accuracy across MLR, Bayesian, or SARIMA frameworks.
Operational Viability: The ability to generate forecasts for any geographic location without local infrastructure supports the expansion of disease surveillance to areas currently lacking weather stations, addressing a critical gap in precision agriculture.
Caveats: The authors modestly note that while temperature and humidity data are reliable substitutes, precipitation data remains a limitation due to poor agreement. Furthermore, the study is limited to Indiana, and the authors state that generalizability across the broader U.S. Corn Belt requires further validation. They emphasize that while satellite forecasts offer improved spatial targeting, they should be deployed with attention to local risk factors rather than as standalone triggers, particularly given the lack of a formally established economic severity threshold for tar spot.