📄 agriculture

Evaluating Machine Learning and Deep Learning Models for Predicting Rice Vegetation Condition in Acadia, Louisiana

This study evaluated seven machine learning and deep learning models using 12,544 rice field-seasons in Acadia, Louisiana, finding that Gradient Boosting achieved the highest predictive accuracy (R² = 0.960) and that the Normalized Difference Water Index (NDWI) was the dominant predictor of rice vegetation condition, far surpassing other environmental variables.

Original authors: Desmond Fosuhene, Dorcas Twumwaa Gyan, Kingsford Kobina Annan, Theodora Mary Fletcher, Bhuwan Thapa, Ebenezer Nana Kwaku Boateng

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Desmond Fosuhene, Dorcas Twumwaa Gyan, Kingsford Kobina Annan, Theodora Mary Fletcher, Bhuwan Thapa, Ebenezer Nana Kwaku Boateng

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

Rice is a staple food for billions of people, yet growing it is a delicate balancing act that depends heavily on water and weather. In many parts of the world, farmers lack the tools to quickly check how their crops are faring, especially when climate patterns shift unpredictably. To solve this, scientists have turned to satellites, which act as eyes in the sky, capturing images of the Earth's surface over time. By analyzing the light reflected from fields, researchers can calculate specific numbers that reveal how green a crop is and how much water it holds. These numbers, known as spectral indices, allow experts to monitor plant health without ever stepping foot in a field. The challenge lies in turning these vast streams of satellite data into clear, reliable predictions about crop conditions, a task that requires sophisticated computer programs capable of finding patterns in complex information.

In the rice-growing heartland of Acadia Parish, Louisiana, a team of researchers set out to test which computer programs work best at predicting the health of rice plants. They focused on a region where rice is grown in flooded fields, a method that creates a unique relationship between the water on the surface and the green leaves above. To do this, they gathered a massive amount of data spanning eighteen years, from 2008 to 2025. They combined satellite images from the Landsat program, which has been photographing the Earth for decades, with local records of rainfall, temperature, and the shape of the land. The team looked at more than twelve thousand separate field-seasons, creating a detailed picture of how rice fields changed from year to year. Their goal was not just to guess the future, but to understand exactly which factors—such as the amount of water on the ground or the temperature of the air—were the most important for explaining why some rice fields looked healthier than others in any given season.

To find the best way to make these predictions, the researchers tested seven different types of computer models. Some of these models were based on simple mathematical lines, while others were complex systems inspired by the way human brains learn, known as deep learning. They also tested powerful tools that build many small decision trees to reach a final answer, a method called ensemble learning. Before running the tests, the team carefully removed the influence of the specific year from their data. This step was crucial because it allowed them to see how individual fields differed from one another within the same season, rather than just seeing how the whole region changed from one year to the next. When they compared the results, every single model performed with high accuracy, but the ensemble methods, particularly one called Gradient Boosting, edged out the others slightly. This top-performing model was able to explain nearly all of the predictable variation in the rice's health, proving that these computer approaches are highly effective for this type of agricultural monitoring.

The most surprising discovery, however, was not about which computer program worked best, but about what the computer actually looked at to make its decisions. The analysis revealed that the single most important factor for predicting rice health was a specific measurement of surface water, known as the Normalized Difference Water Index. This index measures how much open water is visible on the field. In contrast, other factors that one might expect to be critical, such as the total amount of rain that fell during the growing season or the average temperature, had almost no impact on the model's predictions once the year-to-year differences were accounted for. The researchers found that in this specific environment, where farmers carefully manage the flooding of their fields, the amount of water sitting on the surface tells the whole story of the crop's condition. The computer learned that as the rice grows and covers the water, the signal from the surface water changes in a predictable way that perfectly matches the health of the plants.

This study demonstrates that for flood-irrigated rice, the key to understanding crop health lies in watching the water. The researchers showed that by using long-term satellite records and robust computer models, it is possible to create a reliable system for monitoring rice fields over vast areas. While the most complex deep learning models performed well, the slightly simpler ensemble methods proved to be the most efficient and accurate tools for the job. The findings suggest that farmers and agricultural planners do not need to rely on a wide array of complicated weather data to track rice growth in managed irrigation systems; instead, focusing on the spectral signature of the water itself provides the clearest picture. This approach offers a practical and scalable way to keep an eye on one of the world's most important food sources, ensuring that the information needed to support these crops is available quickly and accurately.

Technical Summary: Evaluating Machine Learning and Deep Learning Models for Predicting Rice Vegetation Condition in Acadia, Louisiana

Problem Statement
Rice production is increasingly vulnerable to climate variability, yet many rice-growing regions lack the ground-observation infrastructure necessary for timely crop-condition assessments. While remote sensing offers a scalable solution, integrating multi-source satellite and environmental datasets remains challenging due to differences in spatial resolution, temporal synchronization, and radiometric characteristics. Furthermore, existing literature often evaluates only a limited number of algorithms or focuses on short-term predictions, leaving a gap in understanding which modeling approaches perform most consistently for long-term vegetation monitoring in irrigated landscapes. Additionally, conventional machine learning (ML) and deep learning (DL) models often fail to explicitly account for the geospatial context, such as spatial dependencies and environmental heterogeneity, which can lead to biased predictions.

Methodology
This study evaluated seven modeling approaches to predict rice vegetation condition (proxied by the Normalized Difference Vegetation Index, NDVI) in Acadia Parish, Louisiana, using data from 2008 to 2025.

Data Sources: The analysis integrated 12,544 field-season observations derived from:
- Satellite Imagery: Landsat Collection 2 Level-2 surface reflectance (Landsat 5, 7, 8, and 9) to calculate spectral indices: NDVI, Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDMI), and Normalized Difference Water Index (NDWI).
- Environmental Variables: Monthly rainfall and 2-m air temperature from ERA5-Land reanalysis, and terrain slope from the SRTM digital elevation model.
- Crop Data: Rice field boundaries derived from the USDA-NASS Cropland Data Layer (CDL).
Preprocessing and Feature Engineering:
- Data were processed in Google Earth Engine, with cloud masking and monthly median composites generated.
- A "year fixed-effects" (demeaning) procedure was applied to all variables. This removed inter-annual confounds (e.g., sensor differences, long-term climate trends) to isolate within-year variations in vegetation condition relative to environmental predictors.
- The final dataset included field area, NDWI, NDMI, cumulative growing-season rainfall, mean growing-season temperature, and terrain slope as predictors.
Models Evaluated: Seven models were trained and tested using five-fold random cross-validation:
- Linear Models: Linear Regression, Ridge Regression.
- Ensemble Tree-Based Models: Random Forest, Gradient Boosting, XGBoost, LightGBM.
- Deep Learning: A feed-forward Deep Neural Network (Multilayer Perceptron) with three hidden layers.
Evaluation Metrics: Performance was assessed using the coefficient of determination ( $R^2$ ), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). Variable importance was determined using both impurity-based metrics (from tree models) and permutation feature importance on an independent hold-out dataset.

Key Results

Model Performance: All seven models achieved high predictive accuracy. Ensemble tree-based methods outperformed linear and deep learning approaches, though the differences were marginal.
- Gradient Boosting achieved the highest performance ( $R^2 = 0.960 \pm 0.002$ ).
- LightGBM, Random Forest, and XGBoost followed closely ( $R^2 \approx 0.958–0.959$ ).
- The Deep Neural Network achieved $R^2 = 0.956 \pm 0.002$ .
- Linear and Ridge Regression yielded the lowest performance ( $R^2 \approx 0.947–0.948$ ).
Variable Importance: Permutation feature importance analysis revealed a dominant predictor:
- NDWI (Normalized Difference Water Index) was the overwhelmingly dominant predictor, with a permutation importance ( $\Delta R^2$ ) of 1.708.
- NDMI was the second most important ( $\Delta R^2 = 0.034$ ).
- All other variables (field area, rainfall, temperature, slope) had negligible contributions ( $\Delta R^2 \leq 0.002$ ).
Correlation Structure: A strong negative correlation ( $r = -0.963$ ) was observed between NDVI and NDWI, reflecting the phenological shift from flooded fields (high NDWI, low NDVI) to dense vegetation (low NDWI, high NDVI).

Key Contributions

Comparative Model Evaluation: The study provides a systematic comparison of multiple ML and DL algorithms for long-term rice vegetation monitoring, demonstrating that ensemble tree-based methods offer a marginal but consistent accuracy advantage over deep learning and linear models for this specific tabular geospatial task.
Identification of Dominant Predictors: The research quantifies the relative importance of environmental drivers, establishing that surface-water spectral dynamics (captured by NDWI) account for nearly all predictable variation in rice vegetation condition in flood-irrigated systems after controlling for temporal confounds.
GeoAI Framework Application: The study situates predictive modeling within a GeoAI framework, highlighting the necessity of controlling for temporal confounds (via year fixed-effects) to accurately assess the explanatory power of environmental variables.

Significance and Claims
The authors claim that these findings offer practical guidance for selecting robust GeoAI models and spectral predictors for satellite-based agricultural monitoring. Specifically:

Operational Utility: Ensemble tree-based models are identified as particularly well-suited for operational decision support due to their balance of accuracy, computational efficiency, and interpretability.
Monitoring Strategy: The dominance of NDWI suggests that routine satellite observations of surface water dynamics can effectively characterize crop development in managed irrigation systems without heavy reliance on auxiliary climate or terrain data.
Scalability: The study demonstrates that integrating long-term satellite observations with machine learning provides a scalable approach for monitoring rice vegetation conditions in irrigated agricultural landscapes.

The paper maintains a modest tone regarding its limitations, acknowledging that the use of NDVI may saturate under dense canopies, that the year-fixed-effects approach limits direct forecasting of unseen future seasons, and that the spatial resolution mismatch of input data may have obscured the contribution of terrain and climate variables. The authors suggest that future work should explore spatially explicit modeling approaches (e.g., geographically weighted regression) and validate these findings in regions with different irrigation practices and topographies.

Technical Summary: Evaluating Machine Learning and Deep Learning Models for Predicting Rice Vegetation Condition in Acadia, Louisiana

More like this