📄 agriculture

Sector-Specific Machine Learning Models for Short-Term Sugarcane Yield Forecasting Using NDVI at Plot Level

This study demonstrates that sector-specific machine learning models utilizing age-segmented Sentinel-2 NDVI time series outperform global models and traditional expert estimates in providing accurate, early, plot-level sugarcane yield forecasts, thereby enabling more effective agricultural management and operational planning.

Original authors: JOEL MORALES, David Mera, Juan Francisco Pec, José Luis Quemé

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: JOEL MORALES, David Mera, Juan Francisco Pec, José Luis Quemé

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 sugarcane farmer in Guatemala. You have thousands of tiny plots of green gold stretching across the landscape. Your job is to guess how much sugar you'll harvest from each specific patch before the trucks even arrive. Traditionally, you'd send a manager out with a clipboard and a pair of eyes to make a guess. But the paper by Joel Morales and his team says: "Hold on, let's try something cooler."

They built a digital crystal ball using Machine Learning (smart computer programs) and satellite photos to predict the harvest, called "tons of cane per hectare" (TCH), way before the crops are cut.

The Big Idea: One Size Does Not Fit All

The researchers tested two ways to build their crystal ball.

First, they tried a "Global Model." Imagine trying to teach one single student to predict the weather for every single city in the world at once. You give the computer all the data from every farm, mix it into one giant smoothie, and ask it to guess the yield for everyone. It's a generalist approach.

Second, they tried a "Sector-Specific Model." This is like hiring a tiny, specialized team of experts, where one team knows only Farm A, another knows only Farm B, and so on. Each team learns the specific quirks, soil, and history of their own little neighborhood.

The Verdict: The paper explicitly rules out the idea that the "one-size-fits-all" global model is the best way. The Sector-Specific approach was the clear winner. While the global model stumbled with an error rate (RMSE) of 16.75 TCH, the specialized teams nailed it with an error of just 12.48 TCH. The specialized models were also much more confident, hitting a score (R²) of 0.7840 compared to the global model's 0.5724.

The Secret Sauce: The Satellite's "Green Pulse"

How do these computer teams know so much? They don't look at the soil; they look at the NDVI. Think of NDVI as a "greenness score" that satellites measure. Healthy, happy plants reflect light in a way that makes them glow green on the satellite's screen.

The paper argues against just taking a single snapshot. Instead, they watched the crops grow like a time-lapse movie. They broke the 12-month growing cycle into 45-day chunks (like chapters in a book) and tracked how the "greenness" changed in each chapter.

They found that the Elongation I stage (a specific growth phase where the stalks stretch out) was the most critical chapter. It's like the plant's "teenage growth spurt." If you know how the plant is doing during this specific stretch, you can predict the final harvest much better than if you just look at the plant when it's a baby or when it's fully grown.

The "When" Question: How Early Can We Guess?

A common question is: "Can we know the harvest size right at the start?" The paper suggests that no, you can't get a reliable answer immediately.

At 3 months: The computer is still guessing wildly. The error is huge (24.75 TCH). It's like trying to guess the final height of a baby just by looking at its first few weeks.
At 5 months: This is the magic turning point. The paper shows that by the fifth month, the predictions become reliable enough to beat human experts. The error drops to 14.13 TCH.
At 12 months: As the crop gets older, the prediction gets even sharper, dropping the error to 12.48 TCH.

Beating the Human Expert

Here is the fun part: The paper compares their computer to the actual farm managers who have been doing this for years. The managers usually make their estimates just one month before harvest.

Human Manager Error: 15.47 TCH.
Computer (at 5 months) Error: 14.13 TCH.

The paper demonstrates that the computer can beat the human expert seven months earlier than the human usually makes their guess. And the computer is more accurate, too. The human guesses tend to be influenced by what happened last year or just a "gut feeling," which can be wrong if the weather was weird. The computer, however, looks at the actual "green pulse" of the plants, catching things like water stress that a human eye might miss.

The Catch (and the Future)

The paper is very honest about a limitation. The "Sector-Specific" teams are so good, but they need a perfect movie of the crop's growth. If clouds block the satellite view too often, the computer can't see the "green pulse," and it can't build its specialized team. Currently, this method works for about 52% of the plots because of cloudy skies.

For the rest, the "Global Model" (the generalist) is still useful as a backup plan. But the paper suggests that in the future, they might combine the green photos with radar (which can see through clouds) to make the system work everywhere, all the time.

The Bottom Line

The paper proves that if you want to know how much sugarcane you'll get, you shouldn't use a generic rule for the whole farm. You need to build a custom, neighborhood-specific model that watches the crop's "green heartbeat" through the satellite lens. By the fifth month of growth, this method is already more accurate than the seasoned experts, giving farmers a powerful tool to plan their harvest, save money, and manage their fields like a pro.

Technical Summary: Sector-Specific Machine Learning Models for Short-Term Sugarcane Yield Forecasting

Problem Statement
Accurate, plot-level forecasting of sugarcane yield (measured in tons of cane per hectare, TCH) is critical for optimizing agricultural management, resource allocation, and operational planning in Guatemala's sugar industry. Traditional estimation methods rely on historical data or field biometrics, which are logistically complex and fail to capture the spatial and temporal variability inherent in extensive cultivation areas with staggered cropping cycles. Furthermore, existing macroscopic models often overlook mid-season technological shifts and localized biophysical conditions, while field-based sampling is labor-intensive and difficult to scale. Current Machine Learning (ML) approaches frequently operate at aggregated spatial scales or rely on static observations, limiting their utility for real-time, plot-level decision-making. The core challenge addressed is the inability of existing frameworks to simultaneously integrate plot-level estimation, temporal modeling of vegetation indices across phenological stages, and continuous prediction updates as new satellite data becomes available.

Methodology
The study developed an ML-based framework to forecast short-term sugarcane yields using age-segmented Normalized Difference Vegetation Index (NDVI) derived from Sentinel-2 (S2) satellite imagery.

Data Integration: The researchers constructed an integrated dataset linking productivity records from 2,132 sugarcane plots across six harvest seasons (2016–17 to 2021–22) with S2-derived NDVI time series. The study area covered 46,754 hectares managed by Ingenio Magdalena.
Feature Engineering: NDVI observations were aggregated into 45-day temporal bins aligned with key phenological stages (Tillering, Elongation I, Elongation II, and Ripening). For each bin, five statistical features were extracted: Mean, Median, Quartile 1 (Q1), Quartile 3 (Q3), and Area Under the Curve (AUC). To ensure temporal consistency, only plots with complete NDVI trajectories across all bins were retained.
Modeling Strategies: Two distinct spatial modeling schemes were evaluated:
1. Global Scheme: A single universal model trained on all available observations regardless of location.
2. Sector-Specific Scheme: Independent models developed for individual agricultural sectors (47% of total sectors met the minimum data threshold of five observations) to capture local biophysical variability.
Algorithm Evaluation: Ten regression algorithms were tested, including Boosted Linear Model (BstLm), Support Vector Machine with Polynomial Kernel (svmPoly), Random Forest (rf), Sparse Partial Least Square (spls), and Gaussian Process Regression (gaussprPoly). Hyperparameters were optimized via 5-fold cross-validation.
Validation: Models were evaluated on an independent test set (Zafra 2021–22) using Root Mean Square Error (RMSE) and the coefficient of determination ( $R^2$ ). Performance was also benchmarked against pre-harvest expert estimates provided by farm managers. SHapley Additive exPlanation (SHAP) analysis was applied to interpret feature importance.

Key Results

Model Performance: The sector-specific approach significantly outperformed the global model. On the independent test set, the sector-specific scheme achieved an RMSE of 12.48 TCH and an $R^2$ of 0.7840, compared to the global model's RMSE of 16.75 TCH and $R^2$ of 0.5724.
Algorithm Selection: The most frequently selected algorithms were Sparse Partial Least Square (spls) and Support Vector Machine with Polynomial Kernel (svmPoly). The global scheme favored Gaussian Process Regression with Polynomial Kernel (gaussprPoly), but this was less effective than the tailored sector models.
Feature Importance: SHAP analysis identified Median NDVI as the dominant predictive feature (selected in 51% of models). The Elongation I (EI) phenological stage was determined to be the most influential period for yield estimation, aligning with the phase of peak metabolic activity and water demand.
Temporal Reliability: Reliable forecasts were achievable starting from the fifth month of crop growth (RMSE = 14.13 TCH). Prediction accuracy improved progressively as the crop matured, with the error margin narrowing significantly by the ninth and twelfth months.
Comparison with Experts: The sector-specific framework outperformed traditional expert estimations (RMSE = 15.47 TCH). Notably, the ML system provided higher accuracy earlier in the season (by the fifth month) compared to expert estimates, which are typically conducted only one month prior to harvest.

Significance and Claims
The paper claims that localized, sector-specific ML models combined with temporal NDVI dynamics provide a scalable methodology for precision agriculture that surpasses both global models and traditional expert estimations. The framework enables:

Proactive Management: By generating reliable forecasts from the fifth month of growth, the system allows for earlier adjustments to irrigation, fertilization, and harvesting plans, rather than waiting for end-of-cycle analysis.
Quantification of Variability: The approach captures the complex interactions between environmental conditions and management practices, reducing the uncertainty associated with subjective visual assessments, particularly during periods of water stress.
Operational Utility: The system supports budgeting and operational planning by providing plot-level yield targets that are more accurate than historical averages or subjective manager estimates.

The authors note a current limitation: the reliance on optical satellite data means that persistent cloud cover limits the deployment of the sector-specific approach to approximately 52% of plots. They propose that future work will address this by integrating Synthetic Aperture Radar (SAR) data from Sentinel-1 to ensure continuous monitoring regardless of atmospheric conditions. The study concludes that this paradigm shift transforms yield forecasting from a static, end-of-cycle estimation into a dynamic decision-support tool.