📄 agriculture

Data-Driven Spatial Targeting for Sustainable Intensification of Rice Systems in the Philippines

This study demonstrates that data-driven, spatially targeted nitrogen management strategies, developed using machine learning and validated on-farm, can significantly close rice yield gaps in the Philippines' Iloilo and Nueva Ecija provinces while achieving 86% of potential production gains with only 15% increased fertilizer use compared to blanket approaches.

Original authors: Royette Santos, Hari Nayak, Pompe Sta Cruz, Anton Urfels, Virender Kumar

Published 2026-09-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Royette Santos, Hari Nayak, Pompe Sta Cruz, Anton Urfels, Virender Kumar

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 the daily bread for more than half the world's population, a crop so fundamental that its success or failure dictates the stability of nations and the hunger levels of families. In the Philippines, this grain is not just a food source but a pillar of the economy and culture, feeding over 100 million people. Yet, despite the country's deep reliance on rice, farmers are not harvesting as much as the land and climate could theoretically allow. This shortfall, known as a yield gap, occurs when the actual harvest falls significantly below what is possible with the best available knowledge and resources. Closing this gap is critical for national food security, but doing so requires understanding exactly why some fields thrive while others struggle. The challenge lies in the fact that every farm is different; soil, weather, and how a farmer manages their crops vary from one field to the next, making a single, one-size-fits-all solution ineffective.

A team of researchers set out to solve this puzzle in two of the Philippines' most important rice-growing provinces, Iloilo and Nueva Ecija. Instead of guessing what farmers needed, they gathered detailed information from hundreds of actual rice fields, recording everything from the type of seed planted to the amount of fertilizer used. They then used a sophisticated computer system, a form of artificial intelligence, to sift through this massive amount of data. This system did not just predict future harvests; it learned to explain exactly which factors were holding back production in specific locations. By analyzing the data, the researchers discovered that the most common reason for low yields was simply not using enough nitrogen fertilizer, a key nutrient that helps plants grow. However, the amount needed varied wildly from field to field. Some farmers were applying too little, starving their crops, while others were applying too much, wasting money and potentially harming the environment.

The study tested three different ways of helping farmers improve their harvests. The first was the current government approach, which suggests the same amount of fertilizer for everyone in a region. The second was a data-driven suggestion that applied a single, optimized amount to all farms. The third, and most innovative, was a targeted approach. This method used the computer's analysis to identify only the specific fields that were actually limited by a lack of nitrogen and recommended extra fertilizer only for those spots. The results were striking. The targeted approach predicted that farmers could increase their harvests by more than one ton per hectare in Iloilo and nearly two tons per hectare in Nueva Ecija. Remarkably, this massive gain in food production could be achieved while using significantly less fertilizer overall compared to the other methods. In fact, the targeted strategy could produce 86 percent of the potential extra rice while requiring only a fraction of the additional fertilizer that blanket recommendations would demand.

To ensure these computer predictions were not just theoretical, the researchers took their findings back to the fields. They set up side-by-side trials where one plot followed the farmer's usual methods and the adjacent plot followed the new, data-driven recommendations. The results confirmed the model's accuracy. In Iloilo, the plots using the new advice produced 15 percent more rice, and in Nueva Ecija, they produced nearly 12 percent more. This real-world validation proved that the computer's diagnosis was correct: many farmers were indeed under-fertilizing their crops, and a precise, tailored adjustment could unlock significant productivity without needing to expand farmland or use excessive chemicals.

The implications of this work extend far beyond a single harvest. It suggests a fundamental shift in how agricultural support should be delivered. For decades, policies have often relied on broad subsidies, giving the same amount of fertilizer support to every registered farmer regardless of their specific needs. This study argues that such a uniform approach is inefficient. Instead, resources should be directed like a spotlight, focusing only on the areas where they will have the greatest impact. By using data to pinpoint exactly which fields are struggling with nutrient shortages, governments and aid organizations can maximize the return on their investment. This means more rice for the nation, better incomes for farmers, and less environmental strain from unnecessary fertilizer use.

The researchers also highlighted that this approach is particularly vital given the rising cost of fertilizers and the country's heavy reliance on rice imports. By closing the gap between what farmers currently harvest and what they could harvest with better management, the Philippines could significantly reduce its need to buy rice from other countries. The study demonstrates that the tools to achieve this are already available. By combining extensive on-the-ground data with modern analytical methods, it is possible to move away from guesswork and toward a system of precision agriculture that is both economically smart and environmentally sustainable. The path to food security, the study concludes, lies not in doing more of the same, but in doing the right thing, in the right place, at the right time.

Technical Summary: Data-Driven Spatial Targeting for Sustainable Intensification of Rice Systems in the Philippines

Problem Statement
Rice is the cornerstone of food security in the Philippines, yet the nation remains the world's largest rice importer due to persistent yield gaps that hinder national self-sufficiency. These gaps, defined as the difference between attainable yields (top 10% of farmers) and average farmer yields, are driven largely by suboptimal agronomic management rather than a lack of arable land. While previous studies have utilized frontier-based approaches (e.g., stochastic frontier analysis) to quantify technical efficiency, these methods often struggle to capture the complex, non-linear interactions between management practices and environmental conditions. Furthermore, existing policy interventions, such as the Rice Competitiveness Enhancement Program (RCEP), largely rely on "blanket" top-down recommendations that fail to account for the spatial heterogeneity of production environments. Consequently, there is a critical need for spatially explicit, data-driven interventions that can identify specific productivity drivers and optimize resource use—particularly nitrogen (N) management—without expanding cultivated area.

Methodology
The study employed a bottom-up, data-driven framework integrating interpretable machine learning (ML) with spatial analysis across two major irrigated rice provinces: Iloilo (Visayas) and Nueva Ecija (Central Luzon).

Data Collection: A stratified random sampling survey was conducted across 50 barangays, collecting data from 538 farmers. After rigorous data curation to remove outliers and errors, the final dataset comprised 514 field observations covering both dry (DS) and wet (WS) seasons. Data included detailed management practices (variety, water, nutrient, pest/weed management) and yield outcomes.
Machine Learning Modeling: Province- and season-specific Random Forest (RF) models were developed to predict rice yields based on management inputs. Model performance was evaluated using Out-of-Bag (OOB) $R^2$ and Root Mean Square Error (RMSE).
Interpretability and Driver Identification: The "black box" nature of RF models was addressed using Shapley Additive Explanations (SHAP) to quantify the contribution of individual features to yield predictions. Global feature importance was ranked by mean absolute SHAP values.
Optimization and Scenario Analysis: Partial Dependence Plots (PDP) and Individual Conditional Expectation (ICE) curves were generated to analyze the relationship between Nitrogen application rates and yield. This allowed for the identification of "data-driven optimum N rates" for specific fields.
Ex-Ante Scenario Evaluation: Three management scenarios were simulated to estimate yield gains and economic impacts:
1. Local Government Unit (LGU) Blanket: Applying current standard LGU recommendations (90–120 kg N ha⁻¹ depending on province/season).
2. Data-Driven Blanket: Applying a uniform, economically optimal N rate derived from ICE analysis across all fields.
3. Spatially Targeted: Applying optimal N rates only to fields identified as N-limited (those with negative SHAP values for N), leaving other fields unchanged.
Spatial Mapping: Field-specific yield improvement potentials were aggregated in QGIS to generate spatial maps highlighting high-priority intervention zones.
On-Farm Validation: Side-by-side trials were established in Nueva Ecija (DS 2024) and Iloilo (DS 2025) comparing Farmer Practice (FP) against the data-driven N recommendations (Demo) to validate the ML-derived strategies under real-world conditions.

Key Results

Yield Gaps: Average farmer yields reached only 68–71% of attainable levels, leaving yield gaps of 29–32%. In Nueva Ecija, gaps were larger (up to 3.1 t ha⁻¹ in DS) compared to Iloilo (~1.7 t ha⁻¹), reflecting greater seasonal variability and management heterogeneity in the former.
Dominant Drivers: Nitrogen management was identified as the dominant yield-limiting factor across both provinces and seasons. In Nueva Ecija, N rate was followed by varietal selection and Phosphorus (P); in Iloilo, N was followed by Potassium (K) and weed management.
Optimization Thresholds: Yield response curves indicated that yields plateaued at N rates of approximately 105 kg ha⁻¹ (Iloilo) and 140 kg ha⁻¹ (Nueva Ecija) during the dry season. A significant majority of fields (98% in Iloilo, 68% in Nueva Ecija) applied N rates below these thresholds during the dry season.
Scenario Performance:
- Targeted Approach (Scenario 3) proved most efficient. It predicted yield gains of ~1.11 t ha⁻¹ in Iloilo and ~1.96 t ha⁻¹ in Nueva Ecija.
- Crucially, this targeted approach achieved 86% of the potential production increase with only 15% of the increased fertilizer use compared to blanket approaches. In Nueva Ecija, this resulted in a net reduction in N use (~14,800 tons) while increasing production by ~108,000–132,100 tons.
- Blanket approaches (Scenarios 1 and 2) required significantly higher N investments for comparable or lower marginal gains.
Validation: On-farm trials confirmed the model's efficacy. Data-driven recommendations increased grain yields by 15.1% (0.69 t ha⁻¹) in Iloilo and 11.7% (0.76 t ha⁻¹) in Nueva Ecija compared to farmer practices.

Significance and Claims
The paper claims to provide the first integrated application of interpretable machine learning, spatially targeted nitrogen optimization, and on-farm validation specifically for Philippine rice systems. Its primary significance lies in demonstrating that spatial intelligence can transform agricultural policy from inefficient, uniform input subsidies to precision interventions.

The authors argue that shifting to targeted N management allows for:

Resource Efficiency: Maximizing yield gains while minimizing fertilizer inputs and costs, a critical factor given rising global fertilizer prices.
Policy Optimization: Providing a scientific basis for redesigning subsidy programs (like RCEP) to focus resources on "high-impact zones" where returns on investment are highest, rather than distributing them uniformly.
Sustainable Intensification: Offering a scalable pathway to close yield gaps and reduce import dependence without expanding land use or exacerbating environmental risks associated with over-fertilization.

The study concludes that while machine learning models have inherent limitations in capturing all socio-economic and environmental variability, the integration of these tools with field validation offers a robust framework for designing actionable, locally adapted strategies to sustainably bridge rice yield gaps in the Philippines.

More like this