Abstract. Rice is the most important staple food in Asia. However, high-spatiotemporal-resolution rice yield datasets are limited over this large region. Based on annual rice maps in Asia, we incorporated multisource predictors into three machine learning (ML) models to generate a high-spatial-resolution (4 km) seasonal rice yield dataset (AsiaRiceYield4km) for the 1995–2015 period. The results showed that AsiaRiceYield4km achieves good accuracy for seasonal rice yield estimation (single rice: $R^2=0.88$; double rice: $R^2=0.91$; and triple rice: $R^2=0.93$). AsiaRiceYield4km is a novel long-term gridded rice yield dataset that can fill the unavailability of high-spatial-resolution seasonal yield products across major rice production areas.
Introduction
Rice provides over a quarter of the calories for half the global population, making it the lifeblood of Asia. Assessing the impacts of climate change on this staple crop has been hampered by a significant "data gap." Existing global yield datasets are often too coarse (10–55 km resolution), provide only annual snapshots that ignore multiple cropping seasons (double or triple rice), or rely on static maps that fail to account for year-to-year changes in rice cultivation areas. To build resilient food systems, researchers and policymakers require granular data on when and where rice is grown at a scale reflecting smallholder farming realities.
Methodological Framework
The AsiaRiceYield4km dataset was generated using a three-step machine learning pipeline designed to capture the complexity of Asian rice production.
- Phenology Detection: Using Leaf Area Index (LAI) data, the researchers identified key growth stages—transplanting, heading, and maturity—for every pixel. This allowed the growing season to be divided into vegetative and reproductive periods.
- Predictor Selection: 50 variables were curated across four categories: Cumulative Growing Predictors (CGPs), Extreme Growing Predictors (EGPs), Constant Environmental Conditions (CECs) such as soil properties and elevation, and Temporal Information (TI) reflecting technological trends.
- Model Optimization: Three ML architectures—Random Forest (RF), XGBoost, and Long Short-Term Memory (LSTM)—were tested. An Inverse Probability Weighting (IPW) method was used to select the optimal model for each of the 27 country-specific cases, balancing performance across training, validation, and testing sets.
Data and Key Insights
The study integrated a massive array of multisource data, including over 45,000 yield records from 1,400 administrative units, annual 500 m resolution rice maps (APRA500), and remote sensing data (GLASS AVHRR LAI).
A key finding was that Constant Environmental Conditions (CECs)—including soil pH, clay content, and geographic location—contributed the most (~45%) to yield prediction. This suggests that while weather fluctuates, inherent land quality remains the strongest anchor for yield potential. Furthermore, predictors during the reproductive period (heading to maturity) had a significantly higher impact on final yield than those during the vegetative period. In terms of modeling, XGBoost and Random Forest consistently outperformed the deep-learning LSTM model for this gridded spatial task.
Performance and Results
AsiaRiceYield4km represents a major upgrade over existing benchmarks like the Spatial Production Allocation Model (SPAM). The dataset achieved high accuracy across different cropping systems, with $R^2$ values ranging from 0.88 to 0.93.
Compared to SPAM, AsiaRiceYield4km improved the $R^2$ by 0.20 and reduced the Root Mean Square Error (RMSE) by an average of 618 kg/ha. The dataset also revealed regional trends: while most of Asia saw yield increases from 1995–2015 (e.g., Cambodia at 2.65% annual growth), Japan experienced a slight decline (-0.01%). Crucially, the high-resolution mapping successfully captured yield "hotspots" in the Indo-Gangetic Plain and East China that coarser models often smooth over.
Limitations and Future Directions
Despite its advancements, the dataset has limitations. Mixed-pixel effects in fragmented smallholder landscapes can introduce noise, and cloud interference during monsoon seasons often obscures optical satellite data. Additionally, as a machine learning approach, the models lack the biological traceability of process-based crop models. Future work may involve "knowledge-guided" machine learning to incorporate physiological constraints and utilizing higher-resolution input data to further improve spatial precision.
AsiaRiceYield4km is available at: https://doi.org/10.5281/zenodo.6901968
Summary updated