Abstract
Against the backdrop of a severe global food security situation, accurate prediction of crop yields is crucial for ensuring stable food supply and maintaining socio-economic stability. Heilongjiang Province is an important region for rice production in China, and its rice yield is of great significance to the stable supply of national grain. This study uses multiple linear regression and random forest regression models to conduct in-depth research and predictive analysis on rice production in various cities of Heilongjiang Province. The data is sourced from authoritative databases and field research, after data cleaning, which includes removing outliers, handling missing values, and correcting errors, and standardization, which normalizes data to a common scale (such as using z-score normalization or min-max scaling) to ensure consistency and comparability, models are constructed. The multiple linear regression model is based on the assumption of linear relationship, estimates parameters using the least squares method, and rigorously tests and evaluates them; The random forest regression model, as an optimization model, utilizes multiple decision trees for ensemble learning to improve prediction accuracy and reliability. Comparing the predicted values of the two models, it was found that they each have their own advantages, providing diverse and powerful support for predicting rice yield in Heilongjiang. This not only contributes to the scientific planning and sustainable development of the rice industry in Heilongjiang Province, but also has profound significance for ensuring national food security and promoting the modernization of agriculture.
Introduction
Heilongjiang Province is vital to China’s national food security, yet predicting its rice yield is notoriously difficult due to the complex interplay of environmental, economic, and policy factors. While previous research has applied linear modeling to agriculture, these studies often lack regional specificity or fail to account for the non-linear complexities introduced by climate change and globalized markets.
The authors seek to bridge this gap by creating a more comprehensive predictive framework. They aim to provide agricultural departments and local farmers with a reliable way to forecast yields based on inputs like planting area, fertilizer usage, and irrigation, ultimately helping to stabilize the food supply and optimize farming strategies.
Methodology
The study employs a two-pronged modeling approach to analyze the rice industry. First, Multiple Linear Regression (MLR) is used as the primary analytical model to establish a baseline. It assumes a linear relationship between rice yield ($Y$) and ten independent variables, estimating parameters via the Ordinary Least Squares (OLS) method. Second, Random Forest (RF) Regression is introduced as an optimization model. As an ensemble learning technique, it constructs multiple decision trees and averages their results to improve prediction reliability and handle potential non-linearities in the agricultural data.
Notably, the researchers implemented their regression coefficient solvers using the C programming language, prioritizing computational efficiency. The dataset spans 2015 to 2023, integrating macro data from the Heilongjiang Rural Statistical Yearbook and micro data from field research in hubs like Yichun and Wuchang. To ensure consistency, the team utilized Z-score standardization to normalize dimensions across variables such as fertilizer usage, irrigation volume, and labor input.
Results and Discussion
A critical finding in the MLR analysis was that only the constant term (intercept) showed high statistical significance ($t = 10.0273$), while other individual coefficients had $t$-values near zero. This suggests that in complex agricultural systems, individual factors rarely act in a simple, independent linear fashion.
The "aha moment" comes from the error analysis, where the Random Forest model consistently corrected the "overshoot" or "undershoot" of the linear model. By utilizing multiple decision trees, the RF model better captured the nuanced, multi-factor characteristics of the Heilongjiang rice industry. As shown in the comparison of predicted versus actual data, the RF model tracks the actual yield fluctuations much more closely than the MLR baseline.
The Random Forest model proved superior across all evaluated error metrics:
- Mean Square Error (MSE): RF (110.20) vs. MLR (255.06)
- Mean Absolute Error (MAE): RF (8.17) vs. MLR (11.82)
- Root Mean Square Error (RMSE): RF (10.50) vs. MLR (15.97)
- Mean Relative Error (MRE): RF (0.0193) vs. MLR (0.0272)
Beyond mere prediction, the paper links these models to "Benefit Evaluation." By plugging predicted yields into a secondary income model, farmers can simulate the financial impact of different seed varieties or fertilizer adjustments before the season begins.
Conclusion
While the current models are effective, future research should expand the time span of the data and include a wider array of influencing factors, such as granular soil health data or specific policy shift indicators. Moving toward more flexible data science environments might also allow for easier integration of real-time satellite or IoT sensor data from the fields. Ultimately, the transition from simple linear models to ensemble methods like Random Forest provides a more robust tool for regional food security planning and farmer decision-making in Heilongjiang.
Summary updated