Introduction
Global challenges threaten food security. In Iran, rice is a staple, with 770,000 ha under cultivation and an annual production of 3.75 million tonnes. The northen provinces of Guilan and Mazandaran dominant rice production, with Guilan only supplying about 50% of the nation’s demand. Advanced computational techniques such as metaheuristic algorithms and artificial intelligence offer powerful tools for problem-solving and modeling inspired by the adaptability of living organisms', and are increasingly applied to support agricultural management and sustainability. However, multi-gene genetic programming (MGP) has not yet been used for rice yield modeling in Iran. This study addresses that gap by evaluating the effectiveness of MGP and exploring its potential to enhance agricultural decision-making.
Materials and Methods
This research examines the yield performance of local rice cultivars, namely Hashemi and Ali Kazemi alongside high-yielding cultivars, including Fajr and Shiroudi in Rasht County, Iran. It utilizes library documents, face-to-face interviews, and MGP analysis. Data were collected from 385 randomly selected farmers and landowners in the region, during the 2020-2024 rice production years. Inputs from the energy sector, including fuel and electricity, water pumping equipment, agricultural machinery, fertilizers, pesticides, organic materials, and the energy output of paddy production, were examined. Energy equivalents were used to convert various types of energy into a common unit. To predict the yield of the two types of paddy cultivars using MGP, the structure of MGP trees was first designed with two objectives: maintaining the model accuracy and avoiding structural complexity. Parameter setting include population, generation and tournament sizes, gene limits, tree depth and size, and probabilities for elitism, crossover, and mutation. Additionally, various mathematical functions were utilized.
Results and Discussion
In examining total energy consumption and production in local and high-yielding cultivar farms, the results indicated a significant difference between the two varieties. In farms producing local varieties, the average energy consumption was 41,081.4 MJ, while the average energy production in these farms reached 23,771.9 MJ. In contrast, the average energy consumption in high-yielding cultivar farms was estimated at 41,118.8 MJ, whereas the average energy production in these farms was 42,220.04 MJ. The energy ratio, energy productivity, and specific energy indices for high-yielding varieties were 76.47%, 76.92%, and 77.70% higher, respectively, compared to local varieties, with the net energy gain index showing an improvement of more than 15 times. The higher energy ratio and energy productivity, along with lower specific energy and net energy gain, indicate that the high-yielding cultivar is more energy-efficient in terms of energy consumption. The MGP model converged after 100 iterations, providing the optimal solution. Changes in the best and mean fitness values indicated that as the iterations increased, the error gradually decreased and eventually stabilized, reflecting continuous improvement of the model during the training process and parameter tuning. Through cross-validation with varying training data set sizes, the findings revealed that the MGP model, when utilizing 65% of the total dataset, generated results that were remarkably similar to those achieved with 80% of the data. Hence, 65% is established as the optimal proportion for the training dataset. The coefficient of determination (R²) for the regression line in the training data set was higher than in the test phase for both varieties. In evaluating the MGP equations to assess the accuracy of the proposed model, the tree depth was increased from 4 to 12. For the local cultivar, the highest coefficient of determination (R²) at a tree depth of 4 was 0.95, while for the high-yielding cultivar, it was 0.94. The simpler structure at depth 4 resulted in a simpler mathematical equation. Finally, the effects of independent variables on the dependent variable, paddy yield, were examined. It was found that organic materials, such as compost, seed, rice straw and husks, were the most significant factors influencing the estimation of paddy yield in both varieties.
Conclusion
Since modeling focuses on predicting crop yield and making data-driven scientific and practical decisions, the results of this study represent an important step toward advancing sustainable agriculture. It is recommended that farmers seek up to date insights from consultants and participate in workshops to increase their sustainable yields.