Agriculture is a backbone for national stability, food security, and economic diversification. It fosters rural development, employment, and international trade. Sustainable practices benefit both the environment and national identity. Punjab relies heavily on wheat production but faces market uncertainties due to various factors. Traditional forecasting methods fall short.
Our innovative Economic Forecasting Model for Agriculture focuses on predicting wheat prices in Punjab, leveraging a comprehensive dataset. To handle its complexity, we use the Random Forest Regressor and Genetic Algorithm Optimization, superior to traditional approaches. These AI techniques empower stakeholders with precise predictions, enhancing decision-making for planting, harvesting, and market engagement in agricultural economics.
This dataset, rich with 25 independent variables and the crucial dependent variable Harvesting Price, provides a comprehensive understanding of the multi-dimentional factors influencing wheat pricing in Punjab. It captures the temporal dimension through the Year variable, shedding light on market conditions, climate, and external factors shaping pricing. Furthermore, it offers invaluable geographical insights via variables like Division, District, and Tehsil, allowing us to explore soil types, microclimates, and localized market dynamic's impact on wheat pricing.
Several critical variables, including Harvest Date, Wheat Variety, and Seed Type, play pivotal roles in determining crop quality, quantity, and pricing dynamics. Seed-related factors such as Seed Quantity (kg) and Sowing Date are instrumental in forecasting yield potential, while the Sowing Method significantly influences crop development. Fertilization variables like Organic Manure (Cow Dung) (kg), Urea Fertilizer (kg), and DAP Fertilizer (kg) directly affect crop vitality and economic outcomes. Soil type and irrigation method are crucial in defining the growing medium and hydraulic conditions, impacting crop development and financial results. No. of Watering Times and Residual Weed Material (kg) emphasize the interplay between moisture and weed control, while Previous Crop carries the legacy of past crops, impacting nutrient levels and diseases. Additionally, Seed Treatment, Number of Pest Sprays, and Number of Weed Control Sprays safeguard crops against pests and diseases. Finally, the dataset quantifies the crop's value through Yield (40 Kg/Acre) and Residual Material after Harvest (40 Kg/Acre), underscoring the influence of watering on agricultural outcomes.
With a substantial sample size of 44,424, the dataset is thoughtfully partitioned into a training set (80%) for model development and a testing set (20%) for validation, ensuring its robustness and applicability. This dataset offers a unique opportunity to gain comprehensive insights into the intricate dynamics of wheat pricing in Punjab, making it an invaluable resource for researchers, analysts, and stakeholders in the agricultural sector.
| Variables | Minimum | Median | Maximum |
|---|---|---|---|
| Harvest Date | 0 | 2 | 22 |
| Wheat Variety | 1 | 6 | 13 |
| Seed Quantity (kg) | 0 | 3 | 60 |
| Sowing Date | 0 | 3 | 6 |
| Organic Manure (kg) | 0 | 1 | 4 |
| Urea Fertilizer (kg) | 0 | 100 | 775 |
| DAP Fertilizer (kg) | 0 | 50 | 200 |
| Soil Type | 0 | 1 | 3 |
| No. of Pest Sprays | 0 | 0 | 3 |
| No. of Weed Control Sprays | 0 | 1 | 13 |
| Yield (40 Kg/ Acre) | 0.01452 | 35.211 | 82.80938 |
| Residual Material (40 Kg/ Acre) | 0.680625 | 102.0938 | 1592.051 |
| No. of Watering Episodes | 0 | 3 | 10 |
| No. of Watering Episodes | 0.12 | 1400 | 2600 |
Table 2: Comparative Mean Analysis for Variables across the Original Dataset, Training Set, and Testing Set
| Variables | Original Dataset | Training Set | Testing Set |
|---|---|---|---|
| Harvest Date | 2.229835 | 2.230554 | 2.226958 |
| Wheat Variety | 7.052531 | 7.042987 | 7.090703 |
| Seed Quantity (kg) | 3.959883 | 3.999867 | 3.799947 |
| Sowing Date | 2.618514 | 2.618773 | 2.617475 |
| Organic Manure (kg) | 0.997682 | 0.995438 | 1.00666 |
| Urea Fertilizer (kg) | 80.1918 | 80.13702 | 80.4109 |
| DAP Fertilizer (kg) | 49.41929 | 49.53133 | 48.9711 |
| Soil Type | 1.534283 | 1.534263 | 1.534363 |
| No. of Pest Sprays | 0.273601 | 0.27544 | 0.266249 |
| No. of Weed Control Sprays | 0.939078 | 0.940064 | 0.935136 |
| Yield (40 Kg/ Acre) | 34.49118 | 34.49175 | 34.48893 |
| Residual Material (40 Kg/ Acre) | 256.3489 | 255.7909 | 258.5809 |
| No. of Watering Episodes | 2.866116 | 2.868507 | 2.856553 |
Table 3: Comparative Standard Deviation Analysis for Variables across the Original Dataset, Training Set, and Testing Set
| Variables | Original Dataset | Training Set | Testing Set |
|---|---|---|---|
| Harvest Date | 0.805476 | 0.805637 | 0.804879 |
| Wheat Variety | 3.523892 | 3.517537 | 3.549179 |
| Seed Quantity (kg) | 5.811012 | 5.952807 | 5.202641 |
| Sowing Date | 0.742791 | 0.741065 | 0.749707 |
| Organic Manure (kg) | 0.454628 | 0.452991 | 0.461039 |
| Urea Fertilizer (kg) | 32.3681 | 32.40014 | 32.24083 |
| DAP Fertilizer (kg) | 18.03213 | 18.03141 | 18.02924 |
| Soil Type | 0.706228 | 0.705203 | 0.710362 |
| No. of Pest Sprays | 0.480772 | 0.482229 | 0.47486 |
| No. of Weed Control Sprays | 0.400447 | 0.396117 | 0.417324 |
| Yield (40 Kg/ Acre) | 9.810252 | 9.804385 | 9.834339 |
| Residual Material (40 Kg/ Acre) | 386.834 | 385.6889 | 391.3989 |
| No. of Watering Episodes | 1.793422 | 1.790627 | 1.80465 |
Optimal parameter selection for the Random Forest Regressor (RFR) is of paramount importance to minimize the mean squared error. Attaining a proficient combination of parameters significantly contributes to the reduction of mean squared error, as well as enhances the performance across additional evaluation metrics such as root mean square error, mean absolute error, and Durbin-Watson score. The GA-RFR model amalgamates the merits of both the Random Forest Regressor and Genetic Algorithm (GA) methodologies. The procedural framework for implementing the GA-RFR model is delineated as follows:
Step #1: Create an initial population of binary-coded chromosomes with a predefined quantity. Each chromosome encodes the presence (1) or absence (0) of specific features.
Step #2: Assess the fitness of each individual within the population based on the mean squared error resulting from the training of a Random Forest Regression (RFR) model. Individuals with higher fitness values are closer to the desired optimal solution.
Step #3: Employ Genetic Algorithm (GA) operators. Initially, perform selection to choose pairs of chromosomes for reproduction based on their fitness. Construct a parent population from the selected chromosomes. Subsequently, apply crossover and mutation operations using predetermined GA parameters to generate new candidate individuals with altered characteristics.
Step #4: Repeat the GA operations described in Step 3 until a predetermined maximum number of iterations is reached.
Step #5: Once the maximum iteration limit is met or a stopping condition is satisfied, conclude the process. Output the optimal combination of RFR parameters, including n_estimators, min_samples_leaf, min_samples_split, and max_depth, obtained through the evolutionary process of the hybrid model.
This approach employs a Genetic Algorithm to iteratively evolve a population of binary-encoded chromosomes, optimizing the selection of RFR parameters for enhanced model performance. The process involves fitness evaluation, selection, crossover, and mutation operations, leading to the identification of an optimal parameter configuration.
In this research, a set of four well-established metrics was specifically selected to assess the effectiveness of the regression models considered. These metrics include Mean Squared Error (MSE), R-squared (Coefficient of Determination), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). Additionally, the Durbin-Watson score was included to comprehensively evaluate the model's performance.
1: Mean Squared Error (MSE): MSE measures the average squared difference between predicted values and actual values, providing a quantified assessment of prediction accuracy.
2: Mean Absolute Error (MAE): MAE calculates the average absolute differences between predicted and actual values. It assigns equal importance to all errors and is less affected by outliers than MSE.
3: Root Mean Squared Error (RMSE): RMSE, the square root of MSE, measures the average magnitude of errors in predicted values. Lower values of MSE, MAE, and RMSE indicate more accurate predictions, with RMSE being particularly valuable as it shares the unit of the target variable.
4: R-squared Score: R-squared quantifies the proportion of variance in the dependent variable predictable from the independent variables. A higher value (ranging from 0 to 1) indicates a better model fit.
5: Durbin-Watson Score: This score assesses autocorrelation in regression model residuals. Values near 2 signify no significant autocorrelation.
In the study, the success of the genetic algorithm (GA) optimization hinges on parameter selection. The 'varbound' array sets crucial boundaries for the four variables, guiding the search space and ensuring valid solutions. Within the GA, the 'algorithm_param' dictionary defines key parameters such as 'max_num_iteration' (20 iterations) and 'population_size' (explored across a range of values). This analysis, presented in Table 4, highlights the interplay of population size with the GA. Mutation and crossover parameters shape genetic recombination, while 'parents_portion' and 'elit_ratio' maintain solution diversity. 'Max_iteration_without_improv' combats stagnation, limiting non-improving generations. The optimal results at a population size of 5 underscore the effectiveness of parameter settings. These values represent the GA's ability to converge toward solutions aligning with optimization goals.
The results illustrate the relationship between hyperparameters and prediction accuracy, guiding parameter configurations for agricultural economic forecasting. The GA's role in balancing complexity, preventing overfitting, and navigating trade-offs enhances the Random Forest Regressor's predictive power. This optimization framework showcases the synergy of computational optimization and machine learning, providing a robust model refinement approach in agricultural economics.
Table 4: Comparative Analysis for Evaluating Parameters ( R-2, MAE, MSE ,RMSE and Durbin Watson Score) across the population ranging from 1 to 5
| Population | R-2 | MAE | MSE | RMSE |
|---|---|---|---|---|
| 1 | 0.996 | 17.227 | 633.667 | 25.173 |
| 2 | 0.996 | 17.918 | 647.561 | 24.101 |
| 3 | 0.996 | 17.928 | 675.515 | 25.991 |
| 4 | 0.996 | 16.427 | 578.278 | 24.047 |
| 5 | 0.996 | 15.815 | 546.686 | 23.381 |
Table 5: The table presents the optimal parameters of the (RFR) model when applied to a population size of 5
| n_estimators | max_depth | min_samples_split | min_samples_leaf | mean_squared_error |
|---|---|---|---|---|
| 178 | 42 | 3 | 1 | 546.686 |
After training the Random Forest Regressor on the dataset, a thorough evaluation reveals the model's exceptional proficiency in capturing intricate data relationships. This is notably evidenced by an impressive R-squared (R2) score of 0.999, explaining about 99.6% of the dataset's variance and underlining the model's accuracy in aligning predictions with actual values. Moreover, the model's precision in prediction accuracy is evident through its Mean Absolute Error (MAE) of 6.335. In economic forecasting, where minor deviations carry significant decision-making consequences, such precision is invaluable. Additionally, the Mean Squared Error (MSE) of 114.963 reflects the model's ability to consistently generate predictions close to actual outcomes, a crucial factor for reliable and accurate economic forecasts. The Root Mean Squared Error (RMSE) of 10.722, measured in the original units of the dependent variable, underscores the model's proficiency in providing dependable forecasts with minimal deviations from actual outcomes, particularly beneficial in the realm of agricultural economics.
| n_estimators | R-2 | MAE | MSE | RMSE |
|---|---|---|---|---|
| 178 | 0.999 | 6.335 | 114.963 | 10.722 |
The evaluation of the Random Forest Regressor model on the testing dataset provides valuable insights into its performance and adaptability beyond the training phase. The model's R-squared (R2) score of 0.996 indicates a consistent alignment between its predictions and the actual values in the testing data. This remarkable consistency showcases the model's ability to maintain predictive accuracy, avoiding overfitting and ensuring effectiveness with unseen data. In terms of Mean Absolute Error (MAE), the score of 15.815 quantifies the average prediction error magnitude in the testing dataset. While it reflects a moderate level of deviation, it's important to consider the inherent variability in agricultural economics predictions. Notably, the slightly higher Mean Squared Error (MSE) in the testing dataset, registering at 546.686, suggests that the model encounters somewhat larger errors with previously unseen data. This underscores the importance of evaluating performance across various datasets for robust generalization. The Root Mean Squared Error (RMSE) of 23.381 provides a clear measure of prediction accuracy in the original units of the dependent variable. Lastly, the Durbin Watson score of 2.021 plays a role in detecting potential autocorrelation in prediction errors, suggesting minimal positive autocorrelation. This hints at potential temporal patterns or dependencies within the data, with values near 2 indicating a lack of significant autocorrelation. These findings collectively illustrate the model's strong predictive capabilities and its ability to handle unforeseen data with precision.
| n_estimators | R-2 | MAE | MSE | RMSE | DW Score |
|---|---|---|---|---|---|
| 178 | 0.996 | 15.815 | 546.686 | 23.381 | 2.021 |