Paper Title
Crop Yield Prediction Using Multilayer Perceptron Neural Networks: A Comparative Analysis with Traditional Machine Learning Approaches

Abstract
The world faces numerous challenges with regard to ensuring food security, which can be exacerbated by un-predictable climatic conditions and limited resources, thus necessitating more intelligent and sustainable ways of managing agricultural systems. Precision Agriculture (PA) utilizes technology to manage the input of agricultural crops so as to achieve higher crop yields. In PA, Artificial Intelligence (AI) and Machine Learning (ML) are key technologies in support of using AI and ML to make decisions based on large amounts of data from the field. The purpose of this research project was to provide a complete evaluation of the application of a Multilayer Perceptron (MLP) type neural network to predict crop yields, utilizing multi-modal tabular agricultural data representing weather conditions, nutrient content of soils and how farm managers have managed their farms. We designed, implemented and evaluated an MLP model, trained on the multi-modal tabular data mentioned above. Our approach included the application of a systematic grid-search with 5 fold-cross-validation for optimizing the hyper-parameters of the MLP, domain-specific feature engineering for improving the quality of the data, and domain-specific regularization techniques for reducing overfitting and improving generalizability. We used multiple regression metrics, i.e., R-squared (R 2 ), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), to compare the performance of the MLP with five other regression models, i.e., Multiple Linear Regression (MLR), Decision Tree Regressor (DTR), Support Vector Regression (SVR), and Random Forest Regressor (RFR). The experimental results showed that the optimized MLP achieved competitive performance relative to the other regression models; the performance metrics were R 2 = 0.89, MAE = 245.32 kg/ha, and RMSE = 312.45 kg/ha. The results also showed that the MLP performed better than the linear regression and decision tree regression models but similarly to the random forest regression model. A thorough ablation study provided further evidence to validate the effectiveness of each of the architectural choices we made in the MLP model. Finally, the permutation-based feature importance analysis validated the alignment of the features selected by the MLP model with those recommended by established agronomic principles. This research study provides a new and important contribution to the current state-of-the art research literature because it represents the first study to provide a direct head-to-head comparison of foundational neural network architectures (i.e., MLP) and ensemble methods (i.e., RF) to predict crop yields from tabular agricultural data. As such, the results of this study demonstrate that the MLP is a viable, scalable and accessible decision-support tool for precision agriculture applications. Keywords - Crop Yield Prediction, Multilayer Perceptron, Neural Networks, Precision Agriculture, Machine Learning, Deep Learning, Random Forest, Agricultural Data Analytics, Feature Engineering, Hyperparameter Optimization