Optimizing Crop Selection Through Hyperparameter Tuning of Neural Networks: Development and Evaluation of Deep Learning Models for Enhancing Crop Recommendation Systems

Applications of Precision Agriculture are becoming more popular throughout the world due to the possibilities that this technology offers to enhance crop yields as a result of decision-making based on effective data. By leveraging data such as pH, temperature, and rainfall, the models provide precise predictions, enabling farmers to make informed crop selection decisions. In this study, we propose and test two neural network models for crop recommendation based on the soil and or environmental data. The first architecture has 5 parameters, in fact, it consists of two hidden layers; the second one has 300 parameters it consists of five hidden layers so we can observe the difference between the model complexity and performance. Thus, we employed grid search and 5-fold cross-validation to optimize the use of hyperparameters such as learning rate, batch size, and activation function. Using the two-hidden-layer model, we obtained an accuracy of 98.7% percent, thereby closely performing better compared to the five-hidden-layer model that achieved 98.4% accuracy. Analyzing experimental results confirmed that larger learning rate and the Adam optimizer provided the highest average scores in both networks architecture. For example, the architecture with five hidden layers was almost five times slower, pointing to the time-space trade-off in neural network models. The results are useful in the current endeavor of deploying AI for agriculture and recommending crops by implementing this file-based approach. Further studies on the same work will involve accumulating a larger database with the soil characteristics of the Bangladesh region, with the hope of establishing a more refined and specific recommendation for the right crops to be grown.