Design of Explainable Crop Recommendation System for Smart Farming using Machine Learning

Abstract

The infusion of high-end technologies such as Machine Learning (ML), Deep Learning (DL), and the Internet of Things (IoT) into agriculture has paved the way for smart farming, a paradigm shift aimed at enhancing optimal use of resources, increasing harvest productivity and supporting eco-friendly practices. Central to this advancement is the creation of advanced crop advisory systems that provide farmers with datadriven guidance for optimal crop selection based on specific environmental conditions. However, existing models often rely on limited datasets and lack interpretability, leading to less precise recommendations and slow adoption among farmers owing to the non-transparent functioning of AI models. Current literature highlights the need to address several research challenges to improve the effectiveness of smart farming solutions, particularly within the realm of crop recommendation systems. While traditional approaches have primarily focused on leveraging existing datasets for model training, these methods often fall short in accurately reflecting the challenges of actual farming environments. A crucial area of exploration is synthetic data generation, with Generative Adversarial Networks being identified as key technologies capable of generating realistic data sets. This advancement is pivotal in augmenting the learning process of ML models in agriculture, allowing for more robust and adaptable crop advisory system. Furthermore, the review emphasizes the importance of feature selection techniques, such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), which are instrumental in identifying the most relevant variables for building accurate and interpretable predictive models. These advanced techniques address the challenge of extraneous features, which can confuse the models and impede predictive accuracy.