Multimodal Data Fusion in Smart Agriculture: Integrating Soil, Climate, Imaging, and Biological Signals Using Machine Learning: Scoping Review

The evolution of smart agriculture has expedited the adoption of machine learning-based solutions for crop monitoring, disease detection, yield estimation, and environmental management. However, most of the prevailing machine learning-based solutions utilize unimodal data, such as image, soil, or climatic-based information. These solutions face the limitation of effectively handling the complex interactions between the agricultural ecosystems. Multimodal data fusion-based solutions have been recognized as a powerful tool for effectively utilizing the strength of heterogeneous data sources, including soil-based information, climatic-based information, image-based information, and plant physiological-based information. This systematic review focuses on the machine learning-based multimodal data fusion solutions developed using research published between 2020 and 2026. A structured review-based methodology was followed to identify the relevant research. A review of the relevant literature was conducted to determine the effectiveness of the machine learning-based multimodal data fusion solutions. It was found that the effectiveness of the machine learning-based solutions, including early, intermediate, and late fusion-based solutions, attention-based solutions, transformer-based solutions, graph learning-based solutions, and deep ensemble-based solutions, was better compared to unimodal-based solutions. The applications of the machine learning-based multimodal data fusion solutions were found to be effective for applications such as crop disease detection, soil moisture estimation, yield estimation, plant stress detection, and intelligent decision-making. The challenges associated with the machine learning-based multimodal data fusion solutions were identified to be associated with issues such as heterogeneity, synchronization, the unavailability of standardized multimodal datasets, and the limitations associated with real-time applications. The research direction of the machine learning-based multimodal data fusion solutions was identified to be associated with cross-modal transformer-based solutions, explainable machine learning-based solutions, federated learning-based solutions, the utilization of biological-based information with remote sensing-based information, and the utilization of graph learning-based solutions.