Multimodal Machine Learning for Maize Disease Detection: A Systematic Review of Architectures and Deployment Challenges

Abstract

Maize diseases continue to threaten agricultural productivity and food security, particularly in developing regions where early diagnosis remains constrained by limited expert access. While deep learning has enabled automated disease detection systems, most existing approaches rely on unimodal image datasets and cloud-dependent architectures, limiting robustness and deployment feasibility in resource-constrained environments. This study presents a structured systematic review of 38 peer-reviewed studies published between 2020 and 2025, focusing on multimodal machine learning approaches integrating visual and environmental data for maize disease detection. Quantitative synthesis reveals that 58% of reviewed studies employ image-only deep learning models, 26% implement multimodal frameworks, and only 29% conduct validation under real or semi-real field conditions. Furthermore, 32% explicitly address deployment considerations, including edge computing and mobile inference. The findings demonstrate that multimodal architectures improve robustness and contextual modeling compared to unimodal systems by integrating phenotypic and environmental drivers of disease expression. However, increased computational complexity, synchronization challenges, and limited edge optimization remain barriers to scalable deployment. This review advances scientific knowledge by providing a computing-centered synthesis of multimodal architectures, fusion strategies, deployment constraints, and explainability gaps, identifying key research priorities in edge efficiency, real-world validation, and interpretable intelligent systems.