European Journal of Sustainable Agroecosystems

Context: As climate fluctuations, population surges, and political tensions continue to increase, worldwide food security is more vulnerable than ever. To stabilize the flows of the goods through the supply chains, formulate adequate macroeconomic policy, and apply precision farming, one has to be able to predict crop yields correctly. Within the last ten years, ANN models evolved from experimental mathematical curiosities to being essential parts of modern predictive agronomy.

Purpose of the Review: This review paper performs an in-depth critical review and evidence synthesis of the publications on the use of ANN architectures for the yield forecasting from 2015 to 2025. The review shows how ANN technology evolved over the years from shallow networks to deep, multi-modal and physics-based learning methods.

Literature Review: Inestigators executed a systematic literature search in leading global databases, such as Scopus, Web of Science, IEEE Xplore, ScienceDirect, and SpringerLink, and delivered a collection of high-impact peer-reviewed papers.

Methodology of the review: Based on the PRISMA approach, the papers which were chosen, underwent categorical assessment according to the data they used, their structure, optimization methods, and area of application.

Results of the study: The results show that traditional Multilayer Perceptrons (MLPs) are widely used because they are easy to compute, however, they don’t provide accurate results compared to deep Long Short-Term Memory (LSTM) networks for multi-temporal agrometeorological data, and Convolutional Neural Networks (CNNs) for high-dimensional remote sensing images. The new pieces of evidence show that hybrid architectures (the one combining CNN and LSTM methods) and Transformer-based self-attention mechanisms are the best means for capturing nonlinear spatial-temporal interactions.To ensure practical viability in agricultural decision-making, future studies should move away from experimental-modeling. A focus on Explainable AI (XAI) and design of Physics-Informed Neural Networks (PINNs) is required in order to successfully combine mechanistic crop growth models with deep learning.