A Comprehensive Review Of AI-Driven Crop Recommendation Systems Using Deep Learning: Architectures, Datasets, And Future Research Directions

Abstract

Given how quickly AI as well as deep learning methods are developing, the modern-day agricultural sector has undergone many changes through the introduction of intelligence into the decision-making process. In particular, Crop Recommendation Systems (CRS) utilizing AI algorithms are among the useful solutions that help to boost productivity in the agricultural field. This review explores a number of studies dedicated to deep learning algorithms used to create crop recommendation systems, considering such aspects as architecture, dataset, methodological approaches, and future research areas. The research is conducted according to the PRISMA methodology, exploring relevant studies published between 2016 and 2026. CNN, LSTM, BiLSTM, GRU, Transformers, GNN, as well as hybrids like CNN-LSTM as well as CNN-Transformer are examples of deep learning models. Additionally, the study reviews existing agricultural datasets collected on the basis of soil parameters, climatic data, remote sensing, IoT, and satellite data. The study further considers some important problems that arise during implementation, including data sparsity, imbalance, lack of interpretability, computation limitations and cross-region generalizability problems. In addition to that, the study reviews some of the existing solutions to those problems, which include transfer learning, federated learning, explainable AI, TinyML and attention-based methods. It has been found out through the study that Transformer-based architectures and hybrid models perform much better than conventional ML models because of their ability to capture complicated space-time dependencies among agriculture-related variables. Lastly, the study identifies some potential areas where future research can take place in the domain of crop recommendations.