Natural Language Processing-Enabled Crop Growth Estimation and Market-Readiness Forecasting: A PRISMA-Guided Review of Models, Data Fusion Strategies, Validation Practices, and Research Gaps

Abstract

The estimation of crop growth and the determination of when a crop will be available for the market are typically considered to be two distinct categories of prediction tasks. The purpose of this review is to meticulously compile the findings of fifty studies that have been conducted on agricultural natural language processing (NLP), the prediction of crop growth and yield, the prediction of commodity prices, and the identification of crops that are mature or ready to be harvested. The end evidence base consisted of ten studies on natural language processing and knowing intelligence, twenty research on growth and yield, ten studies on price forecasting, and ten studies on maturity and readiness. Traditional ensembles such as Random Forest and XGBoost maintained their competitiveness when applied to structured regional datasets & samples. The utilization of optimized recurrent networks and transformer versions became increasingly prevalent in the course of price investigations. Object recognition, semantic segmentation, and lightweight vision models were the most commonly utilized in studies pertaining to maturity. The research conducted on natural language processing (NLP) led to improvements in the categorization of agricultural texts, the answering of queries, the extraction of information, the construction of ontologies, and the support of large language models. In contrast, it only occasionally contributed characteristics directly to models that predict the statuses of crops or the timing of market events. The architecture, not the algorithms, is the primary area where there is a lack of evidence: none of the studies that were examined generated a fully tested, natural language processing (NLP)-enabled system that could predict crop growth, harvest readiness, market potential, operational risk, and uncertainty. For the purpose of bridging this gap, a multimodal architecture with five layers is proposed. Among its features are natural language processing that takes into account provenance, calibrated prediction services, decision-level fusion, and feedback from actual farm and market outcomes that are updated over temporal instance sets.