Artughrul Gayibov
A unified ranking-based evaluation framework for diverse crop recommendation techniques on the CropIntel Dataset
The rising demand for sustainable agriculture calls for intelligent systems that recommend suitable crops. This study turns site-specific soil, climate, and terrain data into a ranked short list of crops for decision support. The problem is that crop choice is framed as a yes/no suitability task and evaluated with mixed criteria, hindering comparison and adoption. We propose a unified, ranking-based evaluation and benchmark three approaches: tree-based learners, a similarity method that matches new environments to known ones, and a simple clustering baseline using standard top-k metrics (precision, recall, and mean reciprocal rank). Results show that ensemble trees provide the most reliable overall rankings, while the similarity method yields strong early-rank retrieval; feasibility rules based on agro-ecological constraints keep recommendations realistic without lowering quality. These outcomes arise from non-linear patterns captured by ensembles and closely related environments that favor similarity matching. Features include a common top-k protocol, preprocessing, and transparent guardrails. In practice, the framework supports advisory systems that produce short lists for regions with measured profiles; new or shifting regions require geography-aware validation and local calibration in real deployments.
Keywords: Crop recommendation, Precision agriculture, Machine learning, Recommender systems, Land suitability, Agro-ecological constraints, Decision support
DOI: https://doi.org/10.54381/icp.2025.2.06
A unified ranking-based evaluation framework for diverse crop recommendation techniques on the CropIntel Dataset
The rising demand for sustainable agriculture calls for intelligent systems that recommend suitable crops. This study turns site-specific soil, climate, and terrain data into a ranked short list of crops for decision support. The problem is that crop choice is framed as a yes/no suitability task and evaluated with mixed criteria, hindering comparison and adoption. We propose a unified, ranking-based evaluation and benchmark three approaches: tree-based learners, a similarity method that matches new environments to known ones, and a simple clustering baseline using standard top-k metrics (precision, recall, and mean reciprocal rank). Results show that ensemble trees provide the most reliable overall rankings, while the similarity method yields strong early-rank retrieval; feasibility rules based on agro-ecological constraints keep recommendations realistic without lowering quality. These outcomes arise from non-linear patterns captured by ensembles and closely related environments that favor similarity matching. Features include a common top-k protocol, preprocessing, and transparent guardrails. In practice, the framework supports advisory systems that produce short lists for regions with measured profiles; new or shifting regions require geography-aware validation and local calibration in real deployments.
Keywords: Crop recommendation, Precision agriculture, Machine learning, Recommender systems, Land suitability, Agro-ecological constraints, Decision support
DOI: https://doi.org/10.54381/icp.2025.2.06