Work

An interpretable deep embedding framework for data-driven discovery of crop families

Year: 2026

Type: article

Abstract: Understanding functional similarities among crops is necessary for large-scale crop zoning, resource planning, and climate-adaptive agriculture. However, existing classifications depend primarily on botanical taxonomy and fail to reflect the multi-dimensional behavior of crops in complex real-world cultivation environments. To bridge this gap, we propose CROP-LENS, an interpretable data-driven framework that discovers ecological crop families by combining deep embedding, multi-algorithm clustering, and surrogate decision tree explanation. Unlike traditional joint-optimization deep clustering methods that suffer from feature distortion and high computational overhead for tabular data, CROP-LENS utilizes a decoupled architecture to project high-dimensional agricultural data into a compact latent space, apply a topology-preserving projection, and dynamically select the optimal clustering geometry. To ensure operational scalability, a core-set sampling mechanism is implemented by clustering a representative subset of 1,500 data points and assigning the remainder via k-NN. The proposed framework executes in roughly 65 seconds on standard CPU hardware, representing a speedup over foundational baselines. The proposed framework achieved dominant structural separation (Silhouette = 0.870) while maintaining highly stable agronomic variance (ENV-CV = 0.512) when evaluated across real and synthetic datasets. Moreover, an interpretability layer utilizing medoid representatives, statistical profiles, and a global surrogate decision tree (Hold-out Accuracy = 0.95) translates these latent boundaries into actionable human-readable agronomic rules validated against established agricultural standards. Eventually, this framework provides a transparent highly efficient alternative to botanical taxonomy practical in supporting data-driven decision-making in precision agriculture and climate adaptation planning.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.102571

Authors: Ibrahim Nasir Mahmood, Bencsik Gergely, Mustafa Ali Abuzaraida, Zita Szepreti

Institutions: Eötvös Loránd University, Libyan Authority for Scientific Research, Magyar Agrár- és Élettudományi Egyetem

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $1,770