An ensemble framework for farmland quality evaluation based on machine learning and physical models.

Farmland quality (FQ) evaluation is crucial to curb agricultural land's "non-grain" behavior and promote ecological nitrogen trade-off in North China. However, a promising approach to obtain the verified spatial distribution of nitrogen emissions remains to be developed, making it difficult to achieve the precise FQ estimation. Facing this issue, we present a Machine Learning (ML) - Nitrogen Export Verification (NEV) ensemble framework for the precise evaluation of FQ, taking the Beijing-Tianjin-Hebei 200 km traffic zone (zone) as the case. This was done by employing physical models for the precisely spatial estimation of Nitrogen Export (NE) values and then using ML methods to compute the spatial distribution of FQ using the Farmland Quality Evaluation System (FQES) indicators. We found: (1) the ML - NEV framework showed promising results, as the relative error of the NEV method was lower than 5.25 %, and the Determination coefficient of the ML method in FQ evaluation was higher than 84 %; (2) the FQ results within the zone were mainly good-quality areas (~47.25 % and primarily concentrated in the southwest-northeast regions) with improvement significance, with Fractal Dimension, NE values, and unbalanced Irrigation or Drainage Capabilities serving as the primary driving factors. Our results support finely weighing farmland improvement measures based on refined grids, contributing to offering decision supports for Agribusiness Revitalization Plans (i.e., safeguarding grain yield, activating agribusiness development, Etc.) in developing countries.