Scholarlink

<p>Accurate crop yield estimation is vital for global food security and trade. Despite the superior performance of deep learning (DL) models, their “black box” nature undermines trust in critical decision-making. Explainable artificial intelligence (XAI) improves transparency and interactivity, but the ability of XAI models to balance interpretability and predictive accuracy remains insufficiently assessed, limiting their broader application in real-world scenarios. This study proposes the XAI-Crop framework and, using major soybean-producing countries as case studies and a multi-country comparative experimental design, employs multi-source data to evaluate and compare the performance differences between the inherently interpretable DL model—Kolmogorov-Arnold Networks (KAN)—and Multilayer Perceptron (MLP) as well as Random Forest (RF) models. In small-sample settings, KAN achieves predictive accuracy and generalization comparable to MLP and RF while offering improved interpretability. Feature importance analysis reveals significant regional variability in yield-driving factors, with solar-induced chlorophyll fluorescence (SIF) consistently emerging as a highly sensitive predictor across all regions. These findings demonstrate the feasibility of XAI approaches, such as KAN, in bridging the gap between model accuracy and interpretability, paving the way for their integration into agricultural decision-support systems and contributing to sustainable agricultural development.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From data to decisions: the use of explainable AI to forecast soybean yield in major producing countries

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月13日

查看全文
  • Xiangyi Wang,
  • Yingbin He,
  • Huicong Chen,
  • Shanjun Luo,
  • Yifan Jiao,
  • Jiong Ning,
  • Anran Feng,
  • Shengnan Han,
  • Yixuan Duan,
  • Sijie Fan,
  • Jiyao Yin

摘要

Accurate crop yield estimation is vital for global food security and trade. Despite the superior performance of deep learning (DL) models, their “black box” nature undermines trust in critical decision-making. Explainable artificial intelligence (XAI) improves transparency and interactivity, but the ability of XAI models to balance interpretability and predictive accuracy remains insufficiently assessed, limiting their broader application in real-world scenarios. This study proposes the XAI-Crop framework and, using major soybean-producing countries as case studies and a multi-country comparative experimental design, employs multi-source data to evaluate and compare the performance differences between the inherently interpretable DL model—Kolmogorov-Arnold Networks (KAN)—and Multilayer Perceptron (MLP) as well as Random Forest (RF) models. In small-sample settings, KAN achieves predictive accuracy and generalization comparable to MLP and RF while offering improved interpretability. Feature importance analysis reveals significant regional variability in yield-driving factors, with solar-induced chlorophyll fluorescence (SIF) consistently emerging as a highly sensitive predictor across all regions. These findings demonstrate the feasibility of XAI approaches, such as KAN, in bridging the gap between model accuracy and interpretability, paving the way for their integration into agricultural decision-support systems and contributing to sustainable agricultural development.