Scholarlink

<p>This study presents a hybrid multi-output regression framework for environmental forecasting in smart agriculture, integrating multi-output linear regression (MOLR) with random forest-based residual correction. Designed to address the inherent trade-off between interpretability and predictive power, the proposed architecture captures both global linear trends and localized non-linear deviations across a broad set of agro-environmental variables. The model was evaluated on the LH_Data dataset comprising 18 key environmental indicators, including temperature, humidity, pH, turbidity, and light intensity. Experimental results demonstrate significant performance gains, with an average R² of 0.96 and a mean absolute error (MAE) of 0.69—outperforming standalone MOLR, RF, and SVR baselines. Residual error analysis shows substantial variance reduction and improved distribution symmetry, particularly for highly volatile variables such as soil moisture and water quality parameters. To enhance transparency, SHAP-based explainability was applied, revealing the dominant influence of hydrological variables in the residual correction phase. Furthermore, a scalable deployment pipeline was developed using Scikit-learn’s pipeline API, incorporating edge inference for linear estimation and cloud-based refinement for non-linear residual adjustment. Overall, the proposed hybrid framework offers a robust, interpretable, and deployment-ready solution for real-time agro-environmental monitoring. It contributes both methodologically and practically to the development of explainable AI systems for data-driven decision-making in precision farming.</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

Hybrid multi-output regression with residual correction for smart agriculture: a scalable and interpretable approach

  • 期刊论文

  • 发表日期：

    2025年11月7日

查看全文
  • Nguyen Minh Son,
  • Do Si Truong,
  • Thanh Q. Nguyen

摘要

This study presents a hybrid multi-output regression framework for environmental forecasting in smart agriculture, integrating multi-output linear regression (MOLR) with random forest-based residual correction. Designed to address the inherent trade-off between interpretability and predictive power, the proposed architecture captures both global linear trends and localized non-linear deviations across a broad set of agro-environmental variables. The model was evaluated on the LH_Data dataset comprising 18 key environmental indicators, including temperature, humidity, pH, turbidity, and light intensity. Experimental results demonstrate significant performance gains, with an average R² of 0.96 and a mean absolute error (MAE) of 0.69—outperforming standalone MOLR, RF, and SVR baselines. Residual error analysis shows substantial variance reduction and improved distribution symmetry, particularly for highly volatile variables such as soil moisture and water quality parameters. To enhance transparency, SHAP-based explainability was applied, revealing the dominant influence of hydrological variables in the residual correction phase. Furthermore, a scalable deployment pipeline was developed using Scikit-learn’s pipeline API, incorporating edge inference for linear estimation and cloud-based refinement for non-linear residual adjustment. Overall, the proposed hybrid framework offers a robust, interpretable, and deployment-ready solution for real-time agro-environmental monitoring. It contributes both methodologically and practically to the development of explainable AI systems for data-driven decision-making in precision farming.