Scholarlink

<p>Agricultural productivity faces significant challenges from erratic monsoon patterns, soil degradation, and climate variability, necessitating advanced predictive systems for informed decision-making. This study develops and validates a comprehensive machine learning framework for crop yield prediction and introduces a novel digital twin architecture with continuous feedback mechanisms for operational deployment in semi-arid rainfed agriculture. Data spanning five years (2019–2024) were compiled from eight districts covering 55,000 square kilometers, integrating weather parameters (temperature, rainfall, humidity, wind speed), soil properties (pH, organic carbon, NPK nutrients, electrical conductivity), and yield records for major crops including jowar, cotton, groundnut, and pulses. Four machine learning algorithms—Random Forest (RF), Gradient Boosting (GB), Support Vector Regression (SVR), and Artificial Neural Networks (ANN)—were systematically evaluated using five-fold cross-validation and rigorous hyperparameter optimization. Random Forest demonstrated superior performance with R² values of 0.87, 0.84, and 0.82 for jowar, cotton, and groundnut respectively, with corresponding mean absolute errors of 0.94, 2.21, and 1.52 quintals per hectare. Feature importance analysis revealed rainfall distribution during critical growth stages (14–18% importance) and soil organic carbon content (11–15% importance) as dominant predictors across all crops, providing actionable insights for agricultural interventions. The study’s distinctive contribution lies in the proposed digital twin framework enabling sequential prediction updates throughout the growing season through real-time data assimilation and automated model retraining protocols. Preliminary pilot implementation across 50 farmers in Dharwad and Belgaum districts demonstrated operational feasibility with comparable prediction accuracy (R²: 0.85–0.88) and high farmer satisfaction (85%). Spatial validation revealed systematic performance variation with weather station density and climate variability, with R² ranging from 0.79 to 0.90 across districts. This research advances precision agriculture in rainfed semi-arid systems by demonstrating that ensemble machine learning, coupled with continuous learning architectures, can transform agricultural decision-making under increasing climate uncertainty, offering a scalable framework for similar agro-ecological zones globally.</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

Digital twin-enabled crop yield prediction in semi-arid agriculture: a continuous feedback machine learning framework

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Shrikant Malgi,
  • Manjunath M K,
  • Avinash H T

摘要

Agricultural productivity faces significant challenges from erratic monsoon patterns, soil degradation, and climate variability, necessitating advanced predictive systems for informed decision-making. This study develops and validates a comprehensive machine learning framework for crop yield prediction and introduces a novel digital twin architecture with continuous feedback mechanisms for operational deployment in semi-arid rainfed agriculture. Data spanning five years (2019–2024) were compiled from eight districts covering 55,000 square kilometers, integrating weather parameters (temperature, rainfall, humidity, wind speed), soil properties (pH, organic carbon, NPK nutrients, electrical conductivity), and yield records for major crops including jowar, cotton, groundnut, and pulses. Four machine learning algorithms—Random Forest (RF), Gradient Boosting (GB), Support Vector Regression (SVR), and Artificial Neural Networks (ANN)—were systematically evaluated using five-fold cross-validation and rigorous hyperparameter optimization. Random Forest demonstrated superior performance with R² values of 0.87, 0.84, and 0.82 for jowar, cotton, and groundnut respectively, with corresponding mean absolute errors of 0.94, 2.21, and 1.52 quintals per hectare. Feature importance analysis revealed rainfall distribution during critical growth stages (14–18% importance) and soil organic carbon content (11–15% importance) as dominant predictors across all crops, providing actionable insights for agricultural interventions. The study’s distinctive contribution lies in the proposed digital twin framework enabling sequential prediction updates throughout the growing season through real-time data assimilation and automated model retraining protocols. Preliminary pilot implementation across 50 farmers in Dharwad and Belgaum districts demonstrated operational feasibility with comparable prediction accuracy (R²: 0.85–0.88) and high farmer satisfaction (85%). Spatial validation revealed systematic performance variation with weather station density and climate variability, with R² ranging from 0.79 to 0.90 across districts. This research advances precision agriculture in rainfed semi-arid systems by demonstrating that ensemble machine learning, coupled with continuous learning architectures, can transform agricultural decision-making under increasing climate uncertainty, offering a scalable framework for similar agro-ecological zones globally.