Scholarlink

<p>India’s agricultural sector, a cornerstone of its economy and food security, faces persistent challenges in accurately predicting crop yields due to sparse datasets and highly diverse agroclimatic conditions. Conventional models often fail to capture the nonlinear interactions among climatic, soil, and input variables, limiting their scalability and generalization. This study introduces the AgriSynth Framework, a novel hybrid machine learning approach that integrates synthetic data generation, domain-aware augmentation, and ensemble learning (LightGBM) to overcome data scarcity in agricultural prediction. The framework employs a dual-phase augmentation mechanism combining temporal and spatial diversification to expand a limited 102-record dataset into a 1000-entry synthetic dataset. Controlled temporal perturbations (± 15%) were applied to numeric features to simulate realistic variability; while this range is empirically chosen due to limited historical data, future work will statistically validate the perturbation against multi-year agricultural variance. Unlike previous studies that rely on limited regional datasets or generic augmentation, AgriSynth introduces a domain-aware synthetic data generation pipeline tailored for India’s agroclimatic diversity. Enhanced feature engineering and normalization ensure high model interpretability and convergence. Experimental evaluation demonstrates superior accuracy with MAE = 2.15 tons/ha, RMSE = 5.78 tons/ha, and R<sup>2</sup> = 0.92, validated through five-fold cross-validation. The proposed framework presents a scalable, data-efficient paradigm for yield forecasting, providing a reproducible tool for precision agriculture and evidence-based food security planning in data-scarce regions.</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

The AgriSynth framework for hybridized synthetic data generation and deep ensemble regression model for robust, scalable and interpretable crop yield forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Aradhana Sahu,
  • Shubh Raj Gupta,
  • Bhupesh Kumar Dewangan

摘要

India’s agricultural sector, a cornerstone of its economy and food security, faces persistent challenges in accurately predicting crop yields due to sparse datasets and highly diverse agroclimatic conditions. Conventional models often fail to capture the nonlinear interactions among climatic, soil, and input variables, limiting their scalability and generalization. This study introduces the AgriSynth Framework, a novel hybrid machine learning approach that integrates synthetic data generation, domain-aware augmentation, and ensemble learning (LightGBM) to overcome data scarcity in agricultural prediction. The framework employs a dual-phase augmentation mechanism combining temporal and spatial diversification to expand a limited 102-record dataset into a 1000-entry synthetic dataset. Controlled temporal perturbations (± 15%) were applied to numeric features to simulate realistic variability; while this range is empirically chosen due to limited historical data, future work will statistically validate the perturbation against multi-year agricultural variance. Unlike previous studies that rely on limited regional datasets or generic augmentation, AgriSynth introduces a domain-aware synthetic data generation pipeline tailored for India’s agroclimatic diversity. Enhanced feature engineering and normalization ensure high model interpretability and convergence. Experimental evaluation demonstrates superior accuracy with MAE = 2.15 tons/ha, RMSE = 5.78 tons/ha, and R2 = 0.92, validated through five-fold cross-validation. The proposed framework presents a scalable, data-efficient paradigm for yield forecasting, providing a reproducible tool for precision agriculture and evidence-based food security planning in data-scarce regions.