Scholarlink

<p>This is because accurate crop yield prediction is a critical tool for food security and for making optimal use of agricultural resources. Traditional statistical methods and vegetation index-based models are usually unable to capture the complex spatial, spectral, and temporal dynamics of crop growth. While deep learning has exhibited impressive accuracy in some agricultural forecasting tasks, numerous current methods rely solely on observations from a single season, offer limited interpretability, and cannot adapt predictions to different agricultural regions. To overcome these constraints, this work introduces YieldFusionAI, a multi-season, spatiotemporal deep learning model capable of providing explainable, satellite-derived, spectral-based crop yield forecasting. The proposed framework features a hybrid architecture, YieldNet-ST, which integrates Convolutional Neural Networks (CNNs) to capture spatial features with Bidirectional Long Short-Term Memory (Bi-LSTM) networks and attention mechanisms to model crop growth dynamics over time. Multi-season spectral representations are derived from vegetation indices such as NDVI, EVI, and SAVI, as well as spatial transformation features. An explainable module based on the SHAP method is embedded to quantify the contributions of the spectral-temporal predictors, and a region-specific post-inference calibration module is implemented to reduce systematic residual errors and improve prediction reliability, thereby enhancing transparency. The proposed framework outperforms the conventional machine learning and deep learning baselines, as evidenced by experimental evaluation using multi-season Sentinel-2 imagery and crop yield records, with an RMSE of 0.273 t ha⁻¹, MAE of 0.213 t ha⁻¹, R² of 0.894, and MAPE of 6.82%. Additional ablation and calibration analyses, along with geospatial visualisations, further validate the robustness, interpretability, and usability of the proposed framework. The findings suggest that YieldFusionAI offers a promising and practical approach to crop yield prediction and decision-making support in agriculture. The source code, implementation details, and reproducibility resources for YieldFusionAI are publicly available at <a href="https://github.com/yedukondalu8919/YieldFusionAI">https://github.com/yedukondalu8919/YieldFusionAI</a>.</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

Explainable multi season spatio temporal deep learning framework for crop yield forecasting using sentinel 2 remote sensing data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Yedukondalu Gamidell,
  • Mousmi Ajay Chaurasia

摘要

This is because accurate crop yield prediction is a critical tool for food security and for making optimal use of agricultural resources. Traditional statistical methods and vegetation index-based models are usually unable to capture the complex spatial, spectral, and temporal dynamics of crop growth. While deep learning has exhibited impressive accuracy in some agricultural forecasting tasks, numerous current methods rely solely on observations from a single season, offer limited interpretability, and cannot adapt predictions to different agricultural regions. To overcome these constraints, this work introduces YieldFusionAI, a multi-season, spatiotemporal deep learning model capable of providing explainable, satellite-derived, spectral-based crop yield forecasting. The proposed framework features a hybrid architecture, YieldNet-ST, which integrates Convolutional Neural Networks (CNNs) to capture spatial features with Bidirectional Long Short-Term Memory (Bi-LSTM) networks and attention mechanisms to model crop growth dynamics over time. Multi-season spectral representations are derived from vegetation indices such as NDVI, EVI, and SAVI, as well as spatial transformation features. An explainable module based on the SHAP method is embedded to quantify the contributions of the spectral-temporal predictors, and a region-specific post-inference calibration module is implemented to reduce systematic residual errors and improve prediction reliability, thereby enhancing transparency. The proposed framework outperforms the conventional machine learning and deep learning baselines, as evidenced by experimental evaluation using multi-season Sentinel-2 imagery and crop yield records, with an RMSE of 0.273 t ha⁻¹, MAE of 0.213 t ha⁻¹, R² of 0.894, and MAPE of 6.82%. Additional ablation and calibration analyses, along with geospatial visualisations, further validate the robustness, interpretability, and usability of the proposed framework. The findings suggest that YieldFusionAI offers a promising and practical approach to crop yield prediction and decision-making support in agriculture. The source code, implementation details, and reproducibility resources for YieldFusionAI are publicly available at https://github.com/yedukondalu8919/YieldFusionAI.