Scholarlink

<p>Timely and accurate prediction of crop yield is essential for informed decision-making across food systems, ranging from supply chains optimization to national food security planning. Traditionally convolution and sequential modelling-based yield prediction approaches are effective at capturing long-range temporal dependencies across diverse regions. However, these existing approaches often lacks at capturing the non-linear, multi-scale temporal dynamics inherent in the crop yield patterns. To address these challenges, the current approach introduces an informer-based rice yield estimation framework leveraging multi-source inputs, including satellite imagery, historical yield patterns, and geographical inputs. After undergoing structured pre-processing and feature engineering, the derived inputs representations are passed to a deep multi-headed attention architecture to capture rich long-range dependencies, spatial heterogeneity and temporal patterns for improved yield estimation accuracy. Extensive evaluation utilising a multi-year district level dataset is carried out to assess the performance of the proposed approach using a leave-one year-out strategy. From the comparative assessment, it is found that the proposed approaches reduce the average prediction error by average 8%-10% for all years, with comparatively larger reductions for the high yielding states (Punjab, Uttar Pradesh and West Bengal). Furthermore, spatial error maps are generated to validate the robustness, generalizability, and suitability of the proposed approach for real-world agricultural scenarios.</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

Informer-driven yield forecasting with hybrid attention and distributional spatial features from remote sensing and agronomic data

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Akshat Bakshi,
  • Jatin Bedi

摘要

Timely and accurate prediction of crop yield is essential for informed decision-making across food systems, ranging from supply chains optimization to national food security planning. Traditionally convolution and sequential modelling-based yield prediction approaches are effective at capturing long-range temporal dependencies across diverse regions. However, these existing approaches often lacks at capturing the non-linear, multi-scale temporal dynamics inherent in the crop yield patterns. To address these challenges, the current approach introduces an informer-based rice yield estimation framework leveraging multi-source inputs, including satellite imagery, historical yield patterns, and geographical inputs. After undergoing structured pre-processing and feature engineering, the derived inputs representations are passed to a deep multi-headed attention architecture to capture rich long-range dependencies, spatial heterogeneity and temporal patterns for improved yield estimation accuracy. Extensive evaluation utilising a multi-year district level dataset is carried out to assess the performance of the proposed approach using a leave-one year-out strategy. From the comparative assessment, it is found that the proposed approaches reduce the average prediction error by average 8%-10% for all years, with comparatively larger reductions for the high yielding states (Punjab, Uttar Pradesh and West Bengal). Furthermore, spatial error maps are generated to validate the robustness, generalizability, and suitability of the proposed approach for real-world agricultural scenarios.