Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Learning county from pixels: corn yield prediction with attention-weighted multiple instance learning

Created by
  • Haebom

Author

Xiaoyu Wang, Yuchi Ma, Qunying Huang, Zhengwei Yang, Zhou Zhang

Outline

This paper proposes a novel approach that leverages pixel-level analysis and multiple instance learning to overcome the limitations of existing county-level spatial aggregation methods for predicting US corn yields. Specifically, we apply an attention mechanism to automatically assign pixel-specific weights to mitigate the effects of noise, addressing the issue of mixed pixels caused by resolution mismatches between satellite imagery and crop masks. Experimental results demonstrate that our proposed approach outperforms four existing machine learning models based on five years of data from the US Corn Belt, achieving a coefficient of determination (R²) of 0.84 and a root mean square error (RMSE) of 0.83 in 2022. We demonstrate the advantages of our approach from both spatial and temporal perspectives, and we verify its ability to remove noise and capture important feature information by analyzing the relationship between mixed pixels and the attention mechanism.

Takeaways, Limitations

•
Takeaways:
◦
Pixel-level analysis and multi-instance learning provide the potential for more accurate corn yield prediction than traditional county-level analysis.
◦
Solving mixed pixel problems and improving prediction accuracy by leveraging attention mechanisms.
◦
Achieved excellent predictive performance (R²=0.84, RMSE=0.83) in US corn growing regions in 2022.
◦
Validation of the effectiveness of the approach from spatial and temporal perspectives.
•
Limitations:
◦
Because this study used data limited to the U.S. corn-growing region, further research is needed to determine generalizability to other crops or regions.
◦
Prediction performance may be affected by the resolution and data quality of the satellite imagery used.
◦
There is a need to develop more sophisticated models that take into account various environmental factors (weather conditions, soil conditions, etc.).
👍
Made with Slashpage