Attention-based Multi-modal Deep Learning Model of Spatio-temporal Crop Yield Prediction with Satellite, Soil and Climate Data
Attention-based Multi-modal Deep Learning Model of Spatio-temporal Crop Yield Prediction with Satellite, Soil and Climate Data
Authors: Gopal Krishna Shyam, Ila Chandrakar Date: 2026-04-21 Paper ID: openalex:2604.19217
Summary
The paper presents the Attention-Based Multi-Modal Deep Learning Framework (ABMMDLF) for improved spatio-temporal crop yield prediction by fusing satellite, meteorological, and soil data. The model employs a CNN architecture for spatial feature extraction coupled with a Temporal Attention Mechanism to dynamically weigh phenological periods. Experimental results demonstrate that this multi-modal approach achieves an R^2 of 0.89, significantly surpassing traditional static forecasting techniques.
Key Contributions
- Introduced the Attention-Based Multi-Modal Deep Learning Framework (ABMMDLF) for crop yield prediction.
- Integrated heterogeneous data sources including multi-year satellite imagery, high-resolution meteorological time-series, and soil property data.
- Achieved an R^2 score of 0.89, outperforming conventional single-modality baseline models.
Open Questions & Future Work
Archivist Review
I rejected the proposed framework as it is a specific model architecture tailored to a single domain application. I also rejected the blockchain-related open question, as it pertains to infrastructure and data integrity rather than the core predictive methodology or statistical bottlenecks of spatio-temporal forecasting. I approved the federated learning open question because distributed learning and privacy constraints represent a major, non-trivial, and recurring challenge in large-scale spatio-temporal forecasting.
Approved Open Questions
- Federated Learning for Agricultural Privacy: This is technically important as agricultural yield data is often proprietary or subject to strict privacy regulations, and traditional centralized training methods are becoming increasingly untenable for large-scale, cross-regional model deployment.
Rejected Candidates
- [concept] Attention-Based Multi-Modal Deep Learning Framework (ABMMDLF) (
abmmdlf-framework) - not_novel: This is a specific application-layer framework rather than a novel, reusable mechanism. - [open_question] Blockchain for Agricultural Data Integrity (
blockchain-integrity-agricultural-supply-chain) - low_impact: Blockchain integration is an application-level infrastructure concern rather than a core bottleneck in the predictive modeling of time-series data.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.19217
- paper_id
- 2604.19217
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:51:34Z