A spatial hybrid model for crop yield prediction in Western Australia

Authors

Muhammad Ibrahim, Department of Primary Industries and Regional Development, Western AustraliaFollow
Balwinder Singh, Department of Primary Industries and Regional Development, Western AustraliaFollow
Rodrigo Pires, Department of Primary Industries and Regional Development, Western AustraliaFollow
Julia Easton, Curtin University of Technology
Hasnein Tareque, Department of Primary Industries and Regional Development, Western AustraliaFollow

Document Type

Article

Publication Date

11-25-2025

Journal Title

IEEE International Geoscience and Remote Sensing Symposium Proceedings

ISBN

Electronic ISBN: 979-8-3315-0810-4, Print on Demand ISBN: 979-8-3315-0811-1

ISSN

Electronic ISSN: 2153-7003, Print on Demand ISSN: 2153-6996

Keywords

Deep learning, Accuracy, Food security, Estimation, Predictive models, Transformers, Data models, Geospatial analysis, Crop yield, Residual neural networks, Crop yield estimation, hybrid model, MAMBA, Transformer, slot attention, geospatial, weather, Sentinel-2, ResNet-50, WA Paddock Dataset

Disciplines

Agricultural Science | Agronomy and Crop Sciences

Abstract

Accurate crop yield estimation is crucial for ensuring food security and policy making; however, traditional methods are labor-intensive, and machine learning models often struggle to capture the complex interactions between soil, weather, and temporal data. To address these challenges, this paper first introduces a comprehensive resource—named the WA Paddock Dataset—comprising soil, climate, and satellite data for approximately 450,000 paddocks across Western Australia (WA) over a three-year period. This dataset is designed to enable high-resolution and large-scale modeling and supports other diverse research applications. Building on this resource, we present a novel hybrid regression model for predicting paddock-level crop yield that integrates MAMBA blocks, Transformer attention mechanisms, and Slot Attention to effectively capture spatial and temporal intricacies. This architecture effectively captures spatial and temporal complexities by leveraging diverse geospatial data, including soil properties, weather patterns, and Sentinel-2 imagery, to enhance predictive accuracy. Evaluation against classical machine learning models and ResNet50 demonstrates that our hybrid model significantly improves accuracy while achieving faster inference speeds compared to ResNet50 and some traditional approaches. These results establish the proposed method as a robust and efficient solution for precision agriculture and guiding impactful policy development at regional scale. Our code and relevant dataset will be made publicly available through GitHub repository.

Recommended Citation

M. Ibrahim, B. Singh, R. Pires, J. Easton and H. Tareque, "A Spatial Hybrid Model for Crop Yield Prediction in Western Australia," IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium, Brisbane, Australia, 2025, pp. 614-618, doi: 10.1109/IGARSS55030.2025.11243086.