A phenology-guided deep learning framework for near real-time crop monitoring

The interplay between climate change and global population growth poses significant challenges to food security. To tackle this issue, near real-time (NRT) field-level crop monitoring plays an important role in enabling timely assessment of crop status and early warning of food security. With increased availability of satellite datasets, remote sensing provides a promising pathway for NRT crop monitoring. Recent advances in deep learning further open new opportunities in modeling the relationships between crop conditions and environmental factors with remote sensing imagery. Yet the NRT capabilities of current crop monitoring models are limited due to the difficulty of forecasting within-season crop phenological progress. The objective of this dissertation research is to develop a phenology-guided deep learning framework for NRT crop monitoring, leveraging remote sensing, deep learning, and in-situ field observations. Specifically, this research aims to (1) develop a robust hybrid deep learning fusion model to provide remote sensing images with more accurate spatiotemporal information throughout the crop growing season; (2) devise an emergence-based thermal phenological framework (EMET) for NRT crop type mapping with enhanced model scalability over space and time; and (3) develop a phenology foundation model for NRT phenology characterization and yield prediction with limited ground truth labels. With model interpretation methods, the relationship between crop yield and environmental variables in different phenological stages are examined. Results suggest the hybrid deep learning fusion model can better capture the temporal phenological changes among multi-source satellite images. Leveraging the fusion imagery, the EMET framework is able to achieve accurate characterization of crop distributions during the early growing season. The phenology foundation model, through a weak supervision approach, can accurately forecast crop phenological transition dates and crop yield. Model interpretation methods further provide insights into the key phenological stages when crop yield is more sensitive to changes in environmental conditions. This dissertation research proposes a phenology-guided solution for NRT crop monitoring, which provides critical support for agricultural sustainability and food security. The methodologies developed by this dissertation research hold great potential to be transferred to extended geographical regions for a better understanding of how ecosystems respond to the increasingly frequent disturbances induced by climate change.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Geography

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Yang, Zijun

Contributors dc:contributor

• Diao, Chunyuan
• Wang, Shaowen
• Lara, Mark J
• Sivapalan, Murugesu

• Remote Sensing
• Agriculture
• Deep Learning
• Phenology
• Near Real-time

Statement dc:rights

• Copyright 2024 Zijun Yang

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124677

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124677