PhD Final Defense – Yi-Chia Chang

Machine Learning and Remote Sensing for Agriculture

Advisors: Professor Ximing Cai & Professor Arindam Banerjee

ABSTRACT

Accurate large-scale agricultural monitoring is critical for global food security. In particular, crop

type mapping and yield prediction are two fundamental tasks that provide critical insights for

resource management, food production assessment, and the market. However, the scarcity of

high-resolution crop type and yield datasets limits the performance and generalizability of

machine learning models for these tasks. Pre-trained foundation models have the potential to

bridge the gap by leveraging implicit knowledge learned through self-supervised training on

large pretraining datasets. This dissertation investigates machine learning, remote sensing, and

foundation models to improve the accuracy and scalability of crop type classification and crop

yield prediction.

The dissertation is structured into three interconnected parts. The first part investigates the

generalizability of pre-trained Earth observation (EO) foundation models for crop type mapping

across five continents. By analyzing the performance of EO foundation models across diverse

geographic regions, we assess models' ability to generalize in data-scarce areas and examine the

effects of dataset size and class imbalance in both in-distribution (ID) and out-of-distribution

(OOD) settings. Building on the foundation model experiments, the second part investigates

spatiotemporal machine learning for crop yield prediction. We compare state-of-the-art timeseries

models with traditional machine learning baselines to evaluate their effectiveness. The

third part further explores the effectiveness of multimodal foundation models for in-season crop

yield prediction. We evaluate the foundation models by fine-tuning models and integrating

geospatial embeddings extracted from large-scale EO foundation models. By integrating remote

sensing, foundation models, and advanced machine learning techniques, this research improves

the accuracy and generalizability of crop monitoring systems. Via our integrated datasets and

modeling pipelines, we also contribute to the open-source TorchGeo library, enabling

reproducible crop-type mapping research for the broader Earth observation community. The

findings and machine learning pipelines support the development of scalable and transferable

models for agricultural assessment, ultimately supporting food security and sustainable

agricultural practices.