Graduation Year

2026

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Sudeep Sarkar, Ph.D.

Committee Member

Lawrence Hall, Ph.D.

Committee Member

Dmitry Goldgof, Ph.D.

Committee Member

Susana Lai-Yuen, Ph.D.

Committee Member

Joni Downs Firat, Ph.D.

Keywords

Satellite Imagery, Remote Sensing, Deep Learning, Self-Supervised Learning, Foundation Models

Abstract

Research on large-scale geospatial understanding from satellite imagery remains relativelyunderexplored. Much of the field still relies on narrow, task-specific pipelines, and there is lim- ited work on developing true foundation models or on systematically evaluating how well these models generalize under domain shift, geographic variation, and real-world scene diversity. This dissertation advances foundation-model pretraining and evaluation for remote sensing along three complementary axes. First, we develop self-supervised multispectral pretrained models for satellite image classification by adapting BYOL, BarlowTwins, DINO, MoCo, and Simsiam to multispectral inputs, addressing the limited amount of curated satellite datasets relative to ImageNet. Pretraining on Eurosat and BigEarthNet yields strong transfer for classification across datasets, providing multispectral pretrained models for BigEarthNet that have yet to be fully explored in prior work, and demonstrating competitive performance despite using a fraction of the image samples typically used in ImageNet-scale training. Second, to study robustness and generalization of geospatial foundation models beyond narrow demonstrations, we benchmark eight state-of-the-art ViT-based foundation models, spanning self-supervised MAE variants (SatMAE, Scale-MAE, Cross-Scale MAE), the Geospatial Foundation Model, and supervised Multi-Task Pretraining (MTP). Models pretrained on fMoW, SAMRS, Million-AID, or GeoPile are evaluated via end-to-end fine-tuning on four down- stream tasks—classification, segmentation, oriented object detection, and change detection—across DOTA-v2.0, BigEarthNet, and xDB. Results show that performance is strongly task dependent: foundation models match or exceed supervised baselines for classification and segmentation, yet struggle with change detection and oriented object detection, and cross-dataset performance is hampered by differences in imagery from different satellites. Across settings, pretraining data quality, particularly spatial resolution and scene diversity, drives downstream success more than sheer dataset size, and scaling from ViT-Base to ViT-Large yields less than a 1% improvement.

Finally, to enable resource-constrained learning, we introduce Hierarchical Attention Dis- tillation (HAD), a pretraining paradigm that transfers spatial and semantic representations from a large teacher foundation model to a compact student via multi-level attention across feature maps at multiple spatial scales, improving feature-transfer fidelity and accelerating convergence. Unlike conventional supervised knowledge distillation that relies on the teacher’s labeled pretraining data, the student is pretrained on a separate unlabeled dataset (SSL4EO), distinct from the MillionAID dataset used for the teacher, and then fine-tuned on classification and segmentation. The distilled compact model achieves performance closely aligned with current state-of-the-art satellite founda- tion models while substantially improving computational efficiency and fine-tuning time, supporting practical deployment where parameters and latency are constrained

Download

Share

COinS