CrossPatch: A Flexible Patch Embedding Strategy for Foundation Models in Remote Sensing
Lists
Only you see this
Abstract
Remote sensing imagery is pivotal in environmental monitoring, disaster assessment, and agricultural management applications. Pretrained Foundation Models (PFMs), particularly vision transformer-based architectures, achieve state-of-the-art performance in various remote sensing tasks. However, their reliance on fixed-resolution training limits adaptability, causing significant performance drops when applied to downstream tasks with varying resolutions or image sizes. To address this, we introduce CrossPatch, a flexible patch embedding framework centered on an innovative alignment-embedding mechanism to enhance the generalization of patch embeddings across diverse input configurations. This mechanism ensures seamless embedding alignment while preserving semantic integrity by leveraging robust resampling techniques. CrossPatch dynamically adjusts patch sizes to balance fine-grained feature extraction and computational efficiency. Comprehensive experiments across varying resolutions, sizes, and channels demonstrate that CrossPatch significantly improves PFM adaptability and performance in diverse downstream tasks.
BibTeX
@misc{li2025crosspatch,
author = {Xuyang Li and Chen-Yu Li and Danfeng Hong},
title = {{CrossPatch: A Flexible Patch Embedding Strategy for Foundation Models in Remote Sensing}},
year = {2025},
doi = {10.1109/igarss55030.2025.11242337},
}Loading related content...