HiRes-FusedMIM: A High-Resolution RGB-digital surface model (DSM) Pre-trained Model for Building-Level Remote Sensing Applications

[guneetmutreja]

"Remote Sensing Vision Foundation Models"

It is a novel pre-trained model specifically designed to leverage the rich information contained within high-resolution RGB and DSM data.

Paper: https://arxiv.org/html/2503.18540

Contributions:

1. A Large-Scale, High-Resolution, Paired RGB-DSM Dataset: We curated a unique dataset of over 368k paired RGB and DSM images at 0.2-0.5 m resolution
2. A dual encoder SimMIM architecture [40] with separate encoders for RGB and DSM data
3. Experiments showcase superior performance compared to models trained on lower-resolution or single-modality data, validating the benefits of incorporating high-resolution DSMs in the pre-training process. To facilitate further research in this direction, trained model weights will be publically shared.