HiT: History-Injection Transformers for Onboard Continuous Flood Change Detection

TL;DR

This work tackles onboard, real-time flood change detection under the tight memory and compute limits of nanosatellites. It introduces History Injection Transformer (HiT), which embeds a compact History Embedding into a ViT-based Prithvi-EO-2.0-tiny encoder, enabling continual multi-temporal analysis without storing full historical images. HiT-Prithvi achieves a strong balance between memory efficiency and detection performance, reaching 43 FPS on a Jetson Orin Nano while reducing storage by about 99.6% and attaining an F1-score comparable to a bitemporal baseline on STTORM-CD. The approach demonstrates practical viability for autonomous, ground-free disaster monitoring and sets a foundation for extending foundation-models to resource-constrained onboard EO inference.

Abstract

Natural disaster monitoring through continuous satellite observation requires processing multi-temporal data under strict operational constraints. This paper addresses flood detection, a critical application for hazard management, by developing an onboard change detection system that operates within the memory and computational limits of small satellites. We propose History Injection mechanism for Transformer models (HiT), that maintains historical context from previous observations while reducing data storage by over 99\% of original image size. Moreover, testing on the STTORM-CD flood dataset confirms that the HiT mechanism within the Prithvi-tiny foundation model maintains detection accuracy compared to the bitemporal baseline. The proposed HiT-Prithvi model achieved 43 FPS on Jetson Orin Nano, a representative onboard hardware used in nanosats. This work establishes a practical framework for satellite-based continuous monitoring of natural disasters, supporting real-time hazard assessment without dependency on ground-based processing infrastructure. Architecture as well as model checkpoints is available at https://github.com/zaitra/HiT-change-detection

Figures (9)

• Figure 2: HiT mechanism. The History Embedding is transformed to the required dimensionality and fed into the ViT block. The new embedding is extracted from the block's output. Image input tokens are shown as blue, embedding in green (projected version) and yellow. (PE - position encoding, C - concatenation)
• Figure 3: Result of modified CutMix augmentation. Top line shows image sequence, bottom line shows corresponding change mask. A change is introduced in step $t-2$.
• Figure 4: Model performance with varying History Embedding fusion stage and the bitemporal baseline B.
• Figure 5: Model performance with varying History Embedding dimension and the bitemporal baseline B.
• Figure 6: Model performance with varying History Embedding spatial dimension and the bitemporal baseline B.