Accuracy and efficiency of compact foundation models for surface-water mapping during floods
Abstract
Pretrained Earth observation models may reduce the effort needed to map surface water, but their value depends on the alternative against which they are compared. We evaluated Prithvi-EO-2.0-Tiny using the 446 hand-labeled Sentinel-2 chips from 11 flood events in Sen1Floods11. A fine-tuned model, a frozen-encoder model, and an identical randomly initialized architecture were compared with a similarly sized U-Net, a larger Prithvi model, and spectral-index thresholds. Experiments varied the number of training chips, excluded events from model fitting, and measured computational cost and prediction reliability. With 252 training chips, U-Net achieved a water intersection-over-union (IoU) of 0.856 on the official test set, compared with 0.833 for fine-tuned Prithvi-Tiny. Pretraining improved the tiny transformer over random initialization by a reported 0.046 IoU with 25 training chips and 0.023 with 252 chips, but did not reverse the U-Net ranking. Across 11 event-exclusion evaluations, mean IoU was 0.768 for fine-tuned Prithvi-Tiny and 0.777 for U-Net; the paired interval included zero. Fine-tuned Prithvi-Tiny processed a 512x512-pixel chip in 16.1 ms on a GPU and 258 ms on a laptop CPU, compared with 30.6 and 665 ms for U-Net. Errors were concentrated near water boundaries, and high predictive entropy identified many, but not all, erroneous pixels. These results support a specific role for compact pretrained models: lower measured computational demand with competitive accuracy, rather than consistent superiority. Event-dependent performance, shared hyperparameter selection, and the distinction between observed water and newly inundated land delimit the conclusions.
Related
- Feynman Propagators on Exponentially Decaying Perturbations of Stationary Spacetimes with Compact Cauchy Surface
- Sharp lower bounds for the first eigenvalue of Steklov-type eigenvalue problems on a compact surface
- Generating the twist subgroup of the level $2$ mapping class group of a non-orientable closed surface
- Multi-sensor deep-learning mapping of surface and basal fractures across Antarctic ice shelves during the last four decades
- What masking geometry works best for EEG foundation models?
BibTeX
@misc{farajpoor2026accuracy,
title = {{Accuracy and efficiency of compact foundation models for surface-water mapping during floods}},
author = {Parastoo Farajpoor and Mohammadreza Narimani},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.34904},
eprint = {2610.34904},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.34904},
}
RIS
TY - UNPB
TI - Accuracy and efficiency of compact foundation models for surface-water mapping during floods
AU - Parastoo Farajpoor
AU - Mohammadreza Narimani
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.34904
UR - https://arxiv.org/abs/2610.34904
AB - Pretrained Earth observation models may reduce the effort needed to map surface water, but their value depends on the alternative against which they are compared. We evaluated Prithvi-EO-2.0-Tiny using the 446 hand-labeled Sentinel-2 chips from 11 flood events in Sen1Floods11. A fine-tuned model, a frozen-encoder model, and an identical randomly initialized architecture were compared with a similarly sized U-Net, a larger Prithvi model, and spectral-index thresholds. Experiments varied the number of training chips, excluded events from model fitting, and measured computational cost and prediction reliability. With 252 training chips, U-Net achieved a water intersection-over-union (IoU) of 0.856 on the official test set, compared with 0.833 for fine-tuned Prithvi-Tiny. Pretraining improved the tiny transformer over random initialization by a reported 0.046 IoU with 25 training chips and 0.023 with 252 chips, but did not reverse the U-Net ranking. Across 11 event-exclusion evaluations, mean IoU was 0.768 for fine-tuned Prithvi-Tiny and 0.777 for U-Net; the paired interval included zero. Fine-tuned Prithvi-Tiny processed a 512x512-pixel chip in 16.1 ms on a GPU and 258 ms on a laptop CPU, compared with 30.6 and 665 ms for U-Net. Errors were concentrated near water boundaries, and high predictive entropy identified many, but not all, erroneous pixels. These results support a specific role for compact pretrained models: lower measured computational demand with competitive accuracy, rather than consistent superiority. Event-dependent performance, shared hyperparameter selection, and the distinction between observed water and newly inundated land delimit the conclusions.
ER -