Accuracy and efficiency of compact foundation models for surface-water mapping during floods

Parastoo Farajpoor; Mohammadreza Narimani

eess.IVOct 26, 2026

Accuracy and efficiency of compact foundation models for surface-water mapping during floods

Authors: Parastoo Farajpoor, Mohammadreza Narimani

Organizations: Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA

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.

Figures & tables

Figure 1: Geographic coverage and the Bolivia holdout. (A) Locations, image dates, and hand-labeled chip counts for the 11 Sen1Floods11 events; symbol area scales with chip count. Bolivia is the Protocol A event holdout. (B) Bolivia chip locations and the metadata event footprint over Copernicus GLO-90 relief, with Natural Earth boundaries and rivers. (C) Two example chips, their reference-water boundaries, and Copernicus GLO-30 relief; unlabeled pixels are faded. Panel A uses an Equal Earth projection; the regional panels show WGS 84 geographic coordinates. The source cartography uses vertical exaggeration of 3 for regional shading and 12 for local shading, not for quantitative terrain analysis. Footprints describe the supplied event/chip coverage, not independently measured flood extents.

Event	S2 date	Chips	A: train / valid / test	B: train / valid	Water (%)	Unlabeled (%)	Permanent water (%)
Ghana	2018-09-19	53	31 / 11 / 11	221 / 78	6.3	33.8	1.4
India	2016-08-12	68	40 / 14 / 14	212 / 75	12.4	16	2.7
Cambodia	2018-08-04	30	18 / 6 / 6	234 / 83	24.1	9.3	7.6
Nigeria	2018-09-20	18	10 / 4 / 4	242 / 85	23.5	14.3	1.9
Pakistan	2017-06-28	28	16 / 6 / 6	236 / 83	7.5	26.7	2.5
Paraguay	2018-10-31	67	39 / 14 / 14	213 / 75	8.4	5	2.9

Table 1: Event composition and the two evaluation allocations.

Figure 2: Processing, architectures, and computational workflow. (A) Data preparation, evaluation allocations, methods, and diagnostics. The leave-one-event-out shorthand refers to exclusion from the fixed benchmark training/validation pools; global learning rates were selected once under Protocol A and reused. (B) Single-date six-band Prithvi segmentation. Tiny features are taken from blocks 3, 6, 9, and 12 and decoded through a feature pyramid and UPerNet head. The decoder family is shared, but the non-encoder parameter count differs: approximately 2.15 M for Tiny and 5.73 M for the larger model. (C) Local preparation, Colab training, and local analysis. The 106 reported training runs comprise 12 selection runs and 94 evaluation runs.

Figure 3: Dataset composition and input distributions. (A) Chip counts by event and Protocol A partition. (B) Valid-water fraction per chip, with event medians and a square-root horizontal scale; chips without valid labels are omitted. (C) Pixel composition by split; annotations give water as a fraction of labeled pixels. (D) Selected true-color images and hand labels, including the distinct treatment of unlabeled and out-of-swath areas. (E) Training-pool reflectance distributions compared with the reported HLS pretraining statistics; unequal distribution summaries are descriptive, not a formal domain-shift test. (F) Water fractions of the nested event-stratified training subsets. All training events remain represented, but both subset composition and optimization vary with seed.

Model	Encoder initialization	Updated components	Total (M)	Trainable (M)	Learning rate
U-Net	Random	All	7.76	7.76	$3\times 10^{-4}$
Tiny frozen	Pretrained	Neck + decoder + head	7.79	2.15	$10^{-3}$
Tiny fine-tuned	Pretrained	All	7.79	7.79	$10^{-3}/10^{-4}$
Tiny random	Random	All	7.79	7.79	$3\times 10^{-4}$
Prithvi-100M	Pretrained	All	91.97	91.97	$10^{-3}/10^{-4}$

Table 2: Neural architectures and selected learning rates.

Method	Water IoU mean $\pm$ SD	IoU 95% CI	F1	Precision	Recall	ECE
U-Net	0.856 $\pm$ 0.001	0.796–0.897	0.922	0.936	0.909	0.014
Tiny frozen	0.832 $\pm$ 0.004	0.772–0.877	0.908	0.925	0.892	0.002
Tiny fine-tuned	0.833 $\pm$ 0.001	0.772–0.879	0.909	0.932	0.887	0.004
Tiny random	0.809 $\pm$ 0.003	0.748–0.859	0.895	0.926	0.865	0.003
Prithvi-100M	0.841 $\pm$ 0.002	0.783–0.886	0.914	0.936	0.892	0.005
MNDWI fitted	0.787 $\pm$ 0.000	0.711–0.846	0.881	0.869	0.893	—

Table 3: Seen-event test performance with all training labels (Protocol A).

Figure 4: Qualitative segmentation comparisons under Protocol A, all labels, seed 0. Columns show true-color imagery, reference masks, MNDWI and neural error maps, and fine-tuned Tiny water probability. Error colors separate true positives, false positives, false negatives, true negatives, and invalid areas. Rows show a median eligible test example (A), the largest eligible relative Tiny and U-Net advantages (B,C), a median eligible Bolivia example (D), and a low-IoU example (E), with at most one chip per event. Selection used at least 5% reference water and 80% labeled pixels. These restricted and partly extreme examples do not represent a random sample of all chips. The annotated IoUs are single-chip, seed-0 values, not the three-seed pooled values in Tables 3 and 4 .

Method	Water IoU mean $\pm$ SD	IoU 95% CI	F1	ECE
U-Net	0.795 $\pm$ 0.014	0.685–0.867	0.886	0.015
Tiny frozen	0.689 $\pm$ 0.028	0.470–0.796	0.815	0.036
Tiny fine-tuned	0.752 $\pm$ 0.031	0.552–0.844	0.859	0.025
Tiny random	0.706 $\pm$ 0.022	0.427–0.826	0.828	0.021
Prithvi-100M	0.794 $\pm$ 0.007	0.651–0.863	0.885	0.016
MNDWI fitted	0.704 $\pm$ 0.000	0.529–0.806	0.827	—

Table 4: Bolivia event-holdout performance with all training labels (Protocol A).

Figure 5: Accuracy and calibration with all training labels. (A) Pooled water IoU on 90 seen-event test chips (filled) and 15 Bolivia chips (open), with three-seed means, individual seeds, and chip-bootstrap 95% intervals. (B) Precision-recall positions and F1 contours. (C) Paired IoU differences between fine-tuned Tiny and each comparator; filled markers identify intervals excluding zero. (D,E) Reliability diagrams for the seen-event test set and Bolivia, pooled across seeds; legends report mean seed ECE. Fifteen equal-width bins are used, and bins with fewer than 1,000 pixels are omitted from the curves. The displayed intervals concern IoU, not precision or recall. Index methods are omitted from probability-calibration panels.

A. Seen-event test
Method	25 chips ( 10%)	63 chips (25%)	126 chips (50%)	252 chips (100%)
U-Net	0.831 $\pm$ 0.014	0.844 $\pm$ 0.006	0.857 $\pm$ 0.005	0.856 $\pm$ 0.001
Tiny frozen	0.784 $\pm$ 0.052	0.822 $\pm$ 0.003	0.822 $\pm$ 0.010	0.832 $\pm$ 0.004
Tiny fine-tuned	0.814 $\pm$ 0.013	0.834 $\pm$ 0.003	0.834 $\pm$ 0.005	0.833 $\pm$ 0.001
Tiny random	0.768 $\pm$ 0.011	0.811 $\pm$ 0.008	0.813 $\pm$ 0.007	0.809 $\pm$ 0.003
Prithvi-100M	0.820 $\pm$ 0.005	—	—	0.841 $\pm$ 0.002

Table 5: Water IoU under reduced fine-tuning-label budgets (Protocol A).

Figure 6: Sensitivity to the number of labeled fine-tuning chips. (A,B) Seen-event and Bolivia water IoU for nested training subsets, with three-seed means and individual seed points. The validation pool and global hyperparameter selection remain fixed. Prithvi-100M has measurements only at 25 and 252 chips; its connecting line does not imply intermediate experiments. (C,D) Paired differences for pretrained versus random Tiny, frozen versus random Tiny, fine-tuned Tiny versus U-Net, and 100M versus fine-tuned Tiny. Bars are paired chip-bootstrap 95% intervals; filled markers indicate intervals excluding zero. Seed variation combines subset and optimization effects. Table 5 reports means and standard deviations, not confidence intervals for the upper curves.

Method	Parameters total / trainable (M)	GFLOPs / 256-px tile	Training (min)	Training GPU memory (GB)	GPU (ms/chip)	CPU (ms/chip)
U-Net	7.8 / 7.8	24.3	5.6	2.5	30.6	665
Tiny frozen	7.8 / 2.2	9.6	2.7	1.8	15.9	258
Tiny fine-tuned	7.8 / 7.8	9.6	3.3	2.0	16.1	258
Tiny random	7.8 / 7.8	9.6	3.3	2.2	16.8	258
Prithvi-100M	92.0 / 92.0	53.0	8.7	4.7	50.4	1386
MNDWI fitted	0	—	—	—	—	4.9

Table 6: Measured computational cost of the evaluated implementations.

Figure 7: Accuracy and computational cost. (A,B) Three-seed pooled IoU against GPU and CPU latency per 512 $\times$ 512 chip. Filled markers denote seen-event test results and open markers Bolivia; marker area encodes parameter count. Tiny variants are offset horizontally for legibility, and share one CPU timing. (C) Median training time with all labels and time per optimization step. (D) Peak GPU memory for training and full-chip inference. (E) Total/trainable parameter counts and per-tile floating-point operations. GPU timings use a Tesla T4, mixed precision, and nine overlapping tiles per chip; CPU timings use an i9-12900H with float32. Measurements exclude data acquisition and file I/O and do not estimate energy consumption.

Method	Mean event IoU (95% CI)	Worst event (IoU)	Events above U-Net	Unseen - seen IoU (95% CI)	Mean ECE
U-Net	0.777 (0.705–0.843)	Paraguay (0.522)	—	-0.009 (-0.020, -0.001)	0.030
Tiny frozen	0.765 (0.705–0.823)	Paraguay (0.564)	3/11	-0.011 (-0.018, -0.006)	0.010
Tiny fine-tuned	0.768 (0.703–0.828)	Paraguay (0.573)	3/11	-0.004 (-0.014, +0.006)	0.009
Tiny random	0.725 (0.657–0.792)	Paraguay (0.519)	1/11	-0.008 (-0.028, +0.010)	0.013
Prithvi-100M	not run	not run	not run	not run	not run
MNDWI fitted	0.730 (0.671–0.792)	Pakistan (0.602)	2/11	-0.002 (-0.008, +0.003)	—

Table 7: Event-excluded fitting and matched seen/unseen comparisons (Protocol B).

Figure 8: Event-level performance and reliability. (A,B) Protocol B event IoUs and paired Tiny-U-Net differences; the event mean gives equal weight to all 11 targets. These are event-excluded fits with shared Protocol A learning-rate selection, not fully nested tuning. (C) Changes on identical official test chips under seen-event versus event-excluded seed-0 fitting. (D) Recall for JRC-permanent and other reference water; the latter is not verified newly inundated land. (E) Boundary and interior error rates and the fraction of errors near boundaries. (F) Pixel error versus entropy, with error-detection AUROC. Panels D-F use Protocol A, all labels, seed 0. (G) Clear-image examples from Paraguay and Pakistan, two low-performing events; the Paraguay example itself has high IoU.

Flood Inundation Mapping

Flood inundation mapping uses satellite imagery, terrain data, and models to locate flooded areas and estimate water depth for response and planning.

See all 27 papers

Flood Mapping from RGB imagery using a Vision Foundation Model
Jun 23, 2026Vladyslav Polushko, Tilman Bucher, Ronald Rösch +3Remote Sensing Image Segmentation Geospatial Foundation Models
Image Processing Department, Fraunhofer ITWM, Kaiserslautern, Germany · Institut für Optische Sensorsysteme, DLR, Berlin, Germany · ACIDA Lab, Hochschule Darmstadt, Darmstadt, Germany +1
Cross-Architecture Foundation-Model Distillation for Edge Flood Segmentation
Sep 17, 2026Fabian Schmalstieg, Karsten Mueller, Wojciech SamekRemote Sensing Image Segmentation Geospatial Foundation Models
Department of Artificial Intelligence, Fraunhofer Heinrich Hertz Institute, 10587 Berlin, Germany
Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events
Jun 5, 2026Venkatesh Kolluru, Rajat Shinde, Abdelhak Marouane +6Remote Sensing Image Segmentation Geospatial Foundation Models
Earth System Science Center, University of Alabama in Huntsville, Huntsville, AL 35805, USA · Space and Earth Science Data Analysis, 7515 Mission Drive, Lanham, MD 20706 · NASA Marshall Space Flight Center, Huntsville, AL, 35808 USA.

Jun 23, 2026J/K move · Enter open · S save

Flood Mapping from RGB imagery using a Vision Foundation Model

Vladyslav Polushko, Tilman Bucher, Ronald Rösch, Thomas März+2

Image Processing Department, Fraunhofer ITWM, Kaiserslautern, Germany · Institut für Optische Sensorsysteme, DLR, Berlin, Germany · ACIDA Lab, Hochschule Darmstadt, Darmstadt, Germany+1

Remote Sensing Image Segmentation Geospatial Foundation Models Domain Generalization Vision Foundation Model Adaptation Flood Inundation Mapping Remote Sensing

Open paper AskPDF arXiv

Timely, high-resolution maps of flood extent around settlements are essential for emergency response and damage assessment. We consider airborne RGB imagery for flood mapping as it can be collected rapidly at low cost. To produce flood maps, deep learning models for water segmentation are often used. CNN based and small vision transformer models are used. However, they need much data for adaptation to a change of scenery, i.e., another flooding event. Vision foundation models or large vision transformers are known to generalize across domains. Recently, foundation models for Earth observation became available. They are pretrained on satellite data, whose spatial resolution, viewing geometry, and radiometry differ from nadir RGB imagery. Thus, adaptation is required. We investigate how a satellite-pretrained Earth observation foundation model can be adapted to centimeter-scale floodwater mapping from RGB imagery. Specifically, we fine-tune a model we call Prithvi-2.0-UPN consisting of the Prithvi-EO-2.0-600M Vision Transformer combined with a UPerNet decoder for binary water segmentation on two RGB datasets (BlessemFlood21, NeuenahrFlood). In a first experiment we observe that Prithvi-2.0-UPN reaches state-of-the-art results on BlessemFlood21 and NeuenahrFlood, when trained on their datasets. In a second experiment we show that Prithvi-2.0-UPN performs better than state-of-the-art baseline models for transfer to a new flood event (trained on BlessemFlood21, tested on NeuenahrFlood) in a zero-shot setting. However, the performance indicates room for improvement. In this respect, we investigate in a third experiment how performance improves when further fine-tuning the models with small shares of NeuenahrFlood training data: Prithvi-2.0-UPN improves the fastest and reaches almost the performance level when fully trained on NeuenahrFlood, indicating transfer capabilities.

Abstract

Flood Inundation Mapping

Explore similar work

Flood Mapping from RGB imagery using a Vision Foundation Model

Cross-Architecture Foundation-Model Distillation for Edge Flood Segmentation

Land cover and flood type govern the detection limits of satellite-based flood mapping across diverse global flood events