Earth Observation Foundation Models for Terrestrial Ecohydrology: From Representation Learning to Process Inference

Yi Yu; Jian Peng; Yucheng Lin; Trevor F. Keenan; Thomas F. A. Bishop

cs.LGAug 15, 2026

Earth Observation Foundation Models for Terrestrial Ecohydrology: From Representation Learning to Process Inference

Authors: Yi Yu, Jian Peng, Yucheng Lin, Trevor F. Keenan, Thomas F. A. Bishop

Organizations: Precision Agriculture, Hydrology and Geoinformation Science Laboratory, School of Life and Environmental Sciences, The University of Sydney, Eveleigh, NSW, Australia · Department of Remote Sensing, Helmholtz Centre for Environmental Research–UFZ, Leipzig, Germany · Institute for Earth System Science and Remote Sensing, Leipzig University, Leipzig, Germany · School of Energy and Environment, City University of Hong Kong, Hong Kong SAR, China ·

Abstract

Earth observation foundation models (EOFMs) are emerging as reusable representation frameworks for data-driven retrieval, prediction and process modelling within ecohydrology, which integrate EO, meteorological forcing and process models to characterise coupled water, energy and carbon dynamics in vegetation and soil across scales. However, there is yet to be an ecohydrology-specific synthesis assessing the EOFM relevance, application evidence or evaluation requirements under uncertain reference data, scale mismatch and temporal dependence. Here, we develop a framework for determining when EOFMs support interpretable inference and identify a mismatch between EOFMs and ecohydrological requirements. Firstly, an observation-to-inference hierarchy shows that relevance depends on target-specific sensing pathways, spatial-temporal support and traceable uncertainty. Secondly, a meta-analysis shows that pretraining is dominated by reflected optical and active-microwave data, with sparse thermal coverage and no passive-microwave-emission sources. Thirdly, our synthesis of ecohydrological applications finds strongest support for spatial context, label-efficient adaptation and hybrid workflows. Evidence declines with inference depth; independent validation of fluxes, coupled dynamics, event trajectories, calibrated uncertainty and decision benefits remains sparse. Fourthly, our benchmark audit finds stronger coverage of fair adaptation and reproducibility in general EOFM suites, and of process targets, direct reference evidence and distribution shifts in ecohydrological evaluations; physical consistency and uncertainty remain weakly assessed. These findings motivate a process-aware framework aligning EOFM design and evaluation with the target variable, observation pathway and process timescale, supporting trustworthy monitoring and interpretation of coupled water, energy and carbon dynamics.

Figures & tables

Publication	Year	Journal	Primary scope and coverage	EO data coverage		Model integration		Evaluation
				EO modalities	Time series	Adaptation/ transfer	Domain knowledge	Comparative benchmarks	Process targets
Jiao et al. [ 15 ]	2023	JSTARS	RS FMs, downstream classification, localisation and understanding, and a brain-inspired framework	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$
Lu et al. [ 16 ]	2025	GRSM	Vision foundation-model architectures, pretraining data and methods, performance comparisons and research directions	$\checkmark$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\checkmark$
Huo et al. [ 17 ]	2025	RS	General and domain-specific vision, prompted and heterogeneous FMs for RS	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$
Xiao et al. [ 18 ]	2025	GRSM	Visual, vision–language and language FMs, EO datasets, technical advances and benchmarks	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\times$
Zhang et al. [ 19 ]	2025	GRSM	FMs for general geoscientific AI, including vision, language, multimodal and agentic systems	$\checkmark$	$\times$	$\checkmark$	$\checkmark$	$\times$	$\checkmark$

TABLE I: Representative reviews and surveys of FMs in RS, EO and geoscience. Binary scope is coded at the EO data, model-integration and evaluation levels. A check mark requires substantive coverage through a dedicated section or subsection, a comparison table or an original comparative experiment; a cross denotes incidental or absent coverage.

Fig. 1: A hierarchical framework linking biophysical organisation, inference distance and dominant observation systems in ecohydrology. The horizontal progression follows the scaling of coupled processes from leaves, through canopies and ecosystems, to the biosphere. The vertical tiers distinguish calibrated signals (T0), retrievals and indices (T1), estimated states and fluxes (T2), and mechanisms, events and decisions (T3); examples show representative quantities at different combinations of scale and inference distance. The upper band shows the transition from proximal sensors and flux towers to airborne and satellite EO, which could ultimately be embedded into EOFMs. Moving among tiers and scales introduces pathway-specific uncertainty from sensing, retrieval, forcing, partitioning, upscaling and model assumptions.

Fig. 2: Illustrative pathway from multisource EO pretraining to ecohydrological inference, using AlphaEarth Foundations as an example. During training, optical and thermal imagery, C- and L-band radar, climate and terrestrial-water-storage data, canopy and terrain measurements, and land-cover and semantic information serve as model inputs or prediction targets for learning a time-conditioned 64-dimensional embedding field; only the designated input sources are required to generate the embedding field at inference [ 61 ] . The general downstream operations comprise similarity search, classification, regression and change detection. The ecohydrological examples expand the regression branch according to Fig. 1 : T1 represents retrievals and indices, T2 estimated states and fluxes, and T3 mechanisms, events and decisions.

Fig. 3: Publication growth and diversification in EOFM research. Annual counts are assigned to the year of earliest public scholarly release and stacked by publication role. The green line indicates the percentage of model releases using at least two coded EO data pathways; reviews/perspectives and benchmarks/evaluations are excluded from this denominator. The 2026 record is partial and includes releases available through 31 July.

Fig. 4: Sampling distance and temporal design of shortlisted EOFMs. The 49 plotted model/version releases are positioned by earliest public scholarly release month and one pre-training sampling-distance value. A single reported ground sampling distance is used directly; when only a numeric interval was available, its geometric midpoint provides a transparent plotting proxy rather than an empirical median. Colour identifies the retained observation-pathway group, marker shape identifies temporal support, and marker area represents reported parameters using a base-10 logarithmic transformation.

First release a	Model name	Backbone and reported parameters b	Pre-training inputs and pathway c	Spatial resolution (m) d	Temporal support e	Ecohydrological relevance f , adaptation g , outputs h
Mar. 2021	SeCo [ 73 ]	ResNet-50; 23.5 M.	200,000 global Sentinel-2 seasonal views. [O]	10–60	PS	LP/FT $\rightarrow$ Cls., Seg. I : Seasonal vegetation and change context; no direct state or flux test.
Jul. 2022	RingMo [ 74 ]	Swin/ViT; 87.8 M.	Two million optical and SAR images. [M]	0.3–30	SN	FT $\rightarrow$ Cls., Det., Seg., CD. I : Optical–SAR supports flood and surface mapping; no continuous-state or flux test.
Jul. 2022	SatMAE [ 36 ]	ViT-L; 307 M.	fMoW-Sentinel 13-band imagery and short temporal stacks. [O]	10–60	PS	LP/FT $\rightarrow$ Cls., Seg. I : Crop, phenology and flood context; short sequences do not establish process memory.
Nov. 2022	RSP [ 75 ]	ViT-B/ViTAEv2 with RVSA; 100 M.	MillionAID optical imagery. [O]	0.5–153	SN	FT $\rightarrow$ Cls., Det., Seg. I : Multiscale landscape stratification; snapshot proxy evaluation provides no process test.
Nov. 2022	Satlas Pretrain [ 76 ]	SatlasNet/Swin; 88 M.	NAIP and Sentinel-2 imagery with multisource label products. [M]	1–10	SN	FT $\rightarrow$ Cls., Det., Seg., CD. I : Supplies land-surface context, but no direct ecohydrological state or flux was evaluated.
Dec. 2022	Scale-MAE [ 77 ]	ViT-L; 322.9 M.	fMoW RGB imagery and resolution metadata. [O]	n/r	SN	LP/FT $\rightarrow$ Cls., Seg. I : Scale-aware transfer may reduce mapping mismatch; temporal and process-sensitive measurements are absent.

TABLE II: Comparison of the 60 eligible EOFM model/version releases by model design, task transfer and directness of reported ecohydrological relevance. Rows are ordered by the month of earliest verified public release and then alphabetically by model name within each month.

Fig. 5: Structure of the 71 eligible primary EOFM application records. (a) Evidence density by model and application family. Circle area and labels show record counts, while colour indicates whether refereed works are at least as numerous as preprints within each model–application combination. (b) Numbers of refereed journal articles, refereed conference papers and preprints by application family. A work with multi-model records contribute to each applicable model group in (a); therefore, the sum of record counts within each row of (a) may exceed the corresponding application-family count in (b).

Fig. 6: Spatial correspondence between selected AlphaEarth embedding dimensions and environmental products across CONUS. Panels (a) and (b) compare AlphaEarth dimension A40 with daytime LST, and panels (c) and (d) compare dimension A48 with EVI. Mapped values are expressed as percentile ranks from 0 to 100. Spearman correlation ( $\rho$ ), Random Forest $R^{2}$ and Transformer $R^{2}$ quantify the corresponding relationships. The paired maps demonstrate geographically coherent information within the representation and agreement with the selected products; independent recovery of the underlying environmental quantities requires separate reference observations. Reprinted from [ 121 ] under a CC BY 4.0 License.

Fig. 7: Qualitative comparison of forest above-ground biomass density (AGBD) estimates for representative EnMAP hyperspectral image patches from North America (top two rows) and South America (bottom two rows). Within each regional block, the first column shows an EnMAP RGB composite overlaid with GEDI-derived reference AGBD values at lidar-shot locations and an infrared false-colour composite. The remaining columns present spatial predictions and reference–prediction scatterplots for the supervised U-Net baseline and four geospatial FMs evaluated with fine-tuned encoders: Spectral-ResNet-50 (Spec. RN50), Spectral-ViT-B, DOFA ViT-B and Panopticon. Dashed lines indicate 1:1 agreement, and AGBD is expressed in Mg ha -1 . The GEDI-derived reference values are retrieved biomass-density estimates; independent field measurements are absent from this qualitative comparison. Reprinted from [ 125 ] under a CC BY 4.0 License.

Fig. 8: Retrospective crop-stress early warning during extreme events. (a) Median alert lead time for a satellite-only Transformer, a late-fusion model and AgriFM. (b) Illustrative maize drought case linking vegetation condition and predicted risk with rainfall deficit and vapour-pressure deficit. (c) Illustrative wheat heat case linking thermal stress and predicted risk with temperature anomaly during the vegetative, flowering and grain-filling stages. Reprinted from [ 145 ] under a CC BY 4.0 License.

Fig. 9: Direction of reported results relative to meaningful baselines across the 71 eligible primary EOFM application records, grouped by application family. Stacked bars show the within-family percentage of records classified as positive, neutral (mixed or conditional), or null/negative; labels give percentages rounded to whole numbers. The categories provide a descriptive record-level synthesis across heterogeneous targets, metrics, baseline types, reference data and validation designs. They do not represent pooled effect sizes or model rankings.

Fig. 10: Mean study-level benchmark coverage across eight evaluation dimensions for general EOFM suites (

n=10

) and ecohydrological evaluations (

n=10

). Ordinal scores of 0, 1 and 2 denote absent, partial or indirect, and explicit coverage, respectively (see Tables III and IV ), and are rescaled to 0–100%. The dashed ring represents complete coverage across all dimensions.

Benchmark	Fair adaptation	Process target	Reference directness	Reproduc- ibility	Temporal dynamics	Transfer/ shift	Uncertainty/ reliability
Cryo-Bench [ 158 ]	2	0	0	2	0	1	0
EarthShift [ 170 ]	2	0	0	2	1	2	0
GEO-Bench [ 166 ]	2	0	0	2	0	0	0
GEO-Bench-2 [ 169 ]	2	1	0	2	1	0	1
Landsat-Bench [ 172 ]	2	0	0	2	0	0	0
MMEarth-Bench [ 34 ]	2	2	1	2	0	2	0

TABLE III: Record-level benchmark coverage for general EOFM suites. Benchmarks and coverage dimensions are ordered alphabetically.

Inference tier	Ecohydrological evaluations	Fair adaptation	Physical checks	Process target	Reference directness	Reproduc- ibility	Temporal dynamics	Transfer/ shift	Uncertainty/ reliability
T1	Hydrologic Mini-JEPAs [ 72 ]	2	1	2	1	0	1	2	0
	SM comparison [ 123 ]	2	1	2	2	0	1	2	0
T2	Global biomass [ 129 ]	1	0	2	1	1	1	2	0
	Hydrological generalisation [ 66 ]	2	1	2	2	1	2	2	0
	Ungauged basin prediction [ 139 ]	2	1	2	2	2	2	2	0
T3	Agricultural reliability [ 147 ]	1	0	2	2	1	2	2	1

TABLE IV: Benchmark coverage for ecohydrological evaluations, using the dimensions and scores defined in Table III . Records are grouped by inference tier and ordered alphabetically within tiers.

Dimension	Coverage (%) General / Ecohydrol.	Current benchmark pattern and evidence	Future priority and demonstrated examples
Fair adaptation	90 / 85	Shared adaptation protocols and controlled comparisons are established in general suites and selected ecohydrological evaluations [ 166 , 168 , 139 ] .	Match encoder access, inputs, downstream heads, label and optimisation budgets, and repeated-run reporting. Examples include shared-protocol and factorial comparisons [ 166 , 139 ] .
Physical checks	0 / 25	Partial checks address SM plausibility, hydrological regimes and sensor–variable coherence [ 123 , 66 , 72 ] .	Evaluate physical bounds, water–energy–carbon closure, partition consistency, cross-variable coherence, perturbation responses and regime-conditioned residuals. Examples demonstrate selected components of these tests [ 123 , 66 , 72 ] .
Process target	20 / 90	Continuous soil, biomass, streamflow, crop-stress and yield targets extend evaluation across T1–T3 [ 34 , 129 , 145 ] .	Define the physical quantity, units, inference tier and spatial–temporal support; distinguish signals, retrievals, states, fluxes, stocks, events and decisions. Examples span continuous environmental and event targets [ 34 , 129 , 145 ] .
Reference directness	5 / 60	General suites primarily use product and mapped references [ 166 , 34 , 168 ] . Ecohydrological studies also use station and gauge observations [ 123 , 66 , 139 ] .	Pair product agreement with matched in situ , field, flux-tower, gauge or inventory observations, and report observation and representation errors separately. Examples use station and gauge references [ 123 , 139 ] .
Reproducibility	85 / 40	Open benchmark assets, fixed splits and executable protocols are common in general suites and present in selected ecohydrological evaluations [ 166 , 168 , 139 ] .	Release versioned data, splits, preprocessing, code, weights or embeddings, hyperparameters, compute requirements and licences. Examples provide reusable benchmark or evaluation resources [ 166 , 168 , 139 ] .
Temporal dynamics	20 / 70	Current tests examine paired temporal shifts, temporal generalisation and event development [ 170 , 66 , 145 ] .	Test seasonal phase, event onset, thresholds, antecedent memory, lag and recovery at process-appropriate cadence. Examples evaluate temporal generalisation and event evolution [ 66 , 145 ] .

TABLE V: Priorities for future EOFM benchmarks in ecohydrology.

Geospatial Foundation Models

Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.

See all 119 papers

Jul 8, 2026

Scalable and Trustworthy Earth Observation Foundation Models

Foundation models (FMs) have transformed machine learning from isolated task-specific model development toward general-purpose models pretrained on broad data and adapted to multiple downstream tasks. Earth observation (EO) is an important domain for this paradigm because satellite and airborne archives are large, high-revisit, and increasingly multimodal, while reliable field labels are often sparse. Remote sensing foundation models (RSFMs) cannot be transferred reliably/optimally without domain-specific adaptation. This is because EO data are governed by measurement physics and operational decision constraints. This chapter reviews the design principles arising from these domain-specific constraints. It first defines the FMs paradigm in remote sensing (RS), then synthesizes the current model landscape, pretraining objectives, architecture designs, downstream adaptation and trustworthiness requirements. The chapter also incorporates recent benchmark evidence showing that no single geospatial foundation model is universally best and that inconsistent evaluation remains a major issue to fair comparison and reliable deployment. In addition, two brief environmental monitoring case studies; physics-informed spectral targeted masking for harmful algal bloom prediction and reinforcement learning for adaptive environmental monitoring station selection to illustrate the FMs domain-guided principles in practice. This chapter posits that next-generation RSFMs should be evaluated not only by benchmark accuracy, but also by modality-aware transfer and physically plausible representations for trustworthy EO decisions.

Syed Usama Imtiaz, Mitra Nasr Azadani, Nasrin Alamdari

Department of Civil and Environmental Engineering, Florida State University

Geospatial Foundation Models Vision Foundation Model Adaptation Remote Sensing Remote Sensing Benchmarks

Ask PDF

Sep 26, 2026

Reuse or Relearn? A Spectral View of Earth Observation Foundation Models

Foundation models are rarely used as generic, frozen feature extractors; instead, they are fine-tuned for the target downstream application. This practice is particularly prevalent in Earth observation (EO), and it raises a question that downstream accuracy alone cannot answer: does fine-tuning reuse the pretrained representation, or does it relearn a new one? We study this with spectral diagnostics that compare a model before and after adaptation, quantifying how well its dominant singular subspaces are preserved, how broadly the weight update is distributed, and how large it is. Using natural image models such as CLIP and DINO as a reference, we find that, under the evaluated fine-tuning settings, EO models undergo far larger, higher-rank updates and retain much less of their pretrained structure, so their downstream performance is often obtained with substantial changes to the pretrained weight structure. The diagnostics further provide insight into how cheaply a model can be adapted: where the pretrained subspaces are preserved, adapting a small fraction of the parameters can match full fine-tuning, and where they are not, it can fall behind. More broadly, foundation models, and EO foundation models in particular, should be assessed not only by benchmark accuracy, but also by how reusable their pretrained representation is.

Mehmet Ozgur Turkoglu, Valerio Marsocci, Dominik J. Mühlematter +3

Agroscope, EO of Agroecosystems · ESA, Φ-lab · ETH Zurich

Fine-Tuning Geospatial Foundation Models Representation Drift Parameter-Efficient Fine-Tuning

Ask PDF

Jun 25, 2026

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting

Earth Observation (EO) forecasting aims to predict future Earth surface dynamics from satellite observations under changing meteorological conditions. In this paper, we view this task as a partially observed, weather-driven world modeling problem, in which weather acts as a conditioning signal, while forecasting remains uncertain due to sparse observations and unobserved land-surface states. However, existing methods do not fully capture this setting: deterministic models collapse uncertainty into a single future prediction, while diffusion-based methods typically treat weather variables as undifferentiated conditioning signals, and existing benchmarks focus mainly on reconstruction accuracy rather than whether forecasts respond correctly to changed weather forcing.We introduce EO-WM, a video diffusion transformer for multispectral EO forecasting. EO-WM incorporates a physically informed conditioning framework that represents meteorological forcing through a climatological baseline, weather anomalies, and cumulative physical stress signals. Specifically, it separates baseline and anomaly through distinct conditioning pathways, and accumulates anomalous forcing over time to capture sustained heat and drought stress. To evaluate weather-response behavior beyond standard metrics, we introduce two diagnostic benchmarks: an Extreme Summer Benchmark for severity-aware prediction of vegetation degradation under extreme weather, and a Seasonal Matched-Pair Benchmark for testing response fidelity under changed weather forcing. Experiments show that EO-WM reduces the error in predicted Normalized Difference Vegetation Index (NDVI) decline amplitude by a relative 5.63% and improves directional hit rate by a relative 7.80%, while remaining competitive on standard pixel-level metrics. The benchmarks and model will be made open-source at https://github.com/Luo-Z13/EO-WM.

Junwei Luo, Shuai Yuan, Zhenya Yang +3

1The University of Hong Kong · 2Wuhan University

World Model Learning Forecasting Benchmarks Probabilistic Forecasting Remote Sensing Spatiotemporal Forecasting Diffusion Models

Ask PDF

Abstract

Geospatial Foundation Models

Explore similar work

Scalable and Trustworthy Earth Observation Foundation Models

Reuse or Relearn? A Spectral View of Earth Observation Foundation Models

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting