Temporal Sensitivity Analysis of Tessera Embeddings

Julia Guerrero-Viu; Alex López-Cifuentes; Ignacio Pérez-Villar; Fabio Pacifici

cs.CVAug 27, 2026

Temporal Sensitivity Analysis of Tessera Embeddings

Authors: Julia Guerrero-Viu, Alex López-Cifuentes, Ignacio Pérez-Villar, Fabio Pacifici

Abstract

Many Earth Observation applications need land-use/land-cover maps that are both precise and frequently updated, yet the strongest Earth Observation foundation models build their embeddings from a full year of observations. We present a controlled study of the temporal sensitivity of Tessera, one of these leading foundation models, for land-use/land-cover mapping. Keeping the encoder frozen, we recompute its embeddings over varying observation windows, from a full year down to a single day. We use them as inputs to a linear probe and a UNet segmentation head, benchmarking both of them against from-scratch networks on LUCAS, DynamicEarthNet, and PASTIS-R datasets. We show that the value of the embeddings is task-dependent. Where classes are separated by phenology, as for the crop types of PASTIS-R, they reach a mean Intersection-over-Union of $58.3$ , about $46\%$ above the best from-scratch model. Where classes are temporally stable (e.g., forests in DynamicEarthNet and LUCAS), embedding-based and from-scratch models match only under full supervision. On both datasets, Tessera embeddings remain markedly more label-efficient. Degradation under shorter temporal windows is gradual and class-dependent. Contracting the window from one year to one month costs $39\%$ of the segmentation accuracy on PASTIS-R but only $5\%$ on DynamicEarthNet. Single-day embeddings still classify land cover in LUCAS at $3.4$ times the chance level. Our study shows that temporal coverage is therefore a tunable cost rather than a fixed prerequisite, opening regimes such as near-real-time mapping and faster land-use/land-cover refresh cycles.

Figures & tables

Fig. 1: Temporal Sensitivity Evaluation Framework: (Top) For each dataset, multi-temporal Sentinel-1/-2 observations are extracted over observation windows of decreasing length defined relative to a dataset-specific reference date. (Bottom) Given all observations in a temporal window, we encode them with the frozen Tessera model into a dense per-pixel embedding map. For each dataset and temporal window, we train a linear classification probe and a UNet segmentation head to evaluate the quality of the Tessera embeddings at that temporal resolution.

		1M				3M				6M				1Y
	Method	Acc	B.Acc	F1	mIoU	Acc	B.Acc	F1	mIoU	Acc	B.Acc	F1	mIoU	Acc	B.Acc	F1	mIoU
DEN	UNet-RGB	62.9 $\pm 0.6$	45.7 $\pm 0.7$	43.5 $\pm 0.5$	30.9 $\pm 0.3$	61.5 $\pm 1.6$	46.4 $\pm 1.4$	44.1 $\pm 0.2$	31.3 $\pm 0.1$	64.1 $\pm 0.5$	48.7 $\pm 1.1$	46.5 $\pm 0.5$	33.8 $\pm 0.3$	65.9 $\pm 0.9$	48.9 $\pm 0.9$	47.6 $\pm 0.5$	35.1 $\pm 0.3$
	UNet-RGB tm	63.6 $\pm 0.9$	45.3 $\pm 1.3$	43.9 $\pm 0.8$	31.2 $\pm 0.6$	63.2 $\pm 0.9$	46.2 $\pm 1.0$	44.7 $\pm 0.3$	31.8 $\pm 0.3$	64.3 $\pm 0.7$	49.3 $\pm 1.3$	46.8 $\pm 0.8$	34.0 $\pm 0.6$	66.5 $\pm 0.4$	49.8 $\pm 0.5$	48.4 $\pm 0.4$	35.8 $\pm 0.2$
	UNet-S2	72.0 $\pm 0.4$	58.6 $\pm 0.7$	55.0 $\pm 0.6$	42.4 $\pm 0.3$	71.7 $\pm 0.4$	57.9 $\pm 1.1$	55.6 $\pm 0.6$	42.6 $\pm 0.4$	72.0 $\pm 0.6$	58.8 $\pm 0.5$	57.0 $\pm 0.6$	43.9 $\pm 0.4$	71.3 $\pm 1.0$	58.5 $\pm 1.0$	56.9 $\pm 0.7$	43.5 $\pm 0.4$
	UNet-S1S2	73.0 $\pm 0.5$	53.7 $\pm 1.2$	52.9 $\pm 0.7$	41.2 $\pm 0.4$	72.6 $\pm 0.5$	56.7 $\pm 0.7$	54.6 $\pm 0.4$	42.0 $\pm 0.3$	73.4 $\pm 0.8$	57.6 $\pm 1.1$	56.9 $\pm 0.8$	44.2 $\pm 0.6$	73.0 $\pm 0.8$	57.8 $\pm 0.8$	56.7 $\pm 0.6$	43.9 $\pm 0.3$
	TESSERA-LP	68.3	50.3	49.8	38.1	69.5	52.5	51.9	39.5	71.4	53.6	54.1	41.4	71.8 $\pm 0.0$	54.9	56.4	42.5

TABLE I: Semantic segmentation results (%) on the DEN and PASTIS-R (19 classes, background included) test sets across temporal windows. TESSERA-UNet and the from-scratch UNets are reported as mean

\pm

std over 5 runs; TESSERA-LP corresponds to a single run. Best result per column and dataset in green , second best in blue .

Fig. 2: Per-class temporal sensitivity of the linear probe (TESSERA-LP): per-class F1 normalized by its maximum across windows. DEN and LUCAS show all classes; for PASTIS-R (19 classes) we highlight the most and least temporally sensitive classes, ranked by the F1 retained at the shortest window relative to 1Y (

r=\mathrm{F1(1M)}/\mathrm{F1(1Y)}

), and show the remaining classes in gray.

Temporal resolution	Acc.	Bal. Acc.	Prec.	F1	F1 w
1 Day	43.0	36.8	29.2	25.8	49.2
1 Week	47.5	42.1	31.8	28.8	53.5
1 Month	52.3	44.6	33.9	31.7	58.3
3 Months	55.5	46.4	35.6	33.9	61.8
6 Months	58.8	49.8	37.3	35.8	64.8
1 Year	59.5	50.8	38.1	36.5	65.4

TABLE II: Classification performance (%) in LUCAS dataset as a function of the temporal resolution of the input. Best value per column in green , second best in blue . F1 w denotes the support-weighted F1 score; all other class averages are macro-averaged.

Model	Frozen	Finetuned
CROMA [ 34 ]	36.24	41.29
DOFA [ 14 ]	25.94	29.55
Prithvi [ 9 ]	34.09	39.52
RemoteCLIP [ 16 ]	20.17	21.03
SatlasNet [ 35 ]	21.02	31.38
Scale-MAE [ 36 ]	23.96	28.51

TABLE III: Semantic segmentation on PASTIS-R (mIoU, full labels). Foundation-model baselines are reported both frozen and finetuned, as in [ 11 ] ; AlphaEarth, Tessera and our models use frozen embeddings with a trained head/probe (single evaluation). Baseline numbers are taken from [ 11 ] . Best in green , second in blue .

Fig. 3: Label efficiency at the 1Y window. Test mIoU (%) as a function of the fraction of training labels (

1

–

100\%

) on PASTIS-R (left) and DEN (right). On PASTIS-R we additionally overlay the V1.1 embeddings decoded with TESSERA-UNet (dashed). V1.1 is evaluated on PASTIS-R only.

Config	Acc.	Bal. Acc.	F1-macro	mIoU
baseline	74.1	48.9	50.1	39.6
+ normalization	73.7	48.8	49.6	39.2
+ inverse_sqrt	73.2	55.7	54.4	42.8
+ 60 epochs (best)	72.8	56.5	55.2	43.1

TABLE IV: Ablation study on DEN 6M for TESSERA-UNet , regarding the incremental impact of input normalization, inverse-sqrt class weighting, and extended training (60 epochs) on segmentation metrics (%). The best configuration, selected by mIoU, is shown in bold.

Config	Acc.	Bal. Acc.	F1-macro	mIoU
baseline	81.5	69.5	70.7	57.2
+ normalization	81.6	69.9	71.1	57.7
+ inverse_sqrt	79.9	76.3	70.7	56.8
+ 60 epochs (best)	80.6	74.0	71.9	58.3

TABLE V: Ablation study on PASTIS-R 1Y for TESSERA-UNet , regarding the incremental impact of input normalization, inverse-sqrt class weighting, and extended training (60 epochs) on segmentation metrics (%). The best configuration, selected by mIoU, is shown in bold.

Embeddings	Acc	B.Acc	F1	mIoU
TESSERA-UNet (V1)	80.6 $\pm 0.1$	74.0 $\pm 0.4$	71.9 $\pm 0.3$	58.3 $\pm 0.4$
TESSERA-UNet (V1.1)	80.6 $\pm 0.1$	73.7 $\pm 0.4$	72.0 $\pm 0.4$	58.4 $\pm 0.3$
TESSERA-LP (V1)	76.2	67.4	62.7	48.5
TESSERA-LP (V1.1)	76.2	67.7	62.8	48.5

TABLE VI: TESSERA V1 vs. V1.1 embeddings on PASTIS-R (1Y; test %), decoded with the TESSERA-UNet head (mean

\pm

std over 5 seeds) and with a linear probe (TESSERA-LP, single run). The larger value per version pair and metric is in bold; for the TESSERA-UNet head the differences are within the run-to-run variability, and for the linear probe they are negligible in absolute terms (at most

0.3

points).

Window	V1	V1.1	$\Delta$	$\Delta\%$
1M	0.036	0.048	+0.012	+33.8%
3M	0.052	0.075	+0.023	+43.6%
6M	0.116	0.106	$-0.010$	$-8.8\%$
1Y	0.284	0.278	$-0.005$	$-1.9\%$

TABLE VII: Linear (LDA) class separability of the frozen embeddings on PASTIS-R (higher is better).

\Delta

and

\Delta\%

are the V1.1

-

V1 change; gains in green , losses in red .

Fig. 4: Qualitative results on the DEN test set (1Y window). Each row shows a test sample (RGB composite and ground truth) followed by the predictions of all compared models.

Fig. 5: Qualitative results on the PASTIS-R test set (1Y window). Each row shows a test sample (RGB composite and ground truth) followed by the predictions of all compared models.

Fig. 6: Qualitative evolution across temporal windows on the DEN test set. For each test sample (RGB composite and ground truth), the predictions of TESSERA-LP and TESSERA-UNet are shown at the 1M, 3M, 6M and 1Y windows.

Fig. 7: Qualitative evolution across temporal windows on the PASTIS-R test set. For each test sample (RGB composite and ground truth), the predictions of TESSERA-LP and TESSERA-UNet are shown at the 1M, 3M, 6M and 1Y windows.

Land Cover Classification

Land cover classification assigns satellite, aerial, and LiDAR observations to classes such as forests, crops, water, and built areas.

See all 35 papers

Earth Embeddings
Aug 4, 2026Adam J. Stewart, Heng Fang, Isaac A. Corley +1Geospatial Representation Learning Land Cover Classification
Chair of Data Science in Earth Observation, Technical University of Munich, Munich, Germany · Munich Center for Machine Learning, Munich, Germany · KTH Royal Institute of Technology, Stockholm, Sweden +1
Better Together: Evaluating the Complementarity of Earth Embedding Models
May 18, 2026Thijs L van der Plas, Jacob JW Bakermans, Vishal Nedungadi +3Geospatial Foundation Models Geospatial Representation Learning
Wageningen University, NL · University College London, UK · University of Bonn, Germany
Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland
Jun 18, 2026Htet Yamin Ko Ko, Clement AtzbergerRemote Sensing Image Segmentation Geospatial Foundation Models
International Space Science Institute, Bern, Switzerland · CYCLOPS MRV, United States

Aug 4, 2026J/K move · Enter open · S save

Earth Embeddings

Adam J. Stewart, Heng Fang, Isaac A. Corley, Xiao Xiang Zhu

Chair of Data Science in Earth Observation, Technical University of Munich, Munich, Germany · Munich Center for Machine Learning, Munich, Germany · KTH Royal Institute of Technology, Stockholm, Sweden+1

Geospatial Representation Learning Land Cover Classification Remote Sensing Earth Science

Open paper AskPDF arXiv

Earth observation is moving from foundation models that users must run themselves toward embedding products that package model feature outputs as reusable data without needing to download and process the imagery used to generate them. Earth embeddings are vectors that summarize locations, image patches, or pixels, letting users analyze compact features instead of repeatedly training or running large models on raw satellite imagery. This chapter explains the main types of Earth embeddings, from implicit location encoders to explicit patch and pixel products, and compares their coverage, resolution, dimensionality, storage cost, licenses, and reproducibility. We review their use in land cover and crop mapping, ecological and hazard modeling, socioeconomic prediction, and semantic search, with evidence on when embeddings improve on conventional features and when pooling, fusion, or spatial transfer limit performance. Two case studies show practical workflows for similarity search and land cover mapping. We close with guidance for choosing, evaluating, storing, compressing, and publishing embeddings, and with open problems in oceanic and atmospheric coverage, uncertainty, and benchmarking.

Abstract

Land Cover Classification

Explore similar work

Earth Embeddings

Better Together: Evaluating the Complementarity of Earth Embedding Models

Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland