Location Is All You Need: Continuous Spatiotemporal Neural Representations of Earth Observation Data

Mojgan Madadikhaljan; Jonathan Prexl; Isabelle Wittmann; Conrad M Albrecht; Michael Schmitt

cs.CVApr 8, 2026

Location Is All You Need: Continuous Spatiotemporal Neural Representations of Earth Observation Data

Authors: Mojgan Madadikhaljan, Jonathan Prexl, Isabelle Wittmann, Conrad M Albrecht, Michael Schmitt

Organizations: University of the Bundeswehr Munich · IBM Research – Europe · Columbia University · German Aerospace Center (DLR)

Abstract

In this work, we present LIANet (Location Is All You Need Network), a coordinate-based neural representation that models multi-temporal spaceborne Earth observation (EO) data for a given region of interest as a continuous spatiotemporal neural field. Given only spatial and temporal coordinates, LIANet reconstructs the corresponding satellite imagery. Once pretrained, this neural representation can be adapted to various EO downstream tasks, such as semantic segmentation or pixel-wise regression, importantly, without requiring access to the original satellite data. LIANet intends to serve as a user-friendly alternative to Geospatial Foundation Models (GFMs) by eliminating the overhead of data access and preprocessing for end-users and enabling fine-tuning solely based on labels. We demonstrate the pretraining of LIANet across target areas of varying sizes and show that fine-tuning it for downstream tasks achieves competitive performance compared to training from scratch or using established GFMs. The source code and datasets are publicly available at https://github.com/mojganmadadi/LIANet/tree/v1.0.1.

Figures & tables

Figure 1: Comparison of model performance as a function of the number of tunable parameters for two pixel-wise classification tasks. A classical UNet architecture (trained from scratch), as well as three commonly applied FM (with varying parameter count, depending on fine-tuning scheme), serve as a baseline. Our approach, LIANet , achieves a high performance with minimal tunable parameters.

	GFM	Precomputed Embeddings	Location Encoders	LIANet (ours)
No raw imagery
at inference	✗	✓	✓	✓
Continuous spatio-
temporal field	✗	✗	Spatial only	✓
Image
reconstruction	✓	✗	✗	✓

Table 1: Conceptual comparison of GFM , precomputed embeddings, location encoders, and LIANet on key modeling axes.

Figure 2: Visualization of the LIANet workflow. Pretraining: The specified $(x,y)$ coordinates (model input) are used to query learned representations from a hash table by interpolating table entries for the neighboring pixels (blue dots) for the given input position (red dots) for all different resulution grids $G_{i}$ (in this illustration $N_{\text{grids}}=3$ ) and adding them to the corresponding temporal encoding vector. This process is applied iteratively to all pixels of the target output image and is implemented in a computationally efficient manner [ 38 ] , leading to a scene representation $\hat{\textbf{g}}_{x,y,t}$ . A decoder network $D$ reconstructs the underlying satellite image $\textbf{I}=D(\hat{\textbf{g}}_{x,y,t})$ . Fine tuning: In this stage, the final layers of $D$ are adapted to generate labels for an arbitrary downstream task. During fine-tuning, the learned table values and the first few decoder layers are frozen, and no input images are required during this stage.

Figure 3: Left plot shows the reconstruction performance of LIANet-Base and LIANet-Large on $\mathcal{A}_{0}$ , $\mathcal{A}_{+}$ , and $\mathcal{A}_{++}$ measured by Structural Similarity (SSIM) of the reconstructed images as a function of pretraining step. The images on the right show a sample ground truth image (first row), reconstruction images of LIANet-Base (second row) and LIANet-Large (third row) from two different seasons. The enhanced reconstruction capability of LIANet-Large compared to LIANet-Base becomes evident in regions with high-frequency features, such as the highlighted areas $A$ and $B$ . Note that the images produced on the right are the result of four adjacent and non-overlapping $128\times 128$ patches stitched together horizontally, where, due to effective continuous reconstruction, no patching effect appears, enabling larger scene reconstructions without any undesired artifacts.

Task	Model / Setting	# Tunable Params (M)	IoU	ACC	F1
Dynamic World	UNet / Micro UNet	17.3 / 0.49	0.75 / 0.66	0.82 / 0.73	0.84 / 0.74
	TerraMind-base (Full / Frozen / Embedding)	102 / 15.5 / 0.39	0.72 / 0.71 / 0.66	0.78 / 0.77 / 0.73	0.81 / 0.80 / 0.73
	Prithvi v2-300 (Full / Frozen / Embedding)	324 / 20.3 / 0.46	0.71 / 0.67 / 0.60	0.77 / 0.74 / 0.68	0.80 / 0.76 / 0.69
	DOFA-Large (Full / Frozen / Embedding)	357 / 20.3 / 0.46	0.70 / 0.64 / 0.47	0.76 / 0.71 / 0.54	0.79 / 0.74 / 0.57
	LIANet-Base	0.5	0.72	0.82	0.81
	LIANet-Large	0.5	0.72	0.81	0.81

Table 2: (Top) Pixel-wise classification task performance measured with Intersection-over-Union (IoU), Accuracy (Acc), and F1-score (all calculated with macro averaging) along with the tunable parameter counts for the area of interest

\mathcal{A}_{0}

. (Bottom) Regression task performance (Mean Absolute Error (MAE) and Mean Squared Error (MSE)) along with the tunable parameter counts for the area of interest

\mathcal{A}_{0}

. The from-scratch trainings and FM are evaluated in different configurations with a varying number of tunable parameters. For all metrics reported, the corresponding top three performances are printed bold . Due to the limited number of comparisons, we are not marking any models for the task of building footprint segmentation.

Figure 4: Performance across five downstream tasks as a function of the encoded area of interest $\mathcal{A}$ and as a function of the two model sizes. For regression tasks, we report $1-\text{MAE}$ to highlight the overall trend more clearly. Note that the regression task of building footprints reported in yellow is the prediction of the percentage of the buildings present in each pixel.

Task	Model / Setting	# Tunable Params (M)	IoU	ACC	F1
PASTIS	UNet / Micro UNet	17.3 / 0.49	0.18 / 0.13	0.26 / 0.19	0.26 / 0.19
	TerraMind-base (Full / Frozen / Embedding)	102 / 15.5 / 0.39	0.26 / 0.23 / 0.24	0.35 / 0.30 / 0.31	0.33 / 0.32 / 0.32
	Prithvi v2-300 (Full / Frozen / Embedding)	324 / 20.3 / 0.46	0.23 / 0.21 / 0.19	0.29 / 0.28 / 0.25	0.32 / 0.30 / 0.27
	DOFA-Large (Full / Frozen / Embedding)	357 / 20.3 / 0.46	0.19 / 0.17 / 0.14	0.23 / 0.24 / 0.19	0.28 / 0.25 / 0.19
	LIANet-Modified	0.87	0.34	0.46	0.46
Burn Scars	UNet / Micro UNet	17.3 / 0.49	0.45 / 0.47	0.65 / 0.62	0.62 / 0.62

Table 3: The pixel-wise classification performance evaluation of two datasets from the PANGAEA benchmark, measured with Intersection-over-Union (IoU), Accuracy (Acc), and F1-score (all calculated with macro averaging), along with the tunable parameter counts. For all metrics reported, the corresponding top three performances are printed bold .

Task	Model / Setting	# Tunable Params (M)	IoU	ACC	F1
Dynamic World	UNet / Micro UNet	17.3 / 0.49	0.75 / 0.66	0.82 / 0.76	0.84 / 0.76
	TerraMind-base (Full / Frozen / Embedding)	102 / 15.5 / 0.39	0.70 / 0.70 / 0.65	0.78 / 0.77 / 0.72	0.80 / 0.80 / 0.72
	Prithvi v2-300 (Full / Frozen / Embedding)	324 / 20.3 / 0.46	0.69 / 0.65 / 0.58	0.77 / 0.73 / 0.68	0.79 / 0.75 / 0.67
	DOFA-Large (Full / Frozen / Embedding)	357 / 20.3 / 0.46	0.68 / 0.62 / 0.46	0.75 / 0.71 / 0.54	0.78 / 0.72 / 0.56
	LIANet-Base	0.5	0.69	0.80	0.79
	LIANet-Large	0.5	0.70	0.80	0.80

Table 4: (Top) Pixel-wise classification task performance measured with Intersection-over-Union (IoU), Accuracy (Acc), and F1-score (all calculated with macro averaging) along with the tunable parameter counts for the area of interest

\mathcal{A}_{+}

. (Bottom) Regression task performance (Mean Absolute Error (MAE) and Mean Squared Error (MSE)) along with the tunable parameter counts for the area of interest

\mathcal{A}_{+}

Model	Train Area ( $\text{\,}{\mathrm{km}}^{2}$ )	IoU	ACC	F1
LIANet-Base	80	0.62	0.70	0.72
	160	0.66	0.75	0.75
	320	0.71	0.79	0.80
	500	0.72	0.82	0.81
	640	0.73	0.82	0.82
LIANet-Large	80	0.63	0.71	0.73

Table 5: Comparison of the pixel-wise land cover classification performance measured with Intersection-over-Union (IoU), Accuracy (Acc), and F1-score (all calculated with macro averaging) with respect to the different training area sizes.

Task	Model / Setting	# Tunable Params (M)	IoU	ACC	F1
Dynamic World	UNet / Micro UNet	17.3 / 0.49	0.76 / 0.68	0.83 / 0.76	0.85 / 0.78
	TerraMind-base (Full / Frozen / Embedding)	102 / 15.5 / 0.39	0.73 / 0.72 / 0.63	0.79 / 0.78 / 0.70	0.82 / 0.81 / 0.71
	Prithvi v2-300 (Full / Frozen / Embedding)	324 / 20.3 / 0.46	0.71 / 0.68 / 0.61	0.78 / 0.75 / 0.70	0.80 / 0.77 / 0.70
	DOFA-Large (Full / Frozen / Embedding)	357 / 20.3 / 0.46	0.67 / 0.64 / 0.42	0.73 / 0.72 / 0.51	0.77 / 0.74 / 0.53
	LIANet-Base	0.5	0.70	0.78	0.79
	LIANet-Large	0.5	0.70	0.78	0.80

Table 6: (Top) Pixel-wise classification task performance measured with Intersection-over-Union (IoU), Accuracy (Acc), and F1-score (all calculated with macro averaging) along with the tunable parameter counts for the area of interest

\mathcal{A}_{++}

. (Bottom) Regression task performance (Mean Absolute Error (MAE) and Mean Squared Error (MSE)) along with the tunable parameter counts for the area of interest

\mathcal{A}_{++}

Figure 5: Rate-Distortion analysis for LIANet’s Sentinel-2 image generation. In summary, our models perform decent (purple dash-dotted line) to excellent (red dashed line) with a distribution of PSNR-values derived from 100 randomly sampled spatial patches of size 128x128 pixels over the areas

\mathcal{A}_{0}

\mathcal{A}_{+}

, and

\mathcal{A}_{++}

. The label LIANet-Large

{}_{\mathcal{A}_{++}}

for instance, refers to the LIANet-Large model that is pretrained on area

\mathcal{A}_{++}

. Blue and orange labels correspond to PSNR statistics for all Sentinel-2 bands and RGB channels, respectively. Dots indicate minimum and maximum values, bars cover the

1.5\sigma

interval, the cross marks the mean, and the open circle indicates the median. The bold solid blue and orange lines provide a logarithmic interpolation PSNR(bpp)

\penalty\ \sim P_{3}\left[\log(1+\text{bpp})\right]

of the median PSNR (RGB) values with

Q_{\text{\it LIANet}}=4

where we fixed PSNR(bpp=0)=0.

P_{3}[z]

denotes a polynomial of third degree in

z

. A vertical dashed line (gray) indicates the bpp of the raw Sentinel-2 data in JPEG2000 format as a reference.

Panel of Figure 6 — Figure 6: Two examples of reconstruction quality of LIANet-Base and LIANet-Large on $\mathcal{A}_{0}$ area. It can be seen that LIANet-Large has an improved performance in reconstructing the high-frequency details, including buildings within the city. It can also be seen that cloudy areas are reconstructed as they are present in the Sentinel-2 image.

Panel of Figure 7 — Figure 7: Three example visualization of LIANet-Base and LIANet-Large reconstruction and prediction performances on downstream applications. The landcover labels have seasonal predictions, whereas the other tasks have a single label for all timestamps. The improvement in the prediction of LIANet-Large can clearly be seen in the tasks concerning the segmentation and density estimation of building footprints.

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 126 papers

GeoSANE: Learning Geospatial Representations from Models, Not Data
Mar 24, 2026Joelle Hanna, Damian Falk, Stella X. Yu +1Geospatial Representation Learning Geospatial Foundation Models
University of St.Gallen · University of Michigan and UC Berkeley · ESA Φ-Lab
MajorTOM:Elliot Pretrain, A Multi-Sensor, Multi-Temporal Pretraining Dataset for Earth Observation
Oct 13, 2026Carmen Oliver Huidobro, Mikolaj Czerkawski, Kevin Monsalvez-Pozo +3Remote Sensing Geospatial Foundation Models
Image and Signal Processing (ISP) Group, Universitat de València, Valencia, Spain · Asterisk Labs, London, United Kingdom
Recoverable Geographic Location Information in Earth-Observation Embeddings
Sep 24, 2026Peiwen Zhang, Kristie Hu, Jovana Knezevic +2Geospatial Foundation Models Geospatial Representation Learning
University of Waterloo · University of Cambridge · Aalto University

Mar 24, 2026J/K move · Enter open · S save

GeoSANE: Learning Geospatial Representations from Models, Not Data

Joelle Hanna, Damian Falk, Stella X. Yu, Damian Borth

University of St.Gallen · University of Michigan and UC Berkeley · ESA Φ-Lab

Geospatial Representation Learning Geospatial Foundation Models

Open paper AskPDF arXiv

Recent advances in remote sensing have led to an increase in the number of available foundation models; each trained on different modalities, datasets, and objectives, yet capturing only part of the vast geospatial knowledge landscape. While these models show strong results within their respective domains, their capabilities remain complementary rather than unified. Therefore, instead of choosing one model over another, we aim to combine their strengths into a single shared representation. We introduce GeoSANE, a geospatial model foundry that learns a unified neural representation from the weights of existing foundation models and task-specific models, able to generate novel neural networks weights on-demand. Given a target architecture, GeoSANE generates weights ready for finetuning for classification, segmentation, and detection tasks across multiple modalities. Models generated by GeoSANE consistently outperform their counterparts trained from scratch, match or surpass state-of-the-art remote sensing foundation models, and outperform models obtained through pruning or knowledge distillation when generating lightweight networks. Evaluations across ten diverse datasets and on GEO-Bench confirm its strong generalization capabilities. By shifting from pre-training to weight generation, GeoSANE introduces a new framework for unifying and transferring geospatial knowledge across models and tasks. Code is available at hsg-aiml.github.io/GeoSANE/.

Abstract

Geospatial Representation Learning

Explore similar work

GeoSANE: Learning Geospatial Representations from Models, Not Data

MajorTOM:Elliot Pretrain, A Multi-Sensor, Multi-Temporal Pretraining Dataset for Earth Observation

Recoverable Geographic Location Information in Earth-Observation Embeddings