GeoSANE: Learning Geospatial Representations from Models, Not Data

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

cs.CVMar 24, 2026

GeoSANE: Learning Geospatial Representations from Models, Not Data

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

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

Abstract

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/.

Figures & tables

Figure 1: Instead of pretraining geospatial (foundation) models from satellite data (top), GeoSANE proposes to leverage publicly available geospatial (foundation) models to train a shared latent representation encapsulating their knowledge to generate model weights tailored for specific downstream task (bottom).

Figure 2: Overview of our approach. (A) A heterogeneous collection of models, including ViTs, Swins, ResNets, UNets, and vision-language models, is gathered from HuggingFace. (B) A weight-space autoencoder is trained to reconstruct and embed these models into a shared latent representation. (C) From this latent space, GeoSANE can generate new models on demand for specific downstream tasks such as flood segmentation, object detection, or land-cover classification.

Figure 3: Our model collection retrieved from Hugging Face is diverse as seen in the distribution of model categories in our dataset.

Dataset	Backbone	from scratch	GeoSANE	$\Delta$
EuroSAT [ 22 ]	ViT-L	95.0	99.1	+4.1
RESISC-45 [ 5 ]	ViT-L	78.0	96.5	+18.5
fMoW [ 6 ]	ViT-L	35.7	58.9	+23.2
Sen12Flood [ 36 ]	ViT-L	80.4	85.2	+4.8
Cal. Wildfires [ 4 ]	ViT-L	88.7	94.9	+6.2
BigEarthNet [ 49 ]	ViT-L	69.8	88.7	+18.9

Table 1: Comparison between training from scratch and finetuning a GeoSANE-generated model across all benchmark datasets. We report accuracy for single-label classification, mAP for multi-label classification, mIoU for segmentation and [email protected] for object detection. Best results are in bold .

\Delta

indicates the absolute improvement over training from scratch.

Model	Backbone	Single-label					Multi-label	Segmentation			Object Det.
Model	Backbone	EuroSAT	RESISC45	fMoW	Sen12Flood	Wildfires	BigEarthNet	DFC2020	SpaceNet	Sen1Floods11	DIOR
SatMAE [ 7 ]	ViT-L	98.9	94.8	58.2	80.3	88.6	86.2	44.1	78.1	–	70.9
Scale-MAE [ 37 ]	ViT-L	99.1	95.7	–	82.6	90.8	87.9	–	78.9	74.1	73.8
RingMo [ 50 ]	Swin-B	–	95.7	–	–	–	–	–	–	–	75.9
CROMA [ 13 ]	ViT-L	99.4	–	59.0	83.4	93.3	88.3	49.8	–	90.9	–
GFM [ 32 ]	Swin-B	–	–	–	77.9	91.9	86.3	–	–	72.6	–

Table 2: Comparison with existing remote sensing foundation models on multiple remote sensing benchmarks. We report Overall Accuracy for single-label classification, mean Average Precision for multi-label classification, mean intersection over union for segmentation and mean Average Precision ([email protected]) for object detection. For GeoSANE, we report the mean

\pm

standard deviation over three independently generated models per prompt (Section 3.3 ); for all baselines, we report the values reported in the respective papers. Best results are in bold , second best are underlined .

\Delta

indicates the absolute improvement over the best baseline

	Dataset	Magnitude Pruning (MP)			Variational Dropout (VP)			Knowledge Distillation (KD)			GeoSANE	$\Delta$
		Initial Models			Initial Models			Teachers
		ViT-L (IN)	ScaleMAE	SatMAE	ViT-L (IN)	ScaleMAE	SatMAE	ViT-L (IN)	ScaleMAE	SatMAE
11M Params	RESISC-45	88.1	87.2	84.9	81.1	81.8	80.6	90.2	90.3	91.1	92.2	+1.1
	EuroSAT	97.7	95.4	96.3	95.5	97.0	94.7	97.9	94.0	94.7	98.7	+0.8
	fMoW	33.7	36.2	35.4	25.2	32.3	27.1	38.6	43.1	41.9	53.5	+10.4
	BigEarthNet	45.6	59.1	62.1	38.2	44.9	49.8	65.7	67.3	66.5	83.7	+16.4

Table 3: Comparison of GeoSANE with pruning and distillation baselines. For Magnitude Pruning and Variational Dropout , we prune pretrained Remote Sensing Foundation Models (RSFMs) (ScaleMAE [ 37 ] and SatMAE [ 7 ] ) and an ImageNet(IN)-pretrained ViT-L to obtain versions with approximately 11M non-zero parameters (ResNet-18) and 5M (MobileNetV2) non-zero parameters. For Knowledge Distillation , the same RSFMs and the ImageNet ViT-L act as teachers, and the student networks are ResNet-18 models with 11M parameters or MobileNetV2 with 3.5M parameters. GeoSANE directly generates models of the target architecture and size. Best results are in bold , second best are underlined .

\Delta

indicates the absolute improvement over the best baseline

Model	Single-label		Multi-label
Model	EuroSAT	RESISC45	BigEarthNet
SatMAE [ 7 ]	98.9	94.8	86.2
Scale-MAE [ 37 ]	99.1	95.7	87.9
Merged Model	96.4	86.1	69.0
GeoSANE	99.1	96.5	88.7

Table 4: Comparison of a model generated by GeoSANE against a merged model obtained by combining RSFMs using DARE [ 63 ] , and against individual RSFMs.

Dataset	Model Prompt	GeoSANE	$\Delta$
EuroSAT	97.8	99.1	+1.3
RESISC45	92.3	96.5	+4.2
fMoW	52.4	58.9	+6.5
Sen12Flood	84.2	85.2	+1.0
Cal. Wildfires	93.5	94.9	+1.4
BigEarthNet	82.6	88.7	+6.1

Table 5: Comparison between finetuning GeoSANE-generated models vs. direct finetuning of the models used as prompts for GeoSANE (ImageNet-pretrained ViT-L for classification and Swin-B for segmentation and detection). Both versions receive the same finetuning budget of 50 epochs.

Method	Backbone	m-eurosat	m-bigearthnet	m-so2sat	m-brick-kiln
SatMAE [ 7 ]	ViT-L	96.6	68.3	57.2	98.4
CROMA [ 13 ]	ViT-L	96.6	71.9	60.6	98.7
DOFA [ 61 ]	ViT-L	96.9	68.0	58.7	98.6
Prithvi-EO 2.0 [ 51 ]	ViT-L	96.5	69.0	54.6	98.6
AnySat [ 1 ]	ViT-B	95.9	70.3	51.8	98.6
Galileo [ 53 ]	ViT-B	97.7	70.7	63.3	98.7

Table 6: Results on four classification tasks from GEO-Bench [ 27 ] . All models are trained for 50 epochs. The reported numbers are overall accuracy (OA). The m-bigearthnet dataset is evaluated using mAP. For GeoSANE, we report the mean

\pm

standard deviation over three independently generated models per prompt (Section 3.3 ); for all baselines, we report the values reported in the respective paper [ 8 ] . Best results are in bold , second best are underlined .

\Delta

indicates the absolute improvement over the best baseline

Figure 4: UMAPs Visualization of the latent weight space of GeoSANE, colored by architecture (left) and modality (right). GeoSANE learns a compact shared latent representation of model weights from neural network models of different architecture and modalities (see Sec. 12 , in the supplementary material for details).

Dataset	MobileNet	ResNet-18	ViT-Base	ViT-Large
Dataset	3.5M	11M	86M	300M
RESISC-45	70.0	92.2	91.4	96.5
EuroSAT	96.2	98.7	98.4	99.1
fMoW	17.7	53.5	56.7	58.9
BigEarthNet	73.3	83.7	86.9	88.7
Sen12Flood	70.2	84.0	83.1	85.2

Table 7: We evaluate diverse GeoSANE-generated models on three tasks. All generated models are fine-tuned for 50 epochs. Please note, all models (for (a), (b), (c)) are generated using the same learned shared latent representation.

Dataset	CV-only	CV+RS	$\Delta$
EuroSAT	97.8	99.1	+1.3
RESISC45	93.4	96.5	+3.1
fMoW	53.1	58.9	+5.8
Sen12Flood	82.9	85.2	+2.3
Cal. Wildfires	92.2	94.9	+2.7
BigEarthNet	83.0	88.7	+5.7

Table 8: Downstream performance of models generated by GeoSANE trained without ( CV-only ) and with ( CV+RS ) remote sensing model weights.

Figure 5: Convergence comparison between GeoSANE-initialized models and models trained from scratch.

Task	Dataset	# of classes	Labels (per image)	Size	Channels	# of samples
Classification	RESISC45 [ 5 ]	45	single	$256\times 256$	RGB	31.5K
	EuroSAT [ 22 ]	10	single	$64\times 64$	Multispectral	27K
	fMoW [ 6 ]	63	single	$512\times 512$	Multispectral	$>1M$
	BigEarthNet [ 49 ]	19	multiple	$120\times 120$	Multispectral	590K
	Sen12Flood [ 36 ]	2 (binary)	single	$256\times 256$	SAR	16K
	California Wildfires [ 4 ]	2 (binary)	single	$224\times 224$	SWIR	20K

Table 9: Overview of datasets, tasks, label types, and channels.

Figure 6: Qualitative Results of the Flood Segmentation task on the Sen1Floods11 dataset

Figure 7: Qualitative Results of the Object Detection task on DIOR dataset

Checkpoint	Checkpoint	Checkpoint
jaychempan__EarthSynth	azdin__llava-onevision-weather-dora	mayrajeo__marine-vessel-yolo
torchgeo__presto	azdin__llava-onevision-weather-qlora	torchgeo__core-dino
DevPanda004__PrithviFlood	azdin__qwen2-vl-weather-adalora	torchgeo__delineate-anything
Mahadih534__YoloV8-VisDrone	azdin__qwen2-vl-weather-dora	torchgeo__delineate-anything-s
Mahadih534__yolov8_ship_det_satellite	banghyunmin__Thermal_Video_Detection	torchgeo__yolo11s_marine_vessel_detection
RedbeardNZ__LatentSync-1.6	banghyunmin__thermal-people-yolov11n	wangyi111__Copernicus-FM

Table 10: Remote Sensing Model checkpoints used to train GeoSANE.

Modality	Number of Models
Multispectral	41
RGB	15
Multimodal	13
SAR	6
DEM	2
Unknown	26

Table 11: Distribution of modalities for the models used to train GeoSANE.

Task	Number of Models
Self-supervised representation learning	36
Semantic segmentation	19
Object detection	7
Classification	4
Generative models	3
Regression	2

Table 12: Distribution of primary tasks for the models used to train GeoSANE.

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

A Composition-Aware Pretraining Framework for Geospatial Foundation Models
Aug 31, 2026Aryan Kashyap Naveen, Abhishek Srinivas, Pranav Moothedath +1Remote Sensing Image Understanding Representation Learning
National Institute of Technology, Karnataka · Surathkal, Karnataka, India · Visa Inc +2
Benchmarking Geospatial Foundation Models for Agriculture Applications
Jun 29, 2026Zhuocheng Shang, Sanmay Das, Ahmed EldawyRemote Sensing Image Segmentation Agricultural Remote Sensing
University of California, Riverside · Riverside, USA
GeoMeld: Toward Semantically Grounded Foundation Models for Remote Sensing
Apr 12, 2026Maram Hasan, Md Aminur Hossain, Savitra Roy +6Multimodal Foundation Models Multimodal Grounding
Indian Institute of Technology Bombay · Space Applications Centre, ISRO · University of Trento +1

Aug 31, 2026J/K move · Enter open · S save

A Composition-Aware Pretraining Framework for Geospatial Foundation Models

Aryan Kashyap Naveen, Abhishek Srinivas, Pranav Moothedath, Shrutilipi Bhattacharjee

National Institute of Technology, Karnataka · Surathkal, Karnataka, India · Visa Inc+2

Remote Sensing Image Understanding Representation Learning Geospatial Foundation Models Geospatial Representation Learning Compositional Representation Learning

Open paper AskPDF arXiv

Geospatial foundation models have emerged as state-of-the-art methods for downstream Earth observation tasks. However, existing pretraining methodologies process imagery through a single-concept lens, failing to capture the highly compositional nature of complex satellite scenes. We propose a composition-aware pretraining framework that explicitly encodes fractional land-cover mixtures. Each satellite image cell is mapped to a histogram representing its fractional land-cover distribution, which we term the "composition target". These targets serve as the primary prediction objective and are distilled into the backbone using Earth Mover's Distance. Experimental evaluation shows that composition-aware pretraining yields substantial gains on region-level understanding tasks requiring semantic similarity judgment, including zero-shot image retrieval and scene classification, while remaining competitive on tasks requiring fine-grained spatial precision, such as segmentation and object detection. With a 36.8M-parameter backbone, our framework outperforms SatMAE and Prithvi-EO-2.0, which contain 303M and 600M parameters, respectively, in most retrieval and scene classification settings. On the fine-grained ForestNet-12 dataset, a rigorous testbed for compositional discrimination, our method boosts baseline mAP@10 from 0.279 to 0.434, a 55.6% relative improvement, providing direct evidence for the effectiveness of explicit composition modeling. The code implementation can be found at https://github.com/05kashyap/GFM_Composition_Pretraining

Abstract

Geospatial Representation Learning

Explore similar work

A Composition-Aware Pretraining Framework for Geospatial Foundation Models

Benchmarking Geospatial Foundation Models for Agriculture Applications

GeoMeld: Toward Semantically Grounded Foundation Models for Remote Sensing