GeoCLOSP: Location-Aware Remote Sensing Retrieval

• GeoCLOSP is a location-aware variant of CLOSP that aligns text, multispectral optical imagery, SAR imagery, and geographic coordinates in a unified latent space.
• The model employs contrastive learning with an additional geographic loss component, using spherical harmonic encoding and a SIREN-based location encoder, to specialize retrieval for spatially confined events.
• Empirical results demonstrate enhanced nDCG for geographically limited classes like earthquake damage and snow/ice, while also highlighting a trade-off between semantic generality and location specificity.

GeoCLOSP is the geographically aware variant of CLOSP, a contrastive framework for text-to-remote-sensing-image retrieval that learns a shared embedding space for text, Sentinel-2 multispectral optical imagery, and Sentinel-1 SAR imagery. GeoCLOSP extends that base system by adding a location encoder so that geographic coordinates are aligned with image embeddings alongside the original text–image alignment. In the formulation introduced for CrisisLandMark, the method is presented not as a separate retrieval paradigm but as a modified architecture and training objective that uses geographic context to specialize retrieval toward location-dependent crisis events and rare geographic features (Cambrin et al., 14 Jul 2025).

1. Definition and task setting

GeoCLOSP is defined within the problem of text-to-remote-sensing-image retrieval beyond RGB sources. The underlying retrieval task learns a scoring function

$f: Q \times D \to \mathbb{R},$

where $Q$ is the query space, $D$ is the remote-sensing corpus, a query $q$ is represented by labels $L_q$, and an image $i$ has labels $L_i$. The base CLOSP model uses contrastive learning to align text with two visual modalities: Sentinel-2 multispectral optical imagery and Sentinel-1 SAR imagery. GeoCLOSP adds a third alignment relation, between image embeddings and geographic coordinates, so that the latent space is shaped both by semantics and by spatial context (Cambrin et al., 14 Jul 2025).

The motivation is specific to geospatial remote sensing. The paper argues that, unlike natural images, geospatial imagery is inherently associated with geographic coordinates, and that some retrieval targets are strongly location-dependent. The examples emphasized are earthquake damage, snow and ice, certain crisis events, and rare geographic features. This yields the paper’s central interpretation of GeoCLOSP as a model that trades some general semantic breadth for geographic specificity. CLOSP is described as better for general semantic retrieval, whereas GeoCLOSP is described as a specialist for location-dependent queries (Cambrin et al., 14 Jul 2025).

This specialization matters because the framework is not limited to one visual sensor. The paper explicitly treats optical multispectral imagery and SAR as distinct sensing modalities with different physics and states that directly aligned optical–SAR pairs are rarely available because Sentinel-1 and Sentinel-2 have different revisit times. GeoCLOSP therefore inherits CLOSP’s use of text as a bridge across unpaired visual modalities, then overlays location-aware supervision on top of that structure (Cambrin et al., 14 Jul 2025).

2. Base CLOSP framework and multisensor data model

CLOSP uses three encoders: a text encoder, a Sentinel-1 encoder for SAR, and a Sentinel-2 encoder for multispectral optical imagery. The encoders are trained to project outputs into a shared latent space by aligning text with satellite imagery associated with the same item. A key design decision is that CLOSP does not directly align Sentinel-1 and Sentinel-2 images to each other. Instead, text is aligned independently with each visual modality, and this shared textual supervision acts as the bridge that unifies the latent space (Cambrin et al., 14 Jul 2025).

The remote-sensing corpus for this framework is CrisisLandMark, which contains over 647,000 images in total: 338,342 Sentinel-1 images and 308,461 Sentinel-2 images. The data come from five source datasets: re-BEN, CaBuAr, QuakeSet, MMFlood, and Sen12Flood. Sentinel-1 imagery is represented as GRD SAR with two channels, $VV$ and $VH$. Sentinel-2 imagery is represented as L2A multispectral data with 12 channels after removing one band post atmospheric correction. All imagery is aligned to 10 m spatial resolution and divided into $120 \times 120$ patches (Cambrin et al., 14 Jul 2025).

The text side does not use free-form captions. Instead, it uses structured labels for land cover, land use, and crisis events. The harmonized label space includes Dynamic World-derived classes—Trees, Crops, Shrub and Scrub, Water, Grass, Built, Flooded Vegetation, Bare, and Snow and Ice—plus crisis-event classes Flooded Area, Earthquake Damage, and Burned Area. To construct a unified corpus, the paper maps fine-grained CORINE annotations to the 9-class Dynamic World system and augments crisis datasets with Dynamic World labels. This means that GeoCLOSP operates over a semantically structured but non-captioned query space (Cambrin et al., 14 Jul 2025).

The base CLOSP objective is a symmetric image–text contrastive loss. For image embedding $Q$0, text embedding $Q$1, batch size $Q$2, and learnable temperature $Q$3, the paper gives

$Q$4

$Q$5

and

$Q$6

GeoCLOSP preserves this objective and adds a parallel image–location alignment term (Cambrin et al., 14 Jul 2025).

3. Location encoder and GeoCLOSP objective

GeoCLOSP extends CLOSP by adding a location encoder that embeds the geographic coordinates associated with each image–text pair. The paper states that this encoder is inspired by SatCLIP and uses spherical harmonic positional encoding together with a Sinusoidal Representation Network (SIREN). The coordinate pipeline is therefore

$Q$7

where $Q$8 is the location embedding. The paper is explicit that coordinates are transformed into learned embeddings rather than inserted as raw numeric features, but it does not explicitly specify the raw coordinate format or the normalization procedure (Cambrin et al., 14 Jul 2025).

Architecturally, the geographic branch is parallel to the original text–image branch. The paper does not describe concatenation of coordinate features into the text encoder or the visual encoders. Instead, geographic information is fused at the objective level by contrastively aligning image embeddings and location embeddings. In that sense, GeoCLOSP is a loss-level extension of CLOSP rather than a new sensor-fusion backbone (Cambrin et al., 14 Jul 2025).

The additional losses are

$Q$9

$D$0

and the total GeoCLOSP objective is

$D$1

This equally weighted averaging is an important part of the method’s empirical behavior. The paper later identifies the fixed weighting between semantic and geographic loss components as a limitation. That observation is consistent with the reported trade-off: location-aware supervision can sharpen retrieval when place is discriminative, but can also dilute purely semantic structure when geography is not the primary signal (Cambrin et al., 14 Jul 2025).

4. Training regime and retrieval evaluation

The training setup combines pretrained components from multiple sources. The paper reports that it uses a vision encoder from SSL4EO, a text encoder from SentenceTransformers, and a location encoder from SatCLIP. For GeoCLOSP and the specialized models, the reported backbone choices are ResNet-50 on the vision side and MiniLM on the text side. The training configuration uses 30 epochs, batch size 64, the Adam optimizer, a cosine annealing learning-rate scheduler with warmup, and a maximum learning rate of $D$2 (Cambrin et al., 14 Jul 2025).

The batch construction remains inherited from CLOSP’s multisensor setup. For a batch of $D$3 elements, one visual modality—either SAR or optical—is selected for each element, and the image embedding is paired with its associated text embedding. GeoCLOSP adds the corresponding geographic coordinate for that same item. The positive pairs are therefore matched image–text and image–location instances, while the negatives are the non-matching in-batch combinations. The paper does not describe hard-negative mining, cross-batch memory, or a queue-based mechanism (Cambrin et al., 14 Jul 2025).

Evaluation uses 2,047 distinct multi-label text queries created from co-occurring label combinations. Relevance between a query and an image is graded with the IoU-based rule

$D$4

The reported retrieval metrics are nDCG, Precision@K, and Recall@K, with $D$5. The paper also evaluates zero-shot multilabel classification and analyzes spatial distance correlation in the learned embedding space (Cambrin et al., 14 Jul 2025).

The zero-shot multilabel classification result is especially informative because it isolates the semantic side of the representation. GeoCLOSP records macro $D$6, whereas CLOSP-RN records $D$7. The paper interprets this as evidence that geographic coordinates do not universally improve semantic discrimination and may introduce a semantic–spatial trade-off rather than a uniform gain (Cambrin et al., 14 Jul 2025).

5. Empirical profile: aggregate gains and class-specific specialization

On the main retrieval benchmark, GeoCLOSP is reported as the best overall model among those tested. The paper’s summary metrics are shown below (Cambrin et al., 14 Jul 2025).

Model	nDCG@10	nDCG@1000
SkyCLIP-T	33.46	37.88
CLOSP-RN	50.50	56.23
GeoCLOSP	51.14	57.76

The full result table also reports $D$8 and $D$9 for GeoCLOSP, compared with $q$0 and $q$1 for CLOSP-RN. The paper emphasizes that GeoCLOSP reaches 57.76% nDCG@1000, roughly 20 absolute points above SkyCLIP-T, while also improving over the strongest CLOSP baseline (Cambrin et al., 14 Jul 2025).

A second empirical result concerns the geometry of the latent space. For CLOSP-RN, the reported distance–embedding correlations are Pearson $q$2 and Spearman $q$3. For GeoCLOSP, they become Pearson $q$4 and Spearman $q$5. The paper interprets this as evidence that the location encoder introduces a moderate monotonic spatial structure into the latent space, even though the linear relationship remains weak (Cambrin et al., 14 Jul 2025).

The most distinctive results are class-specific. GeoCLOSP strongly improves retrieval for some geographically constrained targets, but underperforms on broad, geographically widespread classes. The paper’s own examples are summarized below (Cambrin et al., 14 Jul 2025).

Class	CLOSP-RN nDCG	GeoCLOSP nDCG
Earthquake Damage	0.00	65.50
Snow and Ice	18.47	60.90
Crops	90.53	67.84
Trees	80.58	59.23

These results define the method’s practical identity. Earthquake damage and Snow and Ice are the paper’s clearest demonstrations that geographic priors can rescue retrieval where visual evidence alone is weak, localized, or rare. By contrast, the deterioration on Crops, Trees, and Grass shows that forcing spatial structure into the embedding can separate semantically similar samples simply because they occur in different regions. The paper also reports nuanced or mixed cases: Burned Area changes only slightly, Water is effectively saturated for both models, and Flooded Area does not improve under GeoCLOSP (Cambrin et al., 14 Jul 2025).

6. Interpretation, limitations, and relation to adjacent contrastive geolocation methods

The paper’s own interpretation is that GeoCLOSP should not be regarded as a universal replacement for CLOSP. It is a specialist variant whose advantage emerges when geography is a genuine discriminative factor. That formulation places the method in a broader family of geospatial contrastive systems that align imagery with location, but it occupies a distinct niche because text remains the bridge across unpaired visual modalities. This differentiates GeoCLOSP both from image–GPS retrieval models such as GeoCLIP, which align images directly with continuous GPS embeddings for worldwide geo-localization (Cepeda et al., 2023), and from CSP, which uses self-supervised image–location contrastive pretraining for geo-tagged imagery and retains location during inference for downstream classification (Mai et al., 2023).

Several limitations follow directly from the reported setup. First, the weighting between semantic and geographic losses is fixed: $q$6 The paper explicitly notes this as a limitation and suggests that the balance between semantic and geographic supervision is task-dependent. Second, CrisisLandMark is dominated by re-BEN, which accounts for 88% of the images, and the paper explicitly warns about geographic bias in the source datasets. A plausible implication is that a location-aware model may partially encode dataset geography in addition to phenomenon geography (Cambrin et al., 14 Jul 2025).

Third, the paper is explicit that geographic information is not uniformly helpful. GeoCLOSP does not improve zero-shot classification, and it is weaker than CLOSP-RN on several widespread land-cover classes. This suggests that the gain from geographic coordinates depends less on the mere availability of location metadata than on whether the queried phenomenon is spatially concentrated or climatically constrained. In that respect, GeoCLOSP can be read as an instance of a broader design principle also visible in later geolocation work: spatial structure becomes valuable when the training objective acknowledges that geography is not just an auxiliary label but part of the signal itself (Chen et al., 25 Sep 2025).

Within the larger contrastive geospatial literature, GeoCLOSP can therefore be characterized as a location-aware remote-sensing retrieval model that combines three elements: multisensor text bridging, image–location alignment, and a deliberate trade-off between semantic generality and geographic specificity. That combination is its defining contribution, and the paper’s results indicate that its most appropriate use is specialized retrieval of location-dependent crisis events and rare geographic features rather than universally improved semantic retrieval (Cambrin et al., 14 Jul 2025).