Direct Geolocation: Methods and Applications

• Direct Geolocation is a set of techniques that estimate, verify, or generate geographic positions directly from raw spatial observations.
• These methods leverage diverse modalities such as imagery, radio-frequency signals, network latency, and even biological markers for location inference.
• Key techniques include optimization over location hypotheses, continuous spherical mapping, and grid-search methodologies that bypass traditional staging.

Direct geolocation is a family of localization methods that estimates, verifies, or generates geographic position from primary observations in the target spatial domain. In the surveyed literature, those observations include ground imagery and phone-grade GPS, panoramic images and 2-D cartographic maps, raw RF I/Q samples acquired from low Earth orbit, Internet delay measurements, exposed web pages, and fungal DNA profiles. The term is used with modality-specific meanings: direct regression to latitude/longitude, continuous generation on the sphere, single-step position-domain search from raw signals, delay-based presence verification, and VM-scale self-auditing. This suggests that direct geolocation is best understood not as a single algorithmic template but as a recurring design choice: to infer or verify location in the geographic domain itself rather than only through an intermediate proxy such as 6DoF pose estimation in a reconstructed map, a pure administrative label, or a trusted self-report (Sun et al., 2018, Clements et al., 8 Aug 2025, Abdou et al., 2019).

1. Conceptual scope and recurring mathematical structure

Across the literature, direct geolocation usually replaces a staged pipeline with a location-domain objective. In RF localization, a candidate emitter location is searched over a geographic grid and scored by a position-domain correlation computed from raw samples; in image geolocation, the output may be corrected latitude/longitude, a continuous point on $\mathbb{S}^2$, or the GPS coordinate of a highest-scoring retrieved candidate; in network geolocation, the output may be a point estimate, a verified region, or a falsification of an asserted location (Clements et al., 8 Aug 2025, Wang et al., 23 Mar 2025, Chuzhoy et al., 2 Jun 2026, Abdou et al., 2019).

A common structural pattern is explicit optimization over location hypotheses. Examples include the map-conditioned pose likelihood

$$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$

for camera pose $g=(x,y,\gamma)$ in urban image geolocation, the direct RF search metric $S(\mathbf{p})$ over candidate emitter positions, the route-matching objective

$$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$

and the DeepSpace spatial intensity

$\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$

Although the observations and likelihood models differ, the optimization target is geographic location itself rather than a separate latent pose graph, landmark correspondence set, or database label (Mousavian et al., 2016, Clements et al., 8 Aug 2025, Samano et al., 2019, Grantham et al., 2019).

The resulting systems vary in what “direct” excludes. In autonomous-driving imagery, the exclusion is explicit HD maps, SfM, 2D–3D matching, and PnP/reprojection optimization. In GNSS jammer localization, the exclusion is two-step TDOA/FDOA extraction followed by nonlinear localization. In secure Internet geolocation, the exclusion is blind trust in browser coordinates or IP-address databases. This suggests that “direct” is partly a contrast term whose exact meaning depends on the dominant baseline of the subfield (Sun et al., 2018, Clements et al., 8 Aug 2025, Abdou et al., 2019).

2. Vision-based direct geolocation

In ground-vehicle localization, direct geolocation has been formulated as regression from image evidence and a noisy GPS prior to corrected latitude and longitude. A representative model uses a pre-trained ResNet feature extractor followed by 1–2 LSTM layers and a regression head that predicts a 2D correction $\Delta l=[\Delta lat,\Delta lon]$. Training uses the Smooth $L_1$ loss on the offset between raw phone GPS and survey-grade ground truth, and evaluation converts coordinates to UTM and measures $\|p_{pred}-p_{gt}\|_2$ in meters. On the ACM Multimedia 2017 Grand Challenge dataset, this approach reported mean prediction error $2.47$ m with standard deviation $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$0 m, which the paper interprets as nearly lane-level accuracy; it also reported KITTI errors of $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$1 m, $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$2 m, and $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$3 m on three sequences, all without using an HD map or SfM model (Sun et al., 2018).

A second line of work geolocalizes directly on a 2-D map rather than a 3-D reconstruction. One method estimates planar camera pose $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$4 by combining semantic segmentation, line detection, vanishing points, and building-facade decomposition with a 2-D map containing building footprints and identities. The map projection yields predicted facades

$$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$5

which are matched against observed facade hypotheses $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$6 to form a normalized pose likelihood. On a 270-image campus dataset with 11 buildings, marginalized building assignment improved reference-set building identification from $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$7 to $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$8, and using building identity reduced ambiguity relative to geometry-only localization (Mousavian et al., 2016).

Route-based direct geolocation on maps replaces single-image discrimination with sequence discrimination. A Siamese-like dual-branch model embeds Google Street View panoramas and OpenStreetMap tiles into a shared 16-D space using a weighted soft-margin ranking loss. Because a single image is “not sufficiently discriminatory,” localization is performed by concatenating embeddings along a route and matching them to all possible route descriptors in the map. Using StreetLearn panoramas and Mapnik-rendered OSM tiles, the method reported over $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$9 top-1 localization accuracy for routes of about 20 locations, roughly 200 m in length, and top-1% image-to-map retrieval of about $g=(x,y,\gamma)$0–$g=(x,y,\gamma)$1 (Samano et al., 2019).

Language-grounded and LVLM-based systems extend direct image geolocation by integrating textual priors. “$g=(x,y,\gamma)$2: Geolocation via Guidebook Grounding” uses 3,832 clue sentences extracted from a GeoGuessr guidebook, frozen CLIP RN50x16 image embeddings, frozen RoBERTa Base clue embeddings, and pseudo-label-supervised attention over clues. On a 90-country StreetView benchmark, the full model achieved $g=(x,y,\gamma)$3 Top-1, $g=(x,y,\gamma)$4 Top-5, and $g=(x,y,\gamma)$5 Top-10 country accuracy, compared with $g=(x,y,\gamma)$6 Top-1 for the strongest image-only ISN baseline. GaGA, by contrast, turns geolocation into an interactive LVLM task built on MG-Geo, a 5 million image-text-pair dataset. With a frozen CLIP ViT-L/14-336 vision encoder, a trainable projector, and Llama3-8B, it reported $g=(x,y,\gamma)$7 country accuracy and $g=(x,y,\gamma)$8 city accuracy on reproduced GWS15k, and the paper explicitly noted that direct coordinate prediction is intrinsically difficult for LLMs because long floating-point numbers are tokenized sequentially (Luo et al., 2022, Dou et al., 2024).

Recent worldwide systems differ sharply in how they realize directness. LocDiffusion reframes image geolocation as a direct geolocation generation problem on the sphere,

$g=(x,y,\gamma)$9

and performs diffusion in a hidden Spherical Harmonics Dirac Delta representation rather than on coordinates. Its SHDD decoder is learning-free, and under small Gaussian noise the reported perturbation drift was about $S(\mathbf{p})$0 km for SHDD versus $S(\mathbf{p})$1 km for Sphere2Vec and $S(\mathbf{p})$2 km for rbf. On Im2GPS3k with $S(\mathbf{p})$3, it achieved $S(\mathbf{p})$4 at 1 km, $S(\mathbf{p})$5 at 25 km, $S(\mathbf{p})$6 at 200 km, $S(\mathbf{p})$7 at 750 km, and $S(\mathbf{p})$8 at 2500 km, with markedly stronger generalizability to unseen geolocations. Pinpoint, in contrast, remains database-grounded: it uses a contrastive image–GPS embedder trained on MP16-Pro and OSV-5M, three-channel retrieval, and an attention-based reranker with cross-source support tokens from nearby locations. The final prediction is the GPS coordinate of the highest-scoring candidate. It reported state-of-the-art results on IM2GPS3k, YFCC4k, and OSV-5M, including IM2GPS3k accuracies of $S(\mathbf{p})$9, $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$0, $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$1, $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$2, and $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$3 at 1, 25, 200, 750, and 2500 km, and OSV-5M GeoScore $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$4 with mean Haversine error 743 km; it also reported inference time $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$5 s/image versus 288 s/image for GeoRanker (Wang et al., 23 Mar 2025, Chuzhoy et al., 2 Jun 2026).

3. Radio-frequency direct geolocation and direct position determination

In passive RF localization, direct geolocation denotes a single-step, grid-search-based maximum-likelihood estimator of emitter position. For receiver $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$6, the received baseband signal is modeled as

$$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$7

with delays and Dopplers parameterized by candidate emitter position through receiver geometry. The position-domain correlation

$$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$8

is evaluated over a geographic grid, and the maximizing candidate is the emitter estimate. This formulation is especially attractive in low-SNR regimes and with short captures, conditions emphasized for LEO-based GNSS interference sensing (Clements et al., 8 Aug 2025).

The principal obstacle is computational cost. A 10° by 10° latitude-longitude grid with 0.01° spacing yields about one million position-domain correlation values, and each candidate requires geometric delay and Doppler calculation, time shift, Doppler compensation, and sample summation. A GPU-accelerated implementation exploited the embarrassingly parallel structure of candidate-point processing and reported about $$\hat{s}^x_m=\arg\min_{s^x_m\in {\mathcal{S}^x_m}} DIST(s^x_m,s^y_m),$$9 to $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$0 speedup over CPU on an Intel i7-14700K and Nvidia GTX 4060 Ti platform. The simulations used two LEO receivers, 58 snapshots spaced by 1 s, 50 ms coherent intervals, and 5 MHz complex sampling for GPS L1 spoofing, pure tone, chirp, and sawtooth interferers. Single snapshots exhibited waveform-dependent ambiguities—large streaks for tones, multiple equally prominent peaks for chirps, and an “X” shape for sawtooth interference—but noncoherent accumulation over 58 snapshots left sharp peaks at the true emitter positions (Clements et al., 8 Aug 2025).

Quasi-direct geolocation (QDG) reduces that cost by transforming I/Q samples into a sparse time-frequency representation and searching in a quantized time-frequency domain. Using the STFT

$\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$1

QDG performs noise estimation, thresholding, and compression before an approximate position-domain search. The time-frequency bins above threshold are retained using a CFAR-like rule

$\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$2

and the geolocation metric is evaluated through fast delay-Doppler matching and interferometry on the retained bins rather than through raw-sample interpolation. The paper gives QDG complexity as

$\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$3

contrasted with classical DG/DPD complexity

$\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$4

This is the central scalability claim of the method (Pojani et al., 2 Jul 2026).

Real-data demonstrations used OPS-SAT PRETTY, a repurposed 3U CubeSat with two Earth-facing patch antennas, during Jammertest 2025 in Andøya, Norway. For the FFI “Porcus Major” 50 W EIRP CW jammer near $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$5N, $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$6E, the dataset comprised about 1 GB of I/Q samples for the analyzed pass. With 8-bit requantization and $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$7, QDG reduced this to about 0.76 MB, with only about $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$8 of STFT bins carrying meaningful signal energy; the related low-complexity formulation reported compression from 1080 MB of raw data to 0.34 MB for the final STFT-product representation. The resulting geolocation fell within approximately $\hat s(x)=\arg\max_{s\in\mathcal D}\hat g(s).$9 km² around ground truth, the reported MSA95 at the peak was about 900 m, and the semi-minor axis of the 95th-percentile error ellipse was about 50 m. The same work also reported that time-frequency filtering was more effective than null steering for jammer cancellation, and that weaker emitters became visible after cancellation (Pojani et al., 2 Jul 2026, Pojani et al., 22 Jun 2026).

These RF results make clear that direct geolocation is not confined to coordinate regression. In this literature, it is a physically parameterized search over candidate positions using raw or compressed measurements. The trade-off is explicit: QDG gains orders-of-magnitude reductions in data volume and computational burden, but with lower sensitivity and accuracy than full DG/DPD and with limitations on signal classes, especially wideband noise-like emitters, direct-sequence spread-spectrum signals, matched-spectrum GNSS jammers, and GNSS spoofers (Pojani et al., 22 Jun 2026).

4. Delay-based Internet and IP geolocation

In network geolocation, direct methods infer location from latency structure rather than from commercial databases alone. A central difficulty is weak delay-distance correlation in poor-connected networks. The RTD-Corr model decomposes RTT as

$\Delta l=[\Delta lat,\Delta lon]$0

where $\Delta l=[\Delta lat,\Delta lon]$1 is the propagation-delay ratio, $\Delta l=[\Delta lat,\Delta lon]$2 the tortuosity factor, and $\Delta l=[\Delta lat,\Delta lon]$3 the direct geographical distance. Large-scale measurements over China Internet yielded overall correlation $\Delta l=[\Delta lat,\Delta lon]$4, identifying it as a typical poor-connected network, but intra-ISP correlations were much stronger: 0.689 for CERNET, 0.670 for China Telecom, and 0.611 for China Unicom. Restricting classical GeoGet and CBG to rich-connected sub-networks produced large gains: modified GeoGet improved from 44/116 correct city mappings, about $\Delta l=[\Delta lat,\Delta lon]$5, to 112/116, about $\Delta l=[\Delta lat,\Delta lon]$6, and modified CBG reduced median error from 629.8 km to 315.4 km (Ding et al., 2019).

Hints-based systems use network-observable textual evidence as candidate generation and latency measurement as validation. HLOC extracts location hints from router reverse-DNS names using IATA, ICAO, UN/LOCODE, CLLI, and GeoNames codes, consolidates roughly 448k raw location codes into 5,474 locations, and validates hints with ZMap and RIPE Atlas. A hint is accepted only if there exists a probe close enough to the hint and the observed round-trip time satisfies

$\Delta l=[\Delta lat,\Delta lon]$7

with $\Delta l=[\Delta lat,\Delta lon]$8 km and $\Delta l=[\Delta lat,\Delta lon]$9 ms in the study. On CAIDA-derived datasets of 1.4M IPv4 routers and 183k IPv6 routers, HLOC obtained 45k verified IPv4 locations and 5k verified IPv6 locations among responsive addresses. It also showed that, for verified IPv4 locations, GeoLite was wrong for $L_1$0 and ip2location for $L_1$1, and that ip2location IPv6 matches were wrong for $L_1$2 (Scheitle et al., 2017).

GPS-tagged smartphone datasets provide a high-precision benchmark for evaluating database-based IP geolocation. Using 248 million smartphone location reports across New York City, Chicago, and Philadelphia, with median reported GPS accuracy 17 m in 2020 and 11 m in 2021, one study compared MaxMind and IP2Location against GPS-derived ground truth. On April 2021 fixed-line samples, MaxMind GeoIP2 paid reported median errors of 2.62 km in New York, 3.31 km in Chicago, and 4.02 km in Philadelphia, whereas IP2Location paid reported 12.04 km, 23.95 km, and 20.87 km respectively. The same study found that mobile networks were much worse than fixed-line networks, that universities were often more accurately geolocated than consumer networks, and that most fixed-line IP assignments were stable over two months, although stability varied across ISPs (Saxon et al., 2021).

Secure direct geolocation over the Internet turns latency into a verification primitive rather than a passive estimate. Client Presence Verification (CPV) uses three trusted, synchronized verifiers and Minimum-Pairs one-way delay estimation to test whether a client lies inside the verifier triangle, accepting only if the inferred delay geometry satisfies a triangle-area condition up to tolerance $L_1$3. In PlanetLab experiments with 80 nodes and 34 different triangles, CPV reported about $L_1$4 false accepts and about $L_1$5 false rejects with 600 iterations; under a modeled WiFi scenario, false accepts and false rejects were roughly $L_1$6 and $L_1$7. Server Location Verification (SLV) performs an analogous real-time check for the first machine that terminates the client’s TCP and TLS handshake, with pilot results of $L_1$8 false accepts and $L_1$9 false rejects over about 200 experiments (Abdou et al., 2019).

5. Cloud, web-content, and forensic modalities

Cloud geolocation extends delay-based reasoning to VM-scale auditing in multicloud environments. GeoFINDR assumes a dishonest CSP that may have total control over hardware and data, high-performance internal networking including dark fibers, and multicloud architecture. Its pipeline operates from the VM itself and comprises four stages: a Greedy algorithm selects a distributed audit set $\|p_{pred}-p_{gt}\|_2$0 of nearby RIPE Atlas landmarks; sectorization identifies a set $\|p_{pred}-p_{gt}\|_2$1 of landmarks with distance-delay behavior similar to that of the VM; the estimated location is computed as the barycenter of $\|p_{pred}-p_{gt}\|_2$2; and a do-while correction loop repeats until the distance between successive estimates falls below tolerance. The landmark-dispersion step, Dispoints, has complexity $\|p_{pred}-p_{gt}\|_2$3. In experiments, the best reported accuracy was 22.1 km with $\|p_{pred}-p_{gt}\|_2$4, $\|p_{pred}-p_{gt}\|_2$5, $\|p_{pred}-p_{gt}\|_2$6, and $\|p_{pred}-p_{gt}\|_2$7, with 34.3 s audit time; across multiple declared positions, the reported average accuracy was 22.6 km. When landmarks in the Paris area were removed, accuracy degraded to 303.8 km for the true declaration case and audit time rose to 396.9 s (Ider et al., 25 Apr 2025).

Passive direct geolocation of exposed management interfaces operates on web content rather than packet delay. WebGeoInfer targets remote management devices whose administrative pages leak location clues in heterogeneous, partially structured text. It clusters structurally similar pages using Nilsimsa hashes and Hamming-distance similarity,

$\|p_{pred}-p_{gt}\|_2$8

extracts differential text across aligned DOM subtrees, augments ambiguous fragments with search-engine snippets, applies a weighted multi-LLM ensemble, and resolves ambiguity with IP geolocation databases. From an initial corpus of 205,028 SCADA-tagged IP-port combinations and 89,477 filtered HTML pages, the framework inferred locations for 5,435 devices across 94 countries and 2,056 cities, achieving country-, city-, and street-level accuracies of $\|p_{pred}-p_{gt}\|_2$9, $2.47$0, and $2.47$1 respectively (Yang et al., 16 Aug 2025).

Forensic direct geolocation uses biological measurements as spatial evidence. DeepSpace operates on fungal presence/absence vectors from dust samples and aggregates over an ensemble of deep neural network classifiers trained on random Voronoi partitions. Each classifier estimates region probabilities $2.47$2, which are converted into a spatial intensity surface

$2.47$3

and the predicted origin is $2.47$4. On 1,301 dust samples from the continental United States, the best mixed-partition DeepSpace model reported median absolute prediction error 97.8 km, coverage 96.3%, and state match 60.2%; on a global 28-country dust dataset, the mixed-partition model achieved 89.5% country classification accuracy (Grantham et al., 2019).

Taken together, these systems show that direct geolocation extends well beyond conventional imaging and RF sensing. What remains invariant is the attempt to map high-dimensional evidence to geographic provenance itself, whether the evidence is RTT structure, administrator-customized text, or microbiome composition.

6. Reliability, ambiguity, and technical limits

A recurrent limitation is that direct geolocation quality depends strongly on whether the observation actually contains discriminative geographic signal. In image geolocation, this issue has been formalized as image localizability. Given a base geolocation model $2.47$5, an image is defined as localizable at scale $2.47$6 if $2.47$7. A selective predictor $2.47$8 can then abstain on non-localizable images using scale-aware selection functions derived from the output distribution over geographic cells. Spatial Entropy and Prediction Density were shown to outperform Softmax Response and Monte-Carlo Dropout across datasets and scales, and on Im2GPS3k at city scale (25 km) Prediction Density improved usable accuracy from $2.47$9 to $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$00 on the retained predictions (Panagiotopoulos et al., 2021).

This result highlights a common misconception: direct geolocation is not synonymous with always producing a reliable fine-grained coordinate. Some systems deliberately verify presence inside a region rather than a point, as in CPV and SLV; some solve only country-level classification, as in $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$01; and some expose intrinsic ambiguity through prediction regions, as in DeepSpace. This suggests that abstention, regional verification, or uncertainty sets are not peripheral add-ons but core design choices when the evidence is weak (Abdou et al., 2019, Luo et al., 2022, Grantham et al., 2019).

The dominant failure modes are modality-specific but structurally similar. In visual systems, single-image localization may be unreliable because local map semantics are not sufficiently distinctive, and database-grounded methods are limited to indexed locations in a static index. In LVLM systems, direct coordinate generation is weakened by sequential tokenization of long floating-point numbers. In RF QDG, thresholding and time-frequency quantization reduce sensitivity and restrict the usable signal classes. In delay-based VM geolocation, nearby landmark density can dominate accuracy, and audit time can rise sharply when landmarks are sparse (Samano et al., 2019, Chuzhoy et al., 2 Jun 2026, Dou et al., 2024, Pojani et al., 22 Jun 2026, Ider et al., 25 Apr 2025).

A second recurrent issue is adversarial or distributional mismatch. GeoCLIP-like retrieval performance can collapse when anchors are not spatially aligned with the test distribution, whereas LocDiffusion is explicitly less sensitive to anchor choice and number. Secure Internet geolocation must account for VPNs, proxies, anonymizers, and illicit traffic relaying, leading CPV to incorporate cryptographic client puzzles. Cloud auditing must assume dishonest CSP behavior rather than merely economically rational behavior. Web-content geolocation must handle abbreviations, partial phrases, and cross-lingual variation (Wang et al., 23 Mar 2025, Abdou et al., 2019, Ider et al., 25 Apr 2025, Yang et al., 16 Aug 2025).

The literature therefore points to a broad methodological conclusion. Direct geolocation succeeds when a model, signal processor, or verifier can preserve the geometry most relevant to the target spatial scale while suppressing nuisance variability: visual appearance change, RF noise and bandwidth constraints, network-path tortuosity, landmark sparsity, or misleading textual context. The diversity of successful formulations—regression, route matching, diffusion on $$p(\mathcal{I}\mid g)=\frac{S(Z,\widehat{Z})}{S_{max}(Z,\widehat{Z})}$$02, position-domain RF search, verifier-triangle consistency, barycentric VM estimation, and Voronoi-ensemble spatial intensity—shows that direct geolocation is less a single method than a unifying problem formulation centered on geographic inference itself.