Recovering Hidden Park Visits from Sparse Phone Traces

Maps of "urban green deserts" (neighborhoods whose residents have little usable access to parks) increasingly rely on mobile-phone location data rather than census home addresses. Phone traces are sparse, though. A record appears only when the phone makes a call or wakes its GPS, so a 40-minute walk through a park can fall entirely between two pings. If the gaps are longer or more frequent for some groups, such as peripheral residents on cheaper phones or sparser cell towers, then a district can look like a green desert simply because its people are observed less often. This proposal treats green-desert measurement as a gap-filling problem. We reconstruct what each person did between pings using both the ping before a gap and the ping after it. The output is drawn from a spatial vocabulary that is coarse where data are thin and fine only where it matters for the answer, namely at park boundaries. We then ask whether the corrected exposure maps change which neighborhoods are labeled deserts.

The two-sentence version is this. Sparse location data systematically hides short, off-road park visits, and the hiding is uneven across a city. We fill those gaps with a model that conditions on both ends of each gap and predicts locations at a resolution that adapts to data density and park geometry, and we score it by how well it recovers each person's minutes in green space, not by token accuracy.

Core Claim

Green-exposure estimates computed from sparse traces are biased downward, and the bias differs between dense cores and sparse peripheries. Two additions reduce the bias more than either interpolating along the road network or the existing bidirectional trajectory model on a uniform grid. The first is a bidirectional masked model, meaning one that predicts a hidden location from context on both sides. The second is an output vocabulary built as a density-adaptive quadtree, refined near park polygons. The test is downstream: per-person green dwell time, and the share of residents classified as living in a green desert.

The gap this targets is visible in the file. Current audits work either from aggregated, static population counts or from fixed buffers. Measuring spatial accessibility … mobile phone signaling aggregates about 930,000 users into 827 uniform 500 m hexagons in Hefei, and finds that 64% of the area is in deficit at a 15-minute threshold but in aggregate surplus at 30 minutes. Unequal green uses circular 1000 m and 2 km buffers in Mardan, and its field audits show that mapped "parks" can be unusable. Neither reconstructs individual movement. In both, accessibility depends on a catchment threshold, which is the very quantity the Hefei inversion shows to be unstable. Measuring realized dwell time sidesteps that threshold, but only if the dwell time can be recovered from sparse data.

Research Program

The mechanism comes from bidirectional imputation. BRITS shows that coupling a forward and a backward recurrent pass with a consistency loss lets future observations correct imputed past values. It reports PhysioNet MAE of 0.278 against 0.445 for M-RNN. Multiagent off-screen behavior prediction in football shows that anchoring on both the disappearance point and the reappearance point prevents the drift that forward-only predictors accumulate. The closest existing system is BERT4Traj, which reconstructs masked locations in Kampala mobility data. It reaches 87.1% on coarse cell-tower regions but only 71.4% on 100 m GPS cells. The file attributes that drop to classification over a large uniform vocabulary. BERT4Traj treats resolution as a fixed choice and has no notion of parks. Bi-STDDP supplies a constraint we reuse: travel distance within a gap is bounded by elapsed time. Its finding that context wider than one step hurts warns us not to assume that longer context helps.

The discretization side supplies the vocabulary design. AdaSTEM splits space with quadtrees whose cells shrink where observations are dense, which avoids both over-smoothing and empty cells. PointRend spends computation only on uncertain boundary points, and reports roughly 30 times less memory than a dense prediction head. STARE documents the vocabulary-size versus fidelity tension for hierarchical S2 cell tokens but does not resolve it.

We combine these into one model. The vocabulary is a quadtree built from training ping density. Leaf cells that intersect a park polygon, from OpenStreetMap filtered by Sentinel-2 greenness, are subdivided further down to about 25 m. The model predicts hierarchically. It first predicts a coarse cell, then descends only while its predictive distribution straddles a park boundary, which is PointRend's rule transplanted to geography. Each masked gap is encoded with both anchor pings, the elapsed time, and the reachable set implied by Bi-STDDP's speed bound.

We add one head that existing imputers lack: the probability that any park stop occurred within the gap. The file notes that check-in imputers cannot tell whether an unobserved stop happened at all. For exposure measurement that is the central question, so we evaluate the head for calibration.

Baselines

The naive reference does nothing. It computes green dwell time only from observed pings and treats gaps as time spent nowhere. This is the implicit practice when signaling counts are aggregated, as in the Hefei study. The proposal must show this reference is biased before any gap-filling result means anything.

The geometric baseline routes each gap along the shortest pedestrian path on OpenStreetMap between the two anchor pings and spreads the gap time along that path. It is the strongest cheap competitor. For short gaps it is probably close to correct, and it needs no training.

The strongest learned baseline is BERT4Traj, reimplemented on the same data with its uniform 100 m grid, Space2Vec coordinate embeddings, and home and work anchor tokens. The file reports 71.4% reconstruction accuracy at 100 m on Kampala GPS and 87.1% on Kampala CDR, both from BERT4Traj. There is no reported result on Lacuna for green dwell-time error or desert-classification error for any method, so these downstream numbers must be produced here.

All learned models get the same splits, the same masking procedure, matched parameter counts, and the same hyperparameter search budget. Token accuracy is not comparable across vocabularies of different size, so every method is scored in metres of location error and minutes of green dwell error. Beating BERT4Traj means lower dwell-time error and fewer misclassified desert residents at equal or lower compute. Matching it on location error while winning on dwell error counts as a win, because dwell error is the question asked.

Three ablations isolate the ingredients:

  • the full model with a uniform 100 m vocabulary, which isolates adaptive discretization;
  • adaptive quadtree cells without park-boundary refinement, which isolates task-aware refinement;
  • a forward-only variant with the same vocabulary, which isolates the future anchor.

We also test a BRITS-style forward-backward consistency loss as an optional add-on.

Why not just…?

Why not just interpolate along the shortest walking path? This is a real threat. We expect it to win on gaps whose duration is close to the shortest travel time, because then there is no slack for a detour. A park visit can only hide in a gap whose duration exceeds that travel time. We therefore stratify every result by the slack ratio, defined as gap duration divided by shortest-path travel time. The proposal stands only if the learned bidirectional model clearly beats shortest-path routing at high slack and does not lose at low slack. If it does not, the honest conclusion is that cheap routing is enough, and the contribution reduces to the bias measurement.

Why not just correct the aggregate map by reweighting for ping density? Inverse-sampling weights per neighborhood could fix population-level desert maps without any trajectory model. We test this directly by computing reweighted aggregate exposure and comparing it to ground truth. The method cannot recover the distribution of individual exposure, for example the share of residents with zero park minutes per week, because it rescales means. If policy only needs neighborhood means and reweighting matches our model there, we will report that.

Why not just use BERT4Traj with a coarser or multi-level fixed grid? Coarser cells raise token accuracy, as the gap between 87.1% and 71.4% shows. But cells of 300 m or more cannot separate a park from the street beside it, which is the over-smoothing failure AdaSTEM was built to avoid. A sweep over fixed resolutions at 50, 100, 200, and 400 m, with dwell error plotted against compute, shows whether any single fixed resolution matches the adaptive vocabulary. If one does, adaptivity is unnecessary.

Proposed Experiments

Ground truth comes from dense GPS traces. Kampala GPS from BERT4Traj, with 586 participants, is the first target if access can be arranged. Otherwise we use any dense pedestrian-capable GPS panel that has an IRB-cleared release. We subsample each trace to mimic CDR-like sparsity, using real inter-event-time distributions taken from the paired CDR data where possible. We impose deliberately heavier sparsity in peripheral cells so we can test whether differential bias is recovered.

Park polygons come from OpenStreetMap, filtered by Sentinel-2 vegetation cover. A secondary analysis down-weights parks flagged as unusable, in the spirit of the field audit method in Unequal green, to see whether quality weighting moves desert labels more than imputation does. Vehicle datasets such as Porto and Chengdu from GREEN carry no pedestrian park visits. We use them only to stress-test the cost of the adaptive vocabulary at scale.

The primary metrics are per-person weekly green dwell-time error and the error in the share of residents labeled as living in a green desert, split into core and peripheral cells. Secondary metrics are location error in metres and calibration of the park-stop-occurred head.

Three checks decide the outcome. The first is whether the naive reference shows a core-versus-periphery gap in bias. If it does not, the motivating premise fails. The second is whether the bidirectional, park-refined model closes that gap at high slack better than shortest-path routing and uniform-grid BERT4Traj. The third is whether the corrected maps reclassify any neighborhoods. A method that lowers dwell error but never changes a desert label has not changed a conclusion that planners would act on.