The ground under Jakarta is moving.

Where the city sinks fastest

441,934 people in DKI Jakarta live on ground sinking faster than 2 cm a year. In 16 of 261 measured kelurahan the fastest tenth of the ground drops more than that.

The map colours every kelurahan by its rate over 2017–2023: the fastest decile (p10) shows the hotspot cores the way the literature's peak values do; the median tells the neighbourhood story. Fastest single decile: Pluit (Jakarta Utara), −3.99 cm/yr, −1.77 at the median. Toggle the velocity field to see the measurement itself — including Tangerang and Bekasi, which the kelurahan table does not cover.

It is not only a coastal story any more. Of the 20 fastest-subsiding kelurahan by median rate, 12 are in West Jakarta and only 4 in North Jakarta; just 6 of the 20 sit in the coastal belt. Across the city, sinking rate and low ground are close to independent (rank correlation ρ = 0.149): the fastest ground includes inland West Jakarta kelurahan whose low ground sits several metres up, which is a foundation and buried-pipe problem rather than a flooding one.

Why it sinks: decades of groundwater extraction compacting the young alluvial and marine clays under the coastal plain, with the load of the built city on top — the mechanism is well documented (Abidin et al. 2011; Chaussard et al. 2013; Ohenhen et al. 2026). And it has slowed: on the continuous GNSS record the central-Jakarta station CJKT drops from −8.6 mm/yr over 2010–2016 to −3.8 mm/yr over 2017–2022, and all 3 stations inside this working area slowed, while 12 of 17 elsewhere on the north Java coast got faster. The 10–25 cm/yr figures still quoted for Jakarta are from the 1990s and 2000s.

Where the ground goes next — if nothing changes

On this model 0.39 M people stand on ground below +1 m today, 0.42 M by 2030 and 0.47 M by 2050 if the 2017–2023 rates hold; below mean sea level, 0.12 M → 0.19 M.

Read those as an ordering, not as counts. "Ground" here is the lowest of the roughly nine 30 m elevation pixels in each 100 m cell — a way of finding roads and yards rather than roofs in a surface model, but also a minimum of nine noisy samples. Computed on the cell mean instead, the same grid and the same population give 35,965 below +1 m rather than 391,029 — a factor of 10.87. The published figure is, on the cell mean, the population below +2.93 m. The ranking of kelurahan survives this; the absolute counts do not.

This is a linear extrapolation, not a forecast, and its assumptions are worth more than its signal. Each 100 m cell's ground drops at its measured 2017–2023 rate, year after year; cells without radar coverage are held still. Groundwater policy, the sea wall, drainage, rainfall and sea-level rise are all outside the model. Ground height comes from Copernicus GLO-30 — a 30 m surface model acquired 2011–2015, with ±2–4 m vertical accuracy. At its median rate, Kedaung Kali Angke (Jakarta Barat) — low ground at 0.1 m — reaches mean sea level in about 15 years; 14 kelurahan already have low ground below it, and exactly 1 more crosses it inside the window this clock animates.

Three assumptions, priced. Correcting the twelve-year-old elevation model for subsidence since acquisition moves today's count from 391,029 to 432,584 — a number this clock does not otherwise reach until 2037. Raising the sea-level datum 15 cm, within the range implied by measured Java Sea rise, takes 2050 to 536,337. Letting the rate decay at the halving time the Jakarta GNSS station actually shows, instead of holding it flat, takes 2050 down to 422,303. Two of those make the clock optimistic and one makes it pessimistic: the honest 2050 bracket is 422,303–508,612 on assumptions alone. The review article works through all three.

And one number here is a rounding artefact. The elevation model pins water and near-shore cells to exactly 0.00 m, where 61,713 people live. They sit one arbitrary millimetre above the mean-sea-level threshold, so the below-MSL count jumps 43,886 between 2025 and 2026 — 62% of the entire 25-year rise — and then adds a median of 1,103 a year. Use the below-MSL series for its shape after 2026, not for its 2025 level.

Checks before the story

Three things sit beside the radar field: continuous GNSS stations (Susilo et al. 2023, millimetre-level vertical rates), the published literature on Jakarta's hotspots, and the depositors' own published statistics for the same data. Then the exposure numbers are checked against the census and against floods that actually happened. Two of these checks were misdescribed on this page before the review; the corrections are below the tables.

The plausibility check fails, and the failure is the finding — not an apology. Where Jakarta floods and where Jakarta sinks are different places, and an agency that plans drainage from a subsidence map will defend the wrong neighbourhoods. BPBD's 2021–2024 record is dominated by riverine floods along the Ciliwung and its tributaries in East Jakarta — Kampung Melayu (72 event dates), Cawang (67 event dates), Cililitan (54 event dates) — kelurahan on higher ground that barely subside. Across the city, flood frequency is inversely related to low ground (ρ = -0.135) and only weakly to the sinking rate (ρ = 0.323). Subsidence-driven exposure is a coastal, tidal and drainage problem, and it is the January 2020 extent — one extreme event seen by Sentinel-1 — that overlaps the most-exposed kelurahan disproportionately: 12% of their area against 5% city-wide. We keep the gate red rather than redefine it to pass.

Correction: the GNSS gap is a period gap, not a frame offset. This page previously explained the ≈2 mm/yr difference in the table above as an InSAR reference-frame effect. That was wrong. The depositors transform their field into the same global frame family the GNSS solutions use, and the published GNSS rates are fitted over each station's whole record — at CJKT, 2010–2022 — while the radar covers 2017–2023. Refitting each station on the radar's own window (with annual and semi-annual terms removed) collapses the disagreement from +1.78 ± 0.51 mm/yr to −0.03 ± 0.89. The field is tied far better than the table suggests, and the residual it was hiding is the real signal in chapter 01 — the ground slowed down between the two windows.

The strongest check is a replication. The depositors publish per-municipality statistics for this same field (Ohenhen et al. 2026, table S2). Regridding their point cloud independently — different resampling, different boundary file, different zonal method — reproduces that table across all five Jakarta municipalities, with a largest disagreement in the mean of 0.032 cm/yr. The ingest, filter, gridding and zonal steps are faithful.

The GNSS comparison is a consistency check rather than an independent one: the depositors used stations from this same archive to align their field to the global frame, so it cannot fail unless something is badly broken. Genuinely independent reprocessing — our own LiCSBAS run over LiCSAR frame 098A (2017–2024) against this deposited field — is pending; it needs a 13 GB download and runs after the flagships' bandwidth frees.