The Missing Layer in Coastal Climate Risk Maps

A climate-risk map can be accurate about the sea and still miss the flood

Coastal planning often begins with a familiar question: how high will the ocean rise? On parts of Java’s northern coast, that question is necessary but incomplete. The land itself is moving—sometimes by centimeters each year—and the change is uneven across cities, farms, deltas, and industrial zones.

That matters because residents experience **relative sea level**: the height of the water compared with the ground they stand on. A seawall, drainage pipe, rail line, or home can become more exposed even if the ocean follows the expected climate trajectory, simply because the land underneath it sinks faster.

A study published in *Science Advances* in April 2026 assembled an island-wide picture of that motion. The team processed roughly 4,000 Sentinel-1 radar images acquired from 2017 through 2023, combined ascending and descending satellite views with observations from 28 GNSS sites, and produced vertical and east–west displacement fields at about 75-meter resolution. It then used an unsupervised machine-learning pipeline to group places with similar motion through time and space.

The value of the AI was not that it predicted the next flood or issued an autonomous warning. It helped researchers organize a vast, noisy geophysical record into coherent deformation patterns that could be interpreted alongside geology, land cover, high-resolution imagery, and earlier field studies.

Research references: [1]

What the system found

The satellite analysis estimated average vertical land motion across Java at about **minus 0.5 centimeters per year**, with large local variation. The researchers identified subsidence hotspots in coastal and inland areas including Jakarta, Bekasi, Bandung, Tegal, Semarang, Cilacap, Surabaya, Sidoarjo, and Jember. Across 9 of 12 selected major urban districts, more than one-fifth of the analyzed area was sinking faster than one centimeter per year. In Bandung, Pekalongan, and Demak, more than 30 percent of the analyzed area exceeded five centimeters per year.

The time dimension added a second warning. A quadratic trend analysis found statistically significant acceleration in 17 percent of currently subsiding areas during 2017–2023, while 45 percent showed deceleration. Comparing the newer Sentinel-1 estimates with ALOS-1 observations from 2007–2010, the researchers classified 13 percent of previously identified subsiding zones as having intensified, 22 percent as having slowed, and 65 percent as broadly stable within their chosen threshold.

Those numbers are evaluated observations within the study’s data and methods. They are not proof that one policy caused improvement or one activity caused deterioration.

The machine-learning stage began with principal component analysis to compress the motion time series while retaining at least 90 percent of cumulative variance. K-means grouped similar temporal profiles; DBSCAN refined them into spatially contiguous regions; and multivariate dynamic time warping compared the evolution of horizontal and vertical motion across clusters. The pipeline produced 30 spatial clusters, then excluded background cluster 0, leaving 29 clusters that were organized into six spatiotemporal groups. Researchers interpreted likely drivers only afterward, using land cover, geology, imagery, and prior site-specific studies.

Only after the clustering did the team interpret likely mechanisms using geology, land use, imagery, and prior site-specific research. The authors concluded that intensive groundwater withdrawal was the dominant mechanism in several urban coastal centers, while other clusters were more consistent with seasonal agricultural extraction, natural sediment compaction and loading, or mixed anthropogenic and tectonic processes.

That sequence is important. The algorithm identified patterns; human researchers, using external evidence, inferred plausible drivers.

Research references: [1]

The ground can dominate the near-term sea-level budget

To translate motion into coastal risk, the study created “virtual tide gauges” at five-kilometer intervals along Java’s northern coast. Each gauge combined land-motion estimates with satellite-altimetry observations and probabilistic sea-level projections.

For the present-day period, the study estimated relative sea-level rise rates ranging from **0.2 to 9.0 centimeters per year**, averaging 1.7 centimeters per year along the assessed coastline. Under the intermediate SSP2-4.5 pathway and the study’s linear land-motion assumption, projected rates for 2050 ranged from 0.4 to 9.0 centimeters per year. The authors reported that these rates were two to ten times the rates attributable to sea-level rise alone.

In the study’s high-emissions, medium-confidence decomposition, vertical land motion accounted for 25 to 91 percent of the current relative-sea-level budget across 77 percent of the coastline. The companion open dataset summarizes the 2050 contribution as 27 to 85 percent across more than three-quarters of coastal length.

These figures do not mean climate-driven sea-level rise is unimportant. They mean adaptation can fail if it treats the coastline as a fixed platform. Where subsidence is rapid, a local and partly manageable process can dominate the next few decades even as global warming drives the longer-term baseline upward.

Research references: [1] [3]

The most useful AI result is a better decision map

The practical achievement is not simply a more colorful hazard map. It is a framework for differentiating places that need different responses.

A coastal alluvial zone may require better elevation monitoring and settlement-aware engineering. A residential or industrial cluster associated with groundwater demand may call for metering, enforcement, and reliable piped-water alternatives. An agricultural cluster with strong seasonal motion may need groundwater accounting tied to crop calendars. A mixed tectonic–anthropogenic region may require denser ground instrumentation before officials assign a cause.

The island-wide view also reveals a governance risk the authors call a “balloon effect”: slowing subsidence in a regulated urban core can coincide with emerging hotspots farther out, as extraction and water-intensive activity shift geographically. A policy can look successful at one monitoring point while pressure migrates beyond its boundary.

That is a design target, not an evaluated policy outcome. The study observed that parts of Jakarta moved from rapid subsidence in 2007–2010 toward near stability or uplift in 2017–2023, coinciding with groundwater restrictions, relocation incentives, and expanded piped-water infrastructure. The observational record cannot isolate the effect of each intervention. It does show why citywide or regional monitoring is safer than judging success from a single tide gauge or neighborhood.

What the study does not establish

The paper is unusually clear about limits that should travel with its headline findings.

First, clustering is not causation. Some deformation classes may combine groundwater withdrawal, shallow or deep sediment compaction, and tectonic motion in ways the available data cannot fully disentangle.

Second, the virtual tide gauges make simplifying assumptions. They apply land motion as a deterministic rate and, for future projections, largely assume that contemporary subsidence continues linearly. That may be wrong if pumping, land use, rainfall, infrastructure, or regulation changes. The authors’ sensitivity analysis found that nonlinear motion could shift some 2050 relative-sea-level estimates by 0.3 to 0.7 meters.

Third, the virtual gauges inherit limits from their inputs. The 5-kilometer coastal segments use 0.25-degree satellite-altimetry cells, assigning the closest sea-level grid point to each segment; nearshore conditions can differ at finer scales. The reconstructed series were benchmarked against one long-record PSMSL gauge at Cilacap on Java’s southern coast, where the trend correlation was 0.9 and the rate differed by 0.1 millimeter per year. That is a useful check, not coast-wide in-situ validation.

Fourth, the research evaluates a mapping and attribution framework—not whether a particular agency used it, whether decisions changed, or whether damages fell. There is no operational impact trial, prospective warning evaluation, or cost-benefit analysis. The paper reports support from an LDEO postdoctoral fellowship and U.S. Department of Defense funding for four authors; the authors declared no competing interests.

Those constraints do not erase the result. They define the responsible use: as a high-resolution screening and planning layer that directs field verification, monitoring, and adaptation—not as a substitute for local hydrology, engineering surveys, or accountable public decision-making.

Research references: [1]

A resilience playbook for sinking coasts

For governments and infrastructure operators, the study suggests a disciplined sequence:

1. **Measure the platform, not only the water.** Pair tide gauges and climate projections with repeated land-motion observations.

2. **Map change at the scale of infrastructure.** A regional average can hide neighborhood-level differences that bend roads, rupture pipes, or reverse drainage gradients.

3. **Use machine learning to triage, then verify.** Clusters can focus scarce fieldwork, but mechanism attribution should be tested with groundwater records, boreholes, GNSS, engineering inspections, and local knowledge.

4. **Tie policy boundaries to the pressure system.** Monitor peri-urban and agricultural zones so reduced pumping in one jurisdiction does not simply shift extraction elsewhere.

5. **Recompute plans as motion changes.** Linear extrapolation is convenient; it should not become permanent policy. Update elevation, drainage, and protection standards when acceleration or deceleration appears.

6. **Publish uncertainty with the map.** Planning layers should show observational error, alternative mechanisms, and scenarios for changing land motion.

The broader lesson is modest but consequential: AI can make Earth-observation data more legible, and that can expose a missing part of climate risk. The tool does not remove uncertainty or choose a policy. It helps planners ask a better first question: before deciding how high the water may rise, how fast is the ground moving beneath every neighborhood—and what evidence would justify acting on that map?