The Sinking City
Jakarta sits on a swampy alluvial plain at the mouths of 13 rivers, home to more than 10 million people. Decades of groundwater extraction have set the city on a path that no flood infrastructure can fully reverse: the land itself is sinking.
Satellite radar measurements from 2016 to 2024 show that subsidence is far from uniform. The map reveals where the ground is dropping fastest.
1,651,218 building footprints mapped across Jakarta
To understand who is most vulnerable, researchers built a building-level flood risk model combining hazard, exposure, and vulnerability scores for every structure in the city.
How I built this
Data pipeline: satellite data acquisition (Sentinel-1 SAR, 2016–2024), time series processing via LiCSBAS2, ground-truth validation against GNSS measurements, and interpolation (kriging in QGIS).
Paper and GitHub repository forthcoming.
What the Algorithm Sees
Zooming into North Jakarta, every building carries a composite flood risk score derived from three inputs: how exposed the structure is to flooding, how severe the hazard at its location, and how vulnerable the building itself is.
The colors represent that combined risk. The darkest buildings face the highest convergence of hazard, exposure, and vulnerability.
One limitation is visible in the coloring itself: flood-related metrics were computed at the kelurahan (sub-district) level, producing the blocky, uniform patterns across the map. Buildings within the same kelurahan share the same hazard and exposure values, masking the variation that exists building by building.
Filling the Data Gap
Jakarta's building data has significant gaps in physical characteristics: construction materials, building use, and height. This information is essential for physical vulnerability analysis, which feeds directly into risk modeling. At scale, it could help policymakers prioritize interventions and reduce risk where it is needed most.
To fill this gap, the research team applied three vision-language models to predict building typologies from Google Street View imagery in Kecamatan Penjaringan, North Jakarta. From 83,728 raw Street View images, quality filtering narrowed the set to 29,937 buildings that received predictions.
But coverage was uneven, even within Penjaringan. The model could only predict buildings on streets wide enough for the Street View car to reach. When we zoom into two neighborhoods, the pattern sharpens.
How I built this
End-to-end ML classification pipeline: data acquisition via the Google Street View API (83,728 images), multi-stage data validation (error filtering, spatial buffer, LLaVA-NeXT quality screening), prompt engineering across four strategies, and multi-model evaluation using three vision-language models (Gemini 2.0 Flash, GPT-4o, Claude 3.5 Sonnet). Built in Python.
Paper and GitHub repository forthcoming.
The Gap
Two neighborhoods in Penjaringan, North Jakarta. Both are fishing villages. Both face tidal flooding. Both sit on sinking ground. But the algorithm treats them very differently.
RW 04 Kamal Muara
43.9% of buildings unpredicted
535 buildings, 235 with no prediction
RW 22 Muara Angke
84.3% of buildings unpredicted
728 buildings, 614 with no prediction
Land Tenure, Not Street Width
The gap is not a random technical limitation. RW 04's wider streets exist because the government happened to formalize land ownership there, enabling road construction and infrastructure upgrades. Street View cars could drive in. The models could see.
RW 22's density tells a different story. Many residents were displaced from other parts of Jakarta and resettled on land whose legal status remains unresolved. Without formal tenure, government infrastructure rarely follows. Alleys stay narrow. The camera never comes.
Not every gap follows this pattern. Near PIK, the gated luxury development visible across the water from Kamal Muara, some buildings go unpredicted because the estate management prohibits Street View vehicles from entering. Others sit along narrow commercial lanes that no car can reach. But the dominant pattern is clear: where land tenure is informal, the data thins out.
The algorithm's blind spots trace the contours of structural inequality, not just street geometry. Research and policy need more than remote sensing and AI. They need to engage with communities directly.
Alongside the modeling, a team of researchers was already doing exactly that.
The Flip
Same place. Different way of seeing.
While the modeling work mapped risk from above, a parallel ethnographic study handed cameras to 15 young people aged 18 to 25 in RW 04 and RW 22 and asked them to photograph the places that matter to them. What they captured is not what any satellite or Street View car could see.
Life in the Kampung
Joy and Play
"We call this location a waterfall because the currents from the two areas collide, resulting in a strong current. Many children swim and play in this location, one of the exciting places and enjoy fate because of the continuous flooding."
Where flood models see hazard, the youth of Muara Angke see a swimming hole. They renamed this drainage overflow "Curug Eceng," the waterfall. Down the road, a flooded alleyway becomes "Waterboom Gang 1," borrowing the name of Jakarta's commercial water park. The humor is pointed. The play is real. Risk reimagined as ritual.
"We swim to PIK a lot. Sometimes we get chased by security, but we're used to it. We always make it back."
PIK is the gated luxury development visible across the water from Kamal Muara. Swimming there is recreation, but it is also a small act of crossing a boundary that was drawn without asking.
Care and Empathy
"Now that the shelling area is higher [following the seawall construction], they have to walk farther and climb more. It is tiring for them."
The seawall was built for protection. But for the elderly women who peel mussels for a living, it made their daily commute steeper and longer. Resilience infrastructure does not land equally on everyone.
"This is why I stay. I don't want to leave her."
Resilience in the kampung is relational and intergenerational. One young person stays not because the flooding is manageable, but because leaving means leaving family behind.
Spatial Exclusion
"They build without asking. They just come. We adapt later."
Government infrastructure arrives without consultation. Seawalls, roads, and reclamation projects reshape the landscape around the kampung while the people inside it adjust as best they can. Adaptation, here, is not a strategy. It is a condition imposed from outside.
Quotes and photographs from Darmawan, Adinta, Wimbadi, & Colven (2025). Urban Geography.
Both Ways of Seeing
Pull back to the city scale, and both layers are visible at once: the risk scores that algorithms compute and the lives that people actually lead in those same buildings.
Neither view is complete on its own. The flood model tells us which structures face the highest hazard. The ethnography tells us that hazard is not the whole story, that resilience is woven into relationships, humor, daily routines, and the stubborn act of staying.
When both ways of seeing work together, something else becomes possible: risk assessments that account for what people value, not just what they stand to lose. Infrastructure that serves the people it surrounds. Research that begins by listening.
This story map is part of the Empowering Resilience in a Sinking City research project. The communities of RW 04 Kamal Muara and RW 22 Muara Angke made this research possible. The youth participants who shared their photographs and words are its foundation.
How I built this
This story map serves 1.65 million building footprints as vector tiles (PMTiles generated via tippecanoe), with scroll-driven chapter transitions (Scrollama) and responsive map rendering (MapLibre GL JS). The frontend is vanilla HTML, CSS, and JavaScript, deployed as a static site. Explore the full technical details in the Methods section below.
