A field report from BNPB Jakarta — with a research blueprint for scaling KRB using event footprints and triggering conditions.
Indonesia faces compound risk from tectonics, monsoon variability and rapid urbanisation. Hydrometeorological hazards already constitute the vast majority of recorded disasters, and climate change is projected to intensify extremes and raise coastal water levels while land subsidence amplifies exposure in low-lying cities. These pressures demand risk information that is granular, consistent, and policy‑aligned. In September, ADPC and Geoneon convened a technical workshop at BNPB (Jakarta) under the SDC‑supported ACER‑SEA program to demonstrate how Earth Observation (EO) and AI can sharpen Indonesia’s Kajian Risiko Bencana (KRB) by (i) moving from village averages to building‑level vulnerability, and (ii) automating historical event footprints (floods, landslides, wildfires) and linking them to triggering conditions (rainfall, flows, wind and temperature) to derive likelihood and frequency.
Indonesia sits on the boundary of three major tectonic plates and contends annually with floods, landslides, drought, extreme weather and forest and land fires. BNPB’s national briefing underscores that ~90% of disasters are hydrometeorological — a signal that climate variability and exposure, not just seismicity, dominate annual impacts.
Climate change raises the stakes. Relative sea level in Asia has risen faster than the global average and will continue to rise — a critical concern for Indonesia’s heavily populated coasts (IPCC AR6 WG1). At the same time, extreme rainfall events are projected to increase during the wet season across much of Indonesia this century, heightening flood and landslide potential (Kurniadi et al. 2024). In Java’s coastal cities, land subsidence rates measured by GNSS are several times larger than the present‑day rate of global sea‑level rise, compounding coastal flood risk in places such as Jakarta, Semarang and Pekalongan (Susilo et al. 2023). The World Bank’s Country Climate and Development Report likewise frames flood, coastal inundation and heat among top climate risks that intersect with development choices in energy, land, and urban planning (World Bank 2023)
Against this backdrop, KRB provides the national backbone for risk evidence and governance: Risk = Hazard × Vulnerability ÷ Capacity, with outputs including hazard, vulnerability, capacity and risk maps plus a narrative and matrix. KRB is a core service of Indonesia’s Minimum Service Standards (SPM) and underpins spatial planning and contingency plans. National/provincial hazard rasters are compiled at 100 m, and 30 m at the regency/city level.
Figure 1. KRB risk framework and workflow. BNPB’s schematic formalises Risk = Hazard × Vulnerability ÷ Capacity, with map products (hazard, vulnerability, capacity, risk), a risk matrix, and a documented process from general preparation to legalisation/integration. Source: BNPB.
Table 1 — Standard KRB scales and resolutions.
|
Administrative level |
Typical map scale |
Raster resolution |
Primary uses |
|
National/Province |
1:250,000 |
100 m |
Provincial DM and spatial plans |
|
Regency/City |
1:50,000–1:25,000 |
30 m |
Regency/city DM plans, Destana, contingency plans |
|
Community |
1:10,000–1:5,000 |
Vector or high-res raster |
Community action plans; detailed contingency plans |
Source: BNPB.
The Jakarta workshop, part of the ACER‑SEA initiative led by ADPC and Geoneon with support from SDC, focused on two deliverables: (1) a technical demonstration that enhances social vulnerability inputs to KRB using open building data and dasymetric mapping; and (2) a research blueprint showing how computer vision can turn EO archives into national libraries of event footprints tethered to trigger conditions — flood rainfall/flows, landslide antecedent moisture and slope, and wildfire “fire weather”.
BNPB’s contributions recapped the KRB mandate, process flow, and data responsibilities (BPS, BIG, sector agencies), including worked examples of dasymetric versus choropleth population mapping and the InaRISK portal. This provided the policy and methodological frame our pilot plugs into.
Figure 2. Why dasymetric mapping matters. BNPB’s example contrasts choropleth vs dasimetrik population mapping, showing how redistributing counts to settlement areas resolves within‑unit heterogeneity. Source: BNPB.
Our demonstration area was Padang Panjang (West Sumatra). The governing problem is familiar: village‑level census totals mask steep within‑village gradients in exposure and vulnerability. We therefore redistributed census counts to individual buildings and then aggregated to building clusters, creating map units that better reflect lived settlement patterns.
We merged village polygons with village‑level census (total population and subgroups such as poor, disabled, vulnerable) from local partners, then ingested building footprints from Overture Maps using DuckDB against the GeoParquet distribution. After filtering to residential footprints, we spatially joined buildings to villages and ran dasymetric redistribution: each building receives a share of the village totals proportional to its footprint area. We then applied DBSCAN to form building clusters and generated refined boundaries (Voronoi from cluster seeds) suitable for rasterization at KRB scales. All code, a worked example, and a QGIS project were delivered for replication.
Table 2 — Building footprints by provider in the Padang Panjang AOI (Overture).
|
Provider (Overture) |
Count of buildings |
|
Google Open Buildings |
8,887 |
|
Microsoft ML Buildings |
4,325 |
|
OpenStreetMap |
359 |
These counts inform QA (e.g., areas dominated by ML‑derived roofprints vs OSM edits). Source: Geoneon.
The demonstration scripts and notebook are public:
https://github.com/Geoneon/dasymetric-mapping-example.
Instead of assuming uniform population within a village, population and vulnerable groups concentrate where buildings are, and cluster boundaries align with actual settlement structure. This improves overlays with hazard footprints (flood depth grids, landslide susceptibility, burn scars), and it makes KRB‑derived decisions — e.g., contingency planning, micro‑zoning for mitigation — more defensible at neighborhood scale.
Incomplete height/storey metadata, mixed‑use buildings and institutional populations can bias dasymetric estimates. Our recommendations: integrate authoritative registries (BIG, Bappeda, sector ministries), add simple rules (e.g., exclude very large non‑residential footprints), and use local QA to refine residential classifications.
EO archives allow us to build consistent national libraries of event footprints that span multiple hazards:
Figure 3. Geoneon Wildfire Datastack. A schematic shows inputs (Sentinel imagery, vegetation structure, topography, climate, building footprints), preprocessing, and derived severity and exposure indices—an example of how event‑ and driver‑layers combine into actionable risk products. Source: Geoneon.
The second step is linking each footprint to triggers drawn from national/regional gridded data—e.g., CHIRPS precipitation, GloFAS flood hazard layers, CORDEX extremes—computing antecedent windows (1–3 days, 7–14 days) and threshold exceedances. Statistical or ML models can then estimate probability of occurrence and frequency classes at KRB scales and by season. The regional case study presented by SDC (Laos) illustrates this multi‑source approach (CHIRPS, GloFAS, CORDEX; HRSL/Kontur/LitPop), as well as how triangulating multiple population surfaces yields robust targeting.
Figure 4. Multi‑agency screening logic (Laos case study). A comparative map/table demonstrates how UNOSAT, SDC Hub, and Ageospatial independently screened villages using different datasets and platforms; convergence across methods supports robust targeting (first prioritise 22 villages agreed by all three, then those agreed by two). Source: SDC Regional Hub.
Finally, convert the trigger models to hazard likelihood rasters (100 m/30 m), overlay with exposure (buildings, roads, facilities) and vulnerability (redistributed social indicators), and publish through InaRISK/KRB with lineage metadata. This keeps the innovation inside the national workflow while improving statistical power and granularity.
Our plenary talk parsed where computer vision and foundation models add value across hazard, exposure, and vulnerability:
We thank BNPB for hosting and substantive inputs on KRB standards and data responsibilities; ADPC and SDC for the ACER‑SEA partnership and regional framing; and our university and local government partners in West Sumatra for sharing census data used in the demonstration. Workshop decks cited above include BNPB’s KRB briefings, ADPC/Geoneon project overview and AI talk, Geoneon’s technical demonstration, and the SDC Regional Hub case study.