When the Siren Sounds Too Late: Indonesia’s Flood Early-Warning Gap

On 5 June, heavy rain pushed rivers over their banks in Medan, North Sumatra's capital, flooding around 2,000 houses across four sub-districts and five villages. By the AHA Centre's count, roughly 6,900 people were affected; most rode it out at home, thirty ended up in a church shelter, and six road sections went under water. By local standards, this was a routine event — the kind of flood Medan absorbs most rainy seasons without making international news.

What makes it worth pausing on is the timing. The same province was, six months earlier, the epicentre of the floods and landslides that killed more than a thousand people across Aceh and North and West Sumatra and triggered a wave of official soul-searching about why the warnings hadn't worked. June's flood is the system's first real chance to show whether anything changed. The honest answer, from where I sit, is: a little, and not in the places that matter most.

The Warning That Technically Worked

BPBD Medan did what the post-mortems said agencies should do: it monitored river levels, issued early warning advisories, pre-positioned evacuation equipment, and coordinated food assistance once water rose. On paper, this is the early-warning chain functioning as designed. Compare that to the November-December disaster, where alerts were criticised as too slow, too generic, or simply not reaching people in remote upstream catchments before landslides hit.

So why does a "working" warning system still produce 6,900 displaced people and submerged roads? Because the warning chain was built to answer one question — will the agency know in time? — and largely ignores a harder one: will the household three kilometres from the river office know in time, and will they be able to act on it? An advisory broadcast through official channels can be technically timely and still arrive too late for someone without a radio, a working phone, or a place to go.

Last-Mile Is Not a Delivery Problem, It's a Trust Problem

The ODI's recent assessment of Indonesia's disaster governance after the Sumatra floods made a point I keep coming back to: the gaps are less about whether warning infrastructure exists and more about whether it connects to the people who need to act on it, in a form they trust enough to act on. Indonesia has flood sensors, weather models, BNPB dashboards, and an increasingly sophisticated national apparatus. What it has unevenly is the social infrastructure — the RT/RW neighbourhood networks, the village heads, the mosque loudspeakers — that turns an official advisory into "leave now."

I've watched this play out the same way in flood after flood: the national-level data is often right, and the village-level response is often late, and the gap between those two facts is rarely a data problem. It's that the advisory has to pass through two or three more hops — provincial to district, district to sub-district, sub-district to neighbourhood — and at each hop someone has to decide it's credible enough to act on, with their own name attached if they're wrong. In Aceh, BNPB has reportedly tried to strengthen exactly this RT/RW-level coordination since the 2025 disaster. That's the right instinct. It's also slow, unglamorous work that doesn't show up on a dashboard.

What Gets Funded Versus What Gets Used

There's a predictable pattern after any high-profile disaster: money and attention flow toward sensors, satellite monitoring, AI-driven flood prediction, and command-centre dashboards — the parts of the system that are visible, procurable, and demonstrable to donors and oversight bodies. The RT/RW coordination layer, the community early-warning volunteers, the loudspeaker maintenance budget — these don't make for compelling line items, and they're the first things cut when a province's disaster budget tightens between events.

Medan's June flood didn't kill anyone, as far as the reporting shows, and that's worth crediting to the agencies that responded. But "nobody died in a routine flood" is a low bar, and the system won't be tested against that bar again — it'll be tested the next time a flash flood moves faster than a multi-hop advisory chain can travel. The question that matters isn't whether BNPB's dashboards are good enough. It's whether the warning that BNPB sends can reach a village head, be believed, and turn into people moving, inside the window that a flash flood actually gives you. Until that chain gets the same investment as the sensors feeding it, every "lessons learned" report after the next disaster will read a lot like the last one.