AI in Floodplain Management: the tools we have been building
AI can get a flood practitioner from data to an answer faster than ever. The harder part is getting to an answer you can defend. This three-minute video shows how Zealandia Consulting uses AI in floodplain management, the guardrails we put around it, and the twenty-plus tools we have built for councils and private clients across New Zealand.
In this article: why defensibility matters more than speed, how we use AI responsibly, a tour of our tools, and where we are heading next.
Watch the video
The video is adapted from our workshop AI in Floodplain Management: practical uses, confidence and guardrails, presented to Floodplain Management Australia in November 2026. It covers:
• what AI is, and what it is not
• where AI goes wrong in flood work
• how to use it within clear boundaries
• the tools we have built and use every day
• where we are taking them next
Why this matters: can we defend the decision?
In floodplain management, the question is not "can AI produce an answer?" but "can we defend the decision?" Speed is useful. Confidence and accountability are what let an output support a real decision about people, property and infrastructure.
Recent events show why. In January 2023, Auckland Airport recorded 249 mm of rain in 24 hours, against a previous record of 161.8 mm. In February 2022, the Wilsons River at Lismore peaked at 14.4 m, about 2 m above its record. Both events broke the records that models and forecasts relied on. A tool that returns one confident number makes that kind of uncertainty harder to see, not easier.
How we use AI responsibly
Our working view of AI: a very fast pattern-matcher with an excellent vocabulary, and no judgement. It learns from data, it is good at narrow tasks, its outputs are likely answers rather than verified facts, and it is only as good as its inputs. Wrong rainfall in means wrong flood levels out. A smooth, plausible flood map can still be wrong.
So we work to a few simple rules:
• Check the inputs first. Terrain, rainfall, boundary conditions and roughness set the ceiling on quality.
• Watch for bias. A model trained on city streets can mislead on farmland, and a map that "looks right" is easy to accept.
• Show ranges, not false precision. We highlight the few assumptions that move the result most.
• Prompt for challenge, not reassurance. We ask AI to flag unsupported claims, mismatched units and missing limitations.
We also sort every use into three zones:
Zone | Typical uses | Control |
Green | Drafting, summarising, formatting, coding help | Human review before use |
Amber | Classification, parameter suggestions, scenario screening | Documented validation and expert oversight |
Red | Autonomous life-safety, statutory or investment decisions | Not without approved controls and an accountable owner |
The tools we have built
Over the past few years we have built more than twenty tools and prototypes, grouped by where they sit in the floodplain workflow. Each is a worked example of the rules above: a bounded task, a measurable output and a person accountable for it.
Tool | Stage | What it does | Status |
Waipuke Survey | Observe | Field app and dashboard for flood-impact evidence: depths, floor levels, flow paths and photos. Built with Auckland Council Healthy Waters. | Operational since the 2023 floods |
EnviroEyes™ | Observe | Solar-powered, GPS-enabled cameras with AI image descriptions and condition checks on a multi-site dashboard. | Platform |
Virtual StreamWalker | Observe | Turns a walked stream into a searchable digital twin; AI transcripts flag issues and route them to the responsible team. | Pilot (Lucas Creek) |
Overland flow-path dashboards | Assure | Track region-wide flow-path assessments from field visit to resolved action, comparing mapped and observed flow paths. | Platform |
BEDEM | Detect | Bare-earth detection on 0.3 m aerial imagery for repeatable earthworks monitoring. | Ongoing, Auckland |
Sediment and landslip mapping | Detect | Bare earth before and after Cyclone Gabrielle, intersected with flow paths, rivers and steep slopes. | Regional study published |
SSRIMS | Detect | Sediment risk scoring for unsealed roads: 778 km scored as 79,205 segments, with the top 20 roads prioritised. | Delivered, Auckland |
Regional flow paths | Predict | Automated flow paths, catchments and time of concentration from LiDAR (aerial laser survey). | Delivered, Auckland and Northland |
FloodAI | Predict | Brings rainfall, river, soil moisture and tide data together to predict flood likelihood, answer plain-language questions and draft warnings. A person approves every alert. | Pilot, demonstration only |
Story Maps, portals and AI flood visualisation | Communicate | Help communities see the risk at their own door. AI images are always labelled and matched to modelled depths. | In use |
Beyond water: the same observe, detect and act pattern now supports earthworks compliance and STLKR, a commercial wild-deer recovery platform that protects native forest.
What's next: one connected evidence chain
The next step is not another tool. It is connecting the ones we have, so that every flood event teaches the whole system, not just one project.
one evidence chain · 5 stages and an event feedback loop
We are exploring three moves:
• One assurance record per tool: purpose, data versions, validation results, limitations and owner, published with every output.
• An event feedback loop: after each storm, Waipuke surveys, cameras, StreamWalker and flow-path checks score FloodAI and the regional flow paths.
• A whole-chain pilot catchment: run every tool end to end in one catchment and publish a public performance scorecard.
Five things to take away
1. Use AI for clearly defined tasks, not vague "intelligence".
2. Confidence comes from evidence, validation and independent challenge.
3. Keep assumptions, uncertainty and limitations visible.
4. Match the checking effort to the consequence of being wrong.
5. Accountability remains human.
Talk to us
Wondering where AI could take the friction out of your floodplain work, and what evidence you would need before trusting it? We are happy to talk it through.


Comments