Disaster Damage Surveys: Saving Time with Drones and AI

Disaster damage surveys take a long time because staff walk the site and check things one by one. If you photograph first with a drone and use AI to narrow down damage candidates, you can bring the survey order forward. However, the numbers AI counts must be verified by people using samples.
In short
- Damage surveys take long because the site is wide and the judgment criteria vary.
- If you stitch drone photos into an orthoimage, you can view the site from the office.
- AI counts are for narrowing down survey candidates, and sample verification is a must.
The day after a typhoon passes, calls reporting damage keep coming into the village office. There are far more reported houses than surveyors, and the bridge at the village entrance is still under water. Residents ask when someone will come, but recovery support cannot start until the survey results are out.
Drone photography and AI interpretation are used to shorten this wait. However, the tools do not replace the survey. They are closer to deciding where to look first. This post explains, in turn, why surveys take so long, the drone photography procedure, the limits of AI counts, and how the results lead to recovery support.

Why do disaster damage surveys take so long?
The first reason is that the site is wide and the roads are blocked. Surveyors walk from house to house, take photos and fill in forms, and when access roads are cut by floods or landslides, that alone takes a day. The second is the way records are made. Each surveyor has slightly different criteria for naming damage, so if the same building is judged as complete collapse (fully collapsed) by one person and half collapse (partly collapsed) by another, someone has to go out again. The third is time pressure. If rain returns or residents start cleaning up, traces of damage disappear and the evidence becomes weaker. That is why you need a way to photograph first and judge later. In practice, matching the number of reports received with the number of on-site checks is the biggest bottleneck.
What is the procedure for recording a whole site with a drone?
The procedure has three steps: pre-flight checks, shooting and image stitching. First, check with the guidance of the responsible agency whether the area is a no-fly or restricted zone and whether approval is needed. When shooting, fly in a grid pattern at the same altitude and take photos that overlap each other generously. If the overlap is not enough, holes appear when stitching. These photos are merged in software such as WebODM into one orthoimage (a map-like photo seen from above without distortion). If you place ground control points (GCPs), markers whose positions you know, the positional accuracy improves. Menu names can differ by software version. With one orthoimage, you can zoom in and point out damage locations from the office, which reduces return visits to the site.
How does AI count damaged buildings and farmland?
When you cut the orthoimage into small tiles and feed them to a model, the model finds building roofs or flooded rice fields in each tile and draws their outlines. Finding and counting buildings one by one is called object detection, and attaching a class to every pixel to measure area is called segmentation. Buildings are often counted by number and farmland by area, so the two are used separately. If you have an image from before the disaster, you can also do change detection, comparing the two dates and extracting only what changed. The results come out as point or polygon data, so you can total them by district in GIS. However, these numbers are not confirmed figures. They are survey candidates. With a candidate list, it is easier for surveyors to decide which village to visit first.
Can I trust the numbers AI counts as they are?
No, you should not. A photo taken from above shows only the roof, so a building whose roof is intact but whose wall has collapsed, or one with only the first floor flooded, is read as normal. Houses hidden by trees and shadows are missed, and greenhouse or warehouse roofs are sometimes picked up as buildings by mistake. For farmland, it is hard to tell flooding from a paddy that was irrigated anyway. What you learn only by trying is the size of this error. Choose one area, create a point layer in QGIS with Layer > Create Layer > New GeoPackage Layer, mark the buildings people counted, and compare with the AI result, writing down the number missed and the number wrongly detected separately. With these ratios you decide how much to correct the overall figures and whether they can be used as grounds for judgment.
How do survey results lead to recovery support?
The results made by AI and drones do not immediately become support targets. The flow is candidate extraction, field confirmation, damage confirmation and support decision. Local government staff overlay the damage reports received and the candidates on the map, and send teams first to places where the two do not match. Confirmed damage is finalized according to the survey form, and the total becomes the basic data that decides the recovery plan and the scope of support. The support criteria and document forms differ by disaster type and local government, so you must check the guidelines of the relevant agency. Drone images remain as evidence showing the state at the time of judgment, and can be taken out again when an objection is filed. So you should record the shooting date and the survey date together.
Step by step on screen
The diagram shows the order of the steps and may use a Korean interface. Menu positions are the same in the English interface.

- Get the imagery
In your browser, go to openaerialmap.org > search for the area on the map > select an image and save it as GeoTIFF with Download. - Load the imagery
In QGIS, choose Layer > Add Layer > Add Raster Layer > select the file, click Add, then close. - Create a point layer
Layer > Create Layer > New GeoPackage Layer > set the geometry type to Point, add a text field named "type", then click OK. - Mark damaged buildings
Click the pencil icon (Toggle Editing) > use Add Point Feature and click on a building roof > enter "damaged" or "normal" in the type field, then save. - Count
Right-click the point layer > Open Attribute Table > sort by type and record the number people counted. - Compare with the AI result
Load the AI detection result for the same area and write the number of missed buildings and wrongly detected buildings in a table separately.
Menu names may differ slightly between program versions. The figure above is a diagram that summarizes the order.
Prompts you can copy
Compare a drone orthoimage with the AI detection result and get a draft verification report that summarizes the error
You are a GIS analysis assistant who helps the disaster damage survey staff of [organization name]. After [target disaster], AI detected buildings in a drone orthoimage, and I compared this with a manual count of the sample area [area name] made by people. Please summarize the verification result with the numbers below. Number of buildings counted by people: [number] Number of buildings detected by AI: [number] Number of buildings missed by AI: [number] Number of buildings wrongly detected by AI: [number] 1) Show the miss ratio and the wrong-detection ratio together with the calculation steps. 2) Suggest no more than three possible causes of the error, and for each write how I can check it in the imagery. 3) Write a paragraph of no more than 5 sentences reporting the result to [district-level staff member]. Constraints: do not make up numbers other than those I gave you, and write that you do not know where you do not know. Do not use wording in which AI judges damage grades or support eligibility.
Get a draft checklist for a drone photography plan to use before field surveys
You are an adviser on drone photography planning with practical disaster response experience at [organization name]. Right after [target disaster], I want to make an orthoimage of the damaged area of [area name]. Make a pre-flight checklist in table form. The columns are check item, how to check, person in charge, and done. The items to include are no-fly zones and approval status, weather and battery conditions, division of the shooting area, photo overlap settings, installation of control points, and recording the shooting date. Constraints: do not write law article numbers or permit criteria values; write only the types of agencies that should be consulted. Mark anything uncertain as "check with the agency". What a person must check: check that the calculated ratios match my numbers, that the possible causes AI gave can be seen in the actual imagery, and that the approval procedure in the checklist is the same as the guidance of the responsible agency.
Replace the bracketed parts for your situation. Do not use the answers as they are; check the facts first.
Try it yourself
- Search openaerialmap.org to see whether there is open drone imagery of your area or of a disaster area you know.
- Pick one area of an orthoimage, count the building roofs yourself, and write the numbers of normal and damaged buildings in a table.
- Check on the website of your local government or the responsible department what items are on the damage survey form.
Common questions
Do I need separate approval for a drone damage survey?
It depends on the flight location, the aircraft and the purpose of shooting. First check with the guidance of the responsible agency and official aviation information whether the area is a no-fly zone or controlled airspace, and follow the relevant guidelines even during disaster response.
Can AI also decide the damage grade of buildings?
A photo from above shows only the roof, so wall cracks or first-floor flooding are easy to miss. So use AI results only to narrow down survey candidates, and let a person who has confirmed the damage on site decide the grade.
Can I do a damage survey with satellite imagery instead of drones?
A satellite can see a wide area at once, but the image resolution, the shooting time and the cloud cover differ by satellite and period. A drone shoots a small area in detail but has to be flown by someone, so consider using both according to the extent of the damage.
A damage survey is not about taking many photos. It is about turning the recorded photos into grounds for judgment and support. A drone records the site quickly and AI narrows down where to look, but checking whether the numbers can be trusted and confirming the damage is the work of people. If you count a sample area yourself and measure the error, you get a criterion for how far to use the AI result. Then, how should the damage information confirmed by the survey be organized on a map so that it can be used right away to decide recovery priorities? A later post in this series covers this.
This is an English adaptation of a post in the Korean GeoAI basics series. The original is here, and more learning materials are on GeoAI for Disaster Safety.
This post was drafted with the help of AI tools and reviewed by the author. Tool features and data portals change over time, so check the official documentation when you practice.
Comments
Post a Comment