Pteropus medius · drone thermal survey · Bangladesh · 8–19 Aug 2025

Counting flying foxes by their body heat

The Indian flying fox is the reservoir of Nipah virus in Bangladesh. We flew a thermal drone in orbits around its roost trees in Naogaon, Rajshahi and Faridpur, detected every bat in the footage, de-duplicated repeat passes, and read each animal's body-surface temperature.

Roost atlas

Every roost we flew

Circles are sized by census count and coloured by mean body-surface temperature. To protect the roosts, each circle is drawn several kilometres wide around a shifted centre, so it shows the area a roost is in, not the tree. Select a roost to see its orbit video and readings. Some roosts were flown more than once, so the panel lists every flight at that site.

27 °C35 °C

no video census (stills only)

One orbit per flight

Roost videos

Each clip is the drone's first complete 360° pass around the roost (the camera heading turning a full circle), sped up, with the detector's boxes and a running count of unique bats. Where no orbit closed inside the tracked footage, the clip shows the busiest three minutes and is labelled as a segment.

Body-surface temperature

Temperatures and counts by flight

Body-surface temperature by flight

Select any row for a summary. Line spans the minimum to maximum reading; the bar is the mean ± 1 SD; the dot is the mean. Sorted warmest to coolest. Video readings are calibrated estimates, still-photo readings are radiometric measurements.

Census: every pass vs first orbit

Grey is every unique track in the video; colour is the corrected census. Repeat laps re-count the same colony.

Mean temperature by local time of flight

Each roost was flown in one short session, so time of day and roost are tangled. Read this as a caution, not a daily cycle.

How the numbers were made

Methods and limits

Aircraft. DJI Matrice 4T, paired thermal and RGB video plus radiometric thermal stills. Times are Bangladesh local (UTC+6), corrected from an aircraft clock left on US Central time.
Detection. A custom RF-DETR model (Roboflow) finds bats in each frame; a tracker gives each bat an ID across brief occlusions.
Census. The tracker does not re-identify a bat on the next lap (0–5% cross-lap re-identification on six multi-lap flights). The census therefore counts only the first complete orbit. Flights that never closed an orbit are left uncorrected and should be read as upper bounds.
Temperature. Radiometric stills give measured temperatures (DJI Thermal SDK v1.8). Video carries no radiometric data, so video sightings are estimated from a per-roost colour calibration built on those stills.

Sites vs flights. The survey logged 24 flights. Several were repeat visits to the same trees: flights 008, 009 and 011 are one roost, and so are 003 + 005, 004 + 010, 014 + 016, and 020 + 023 + 035 + the unfiled flight. This page groups flights within 400 m into one site, giving sites.
Totals. The headline census adds the largest single-flight count at each site (). Adding every flight gives , which counts repeat-visited roosts more than once.
Anomalies. A bat is flagged warm or cool only against its own roost. Flags are a screening step, not a diagnosis; a cool “bat” can be a false detection on cold background.
Location privacy. Roost positions are not published. Each site appears as a circle at least 6 km across whose centre is shifted 1–3 km from the roost, the map stops zooming at district scale, and the drone's GPS readout is masked in every clip and still.
Coverage. Five flights (021–024, 035) have radiometric stills but no tracked video, so they have temperatures but no census.