dejanb/beograd

PES2026 — UAV Thermal Imagery for Buried-Object Anomaly Detection

825 radiometric thermal images (16-bit, 640×512) captured by a DJI Zenmuse H20T over a 27 × 35 m plot near Belgrade, Serbia, on 15–16 July 2024. Every image has a photogrammetrically determined flight altitude and RTK position.

This is the subset of a larger capture that falls inside the 1–15 m altitude range usable for detecting objects of ~10 cm. Companion dataset to the PES 2026 paper on convolutional-autoencoder detection of thermal anomalies.

⚠️ Safety. This dataset supports research. Thermal analysis yields a list of candidates for physical verification, never a confirmation that terrain is safe. Absence of a thermal signal is not evidence of absence of an object. Declaring land safe is the exclusive responsibility of accredited demining organisations following established procedure.

---

What makes this dataset unusual

Every image carries a verified altitude. This is rare in UAV thermal datasets and it matters: altitude fixes ground sample distance, which fixes whether an object of a given size is physically resolvable at all. Reference literature reports 97.1 % detection success at 1 m versus 88.8 % at 2–10 m — altitude is the primary sensitivity axis, not a footnote.

Obtaining it was not straightforward. None of the three altitude metadata fields the camera records is reliable on its own:

Source	What it actually measures	Failure mode observed here
Barometer (RelativeAltitude)	height above take-off point	off by +7.8 / +8.3 m where take-off was below the plot
Laser rangefinder (LRFTargetDistance)	distance to ground	unusable below its ~3 m minimum; returned Normal status with wrong values
RTK (AbsoluteAltitude) − terrain	absolute height	requires an independent terrain elevation, which was wrong by ~1.3 m in places

Altitude was therefore derived photogrammetrically, using no altitude metadata at all:

AGL = (d_RTK / Δ_px) · (f / p) = (d_RTK / Δ_px) · (13.5 mm / 12 µm)

where d_RTK is the baseline between consecutive exposures (RTK, 3–5 cm) and Δ_px the scene shift measured by ORB feature matching + RANSAC. Validated against the laser on seven sessions where the laser operates within specification: deviation −2.4 % to +0.8 % across 9.7–26.4 m.

Per-image reliability, against the laser reference:

Independent estimates	Images	Median |laser − photogrammetry|
2	596 (47.8 %)	0.33 m
1	416 (33.4 %)	0.47 m
0 (interpolated)	234 (18.8 %)	0.65 m

The AGL_quality column marks which images are measured and which interpolated.

---

Contents

pes2026/
├── manifest.csv          825 rows, one per image
└── data/
    ├── s2/   118 images   9.72 m AGL   8.6 mm/px
    ├── s8/   110 images   1.74 m AGL   1.55 mm/px
    └── s9/   597 images   1.90 m AGL   1.69 mm/px

Images are 16-bit single-channel GeoTIFF, georeferenced to EPSG:32634 (WGS 84 / UTM 34N) with a ModelTransformation tag that includes gimbal yaw rotation. They open directly in QGIS.

Sessions

Session	Local time (CEST)	n	AGL	GSD	Footprint	Overlap	Daylight
s2	15 Jul 19:56–20:00	118	9.72 m	8.6 mm/px	5.5 × 4.4 m	63 %	before sunset
s8	16 Jul 07:32–07:38	110	1.74 m	1.55 mm/px	0.99 × 0.79 m	55 %	morning
s9	16 Jul 08:59–09:22	597	1.90 m	1.69 mm/px	1.08 × 0.86 m	52 %	morning

Sunset on 15 July was ≈ 20:25, sunrise ≈ 05:10.

manifest.csv columns

Column	Meaning
path	relative path to the GeoTIFF
session	2, 8 or 9 — use this for train/test splits
datetime	ISO 8601, local time (+02:00 CEST)
AGL_m	altitude above ground, photogrammetric — the authoritative value
AGL_quality	merena (measured) or interpolirana (interpolated)
GSD_mm_px	ground sample distance
footprint_w_m, footprint_h_m	ground footprint of the frame
lat, lon, abs_alt_m	RTK position (3–5 cm horizontal, 6–10 cm vertical)
gimbal_yaw	gimbal yaw in degrees; pitch is −90° (nadir) throughout
raw_min, raw_max, raw_mean	raw-value statistics of the image
lrf_dist_m, lrf_status	laser rangefinder reading, kept as an independent check
rel_alt_m	barometric altitude above take-off — retained for comparison, do not use
n_est_foto	number of independent photogrammetric estimates for this image

---

Pixel values are NOT temperature

Values are uncalibrated sensor counts (range 0–24 790 across the set), extracted from the APP3 segment of the original DJI R-JPEG. Converting to °C requires DJI's proprietary calibration coefficients via the closed DJI Thermal SDK. No measured temperature is recorded anywhere in the source files.

An uncooled microbolometer responds approximately linearly, so differences between raw values are proportional to temperature differences. For anomaly detection — where a patch warmer or cooler than its surroundings is the signal — this is sufficient; only the scaling constant is missing.

Two normalisation cautions

1. Normalise globally, not per image. Per-image normalisation stretches every frame to its own range, so a frame containing a strong anomaly and one without end up with identical statistics and the difference of interest is destroyed.
2. Do not compare raw values across sessions without care. The sensor's calibration state changes between flights (a per-flight field in the APP4 segment takes exactly one value per session). Within a session values are directly comparable; across sessions there is an unknown offset. Split train/test by session.

---

Loading

import pandas as pd, numpy as np
from PIL import Image
from huggingface_hub import snapshot_download

root = snapshot_download("dejanb/pes2026", repo_type="dataset")
df   = pd.read_csv(f"{root}/manifest.csv")

# session split avoids both spatial leakage and the cross-session calibration offset
train = df[df.session == 9]
test  = df[df.session == 2]

img = np.array(Image.open(f"{root}/{train.iloc[0].path}"))   # uint16, (512, 640)
print(img.dtype, img.shape, img.min(), img.max())

Global normalisation, fitted on the training split only:

lo, hi = np.percentile(
    np.concatenate([np.array(Image.open(f"{root}/{p}")).ravel() for p in train.path]),
    [1, 99])
def norm(x): return np.clip((x.astype(np.float32) - lo) / (hi - lo), 0, 1)

---

Acquisition

Camera	DJI Zenmuse H20T, uncooled microbolometer 640 × 512
Optics	f = 13.5 mm, F1.0, 12 µm pixel pitch, FOV 31.8° × 25.6°
Gimbal	nadir, −90.0° ± 0.1°, roll 0
Positioning	RTK fixed solution throughout
Site	44.660172 N, 20.348805 E; terrain ≈ 236.9 m a.s.l.; extent 27 × 35 m
Date	15–16 July 2024

Thermal parameters in the source files were left at factory defaults (distance 5 m, humidity 70 %, emissivity 1.00, reflected temperature 23.0 °C). These affect absolute temperature but very little the within-frame relative contrast that anomaly detection uses.

Provenance

Source files were DJI R-JPEGs in which a block of NUL bytes (median 1.0 MB) preceded the JPEG data, making them unreadable by standard tools. The prefix was removed and the 16-bit APP3 layer extracted; correctness of byte order, dimensions and orientation was confirmed by a correlation of 0.9735 against the embedded 8-bit preview. No pixel data was altered.

Limitations

• No ground truth. Target positions on the site are not known, so detection performance cannot be computed from this dataset alone — only candidate lists. For labelled evaluation see the Tenorio-Tamayo dataset below.
• No night imagery in the usable range. The full capture includes 421 further images from 21:42–22:04 and 07:20–07:29, but all at 21–28 m, where an 8.7 cm object spans under 5 pixels. Night is the thermally most favourable window, so this is the main gap.
• Altitude is a rolling median within each session, hence locally smoothed; it would not capture an abrupt genuine altitude change.
• Georeferencing is approximate: it assumes ideal nadir and flat terrain, with no lens distortion model.
• The altitude sensitivity axis is covered at only two points (≈1.8 m and ≈9.7 m) — enough to compare extremes, not to trace a curve.

Citation

@dataset{pes2026_thermal,
  author    = {Blagojević, Dejan and Batanjac, Dejan and Glamočlija, M.},
  title     = {PES2026 — UAV Thermal Imagery for Buried-Object Anomaly Detection},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/dejanb/pes2026},
  license   = {CC-BY-4.0}
}

Related

• Reference dataset with labels — D. Tenorio-Tamayo et al., Dataset of thermographic images for the detection of buried landmines, Data in Brief 49 (2023) 109443, doi:10.17632/732ngnf4r3, CC BY 4.0. Zenmuse XT, 336 × 256, known mine positions and burial depths (0/1/5/10 cm), altitudes 1–10 m.
• Altitude sensitivity baseline — J. Forero-Ramírez et al., Detection of "legbreaker" antipersonnel landmines by analysis of aerial thermographic images of the soil, Infrared Physics & Technology 125 (2022) 104307.
• Checkerboard-free decoder — C. Olah et al., Deconvolution and Checkerboard Artifacts, Distill (2016), https://distill.pub/2016/deconv-checkerboard/.