Adaptive Multispectral Landmine Dataset (AMLID)

• AMLID is a comprehensive dataset of 12,078 annotated images featuring 21 landmine types captured via synchronized RGB and LWIR sensors on UAVs.
• It employs 11 fusion levels and varied sensor parameters including four altitudes, seasonal, and diurnal variations to assess detection performance.
• The dataset supports rigorous benchmarking with YOLO-based evaluations, achieving up to 0.80 mAP, while addressing issues like class imbalance and domain specificity.

The Adaptive Multispectral Landmine Identification Dataset (AMLID) is the first open-source dataset created for unmanned aerial systems (UAS)-based landmine detection that systematically combines Red-Green-Blue (RGB) and Long-Wave Infrared (LWIR) imagery, supporting rigorous development and benchmarking of adaptive detection algorithms without the need for hazardous fieldwork or expensive data collection infrastructure. AMLID encompasses 12,078 expertly annotated images depicting 21 globally representative landmine types—across anti-personnel and anti-tank classes, and both metal and plastic construction—acquired over 11 RGB-LWIR fusion ratios, four discrete sensor elevations, two distinct seasons, and three times of day (Gallagher et al., 21 Dec 2025).

1. Acquisition Protocols and Sensor Instrumentation

AMLID was generated using coordinated RGB and LWIR image capture on a DJI Inspire 2 drone, with a custom 3D-printed housing ensuring robust mechanical co-registration and minimized parallax artifacts. The RGB imagery was collected by a RunCam 5 (Sony IMX377 sensor, 1920×1080 px, 12 MP, 145° FOV, 400–700 nm spectral response). LWIR was acquired using a FLIR Vue Pro R radiometric camera (336×256 px, 6.8 mm lens, 45° FOV, 7.5–13.5 µm spectral band).

Flights were conducted at four fixed altitudes—5 m, 10 m, 15 m, and 20 m above ground—allowing for controlled assessment of spatial resolution and atmospheric effects on both sensor modalities. The effects of increasing altitude include degradation in ground sample distance (GSD) and a reduction in thermal target contrast for small landmine objects. Seasonal and diurnal coverage comprised January (winter) and May (early summer) in Leavenworth, KS, at three illumination conditions: post-sunrise, noon, and pre-sunset. Table 1 summarizes measured environmental conditions during image acquisition:

Condition	January (°C, lux)	May (°C, lux)
Post-sunrise	Ground –10.9°C, Air 9.9°C, 3,716 lx	Ground 21.4°C, Air 22.6°C, 11,500 lx
Noon	Ground 8.5°C, Air 20.7°C, 36,300 lx	Ground 33.5°C, Air 27.4°C, 83,600 lx
Pre-sunset	Ground 2.8°C, Air 18.9°C, 5,498 lx	Ground 34.4°C, Air 24.7°C, 31,300 lx

Average January air and ground temperatures were 16.5°C and 0.13°C, respectively; May averages were 29.8°C and 29.76°C. Illumination spanned an order of magnitude between daily periods, substantially affecting both LWIR thermal contrast and RGB scene shadowing (Gallagher et al., 21 Dec 2025).

2. Multispectral RGB-LWIR Fusion Schema

AMLID introduces 11 systematic levels of RGB-LWIR fusion, achieved by spatially registering and then overlaying the LWIR channel onto the RGB base in 10% increments. The fusion operation for each pixel $(x, y)$ is defined as:

$$I_{fused}(x, y; \alpha) = (1 - \alpha) \cdot I_{RGB}(x, y) + \alpha \cdot I_{LWIR}(x, y),$$

with $\alpha \in \{0.0, 0.1, 0.2, ..., 0.9, 1.0\}$. Here, $\alpha=0.0$ corresponds to pure RGB, while $\alpha=1.0$ yields pure LWIR. Only the LWIR contribution is varied; the RGB baseline remains unchanged. This design allows controlled evaluation spanning the spectrum from spatial to thermal-dominated imagery, supporting detailed analysis of fusion benefits and modality selection under varying environmental stressors (Gallagher et al., 21 Dec 2025).

3. Dataset Composition and Annotation Framework

AMLID comprises 12,078 images, with a train/test split of 10,758 (89%) and 1,320 (11%), respectively. The dataset incorporates 21 distinct landmine simulant types:

• Anti-Tank, Plastic (10 types): TM-62, TM-46, TMP-3, TC/2.4, VS-HCT2, SB-81, M19, TC-3.6, TC-6, CV-PT-MI-BA
• Anti-Tank, Metal (2): TM-57, MON-100
• Anti-Personnel, Plastic (6): POMZ-2M, VS-MK2, VS-50, PP MI-NA, PMN, APERS NO.4
• Anti-Personnel, Metal (3): OZM-72, OZM-3, Type-69

Detection annotations are grouped into four classes:

Class Index	Category	Target Description
0	AT-plastic	Anti-tank, plastic
1	AT-metal	Anti-tank, metal
2	AP-plastic	Anti-personnel, plastic
3	AP-metal	Anti-personnel, metal

Annotations are provided in both Pascal VOC XML and YOLO TXT formats, using axis-aligned bounding boxes specified by normalized (center-x, center-y, width, height) coordinates in the [0,1] range. 100% of images were manually annotated, with a random 10% quality-controlled by a second expert. The class distribution is heavily imbalanced; for the 14,905 test-set instances: AT-plastic accounts for 67.3%, AT-metal 11.9%, AP-plastic 12.0%, and AP-metal 8.9% (Gallagher et al., 21 Dec 2025).

4. Data Organization, Preprocessing, and Benchmarking Protocols

Image files are stored in JPEG format at native sensor resolution (1920×1080 for RGB, 336×256 for LWIR). The directory structure supports stratified access by season, illumination period, altitude, and fusion level:

AMLID/ ├─ train/ │ ├─ January/ │ │ ├─ post-sunrise/alpha_0.0/ │ │ ├─ post-sunrise/alpha_0.1/ │ │ ... │ │ └─ pre-sunset/... │ └─ May/... └─ test/ ├─ January/post-sunrise/alpha_* ├─ January/noon/... └─ ... (each with 55 images at 4 altitudes)

Training images span all combinatorial factors, with the 1,320 test images equally allocated across 24 season-time-altitude cells (55 per cell). Recommended preprocessing includes oversampling of January images and weighted losses to address season imbalance, as well as domain adaptation for terrain generalization due to site homogeneity. Data augmentation strategies include random flips, moderate brightness/contrast shifts, and small rotations, especially for fusion studies (Gallagher et al., 21 Dec 2025).

Evaluation metrics align with prevailing object detection protocols:

• Precision $\frac{TP}{TP + FP}$
• Recall $\frac{TP}{TP + FN}$
• Per-class Average Precision (AP):

$AP_c = \int_{0}^{1} \text{precision}(r) dr$

• Mean Average Precision (mAP) across $C$ classes:

$\mathrm{mAP} = \frac{1}{C} \sum_{c=1}^{C} AP_c$

5. YOLO Detection Benchmarks and Empirical Insights

Twenty-seven YOLO configurations—covering three architectures and all temporal partitions—were trained on the full set of 10,758 fusion-level training images and evaluated on the comprehensive, balanced test set. The principal outcome measure is $\mathrm{mAP}_{0.5}$ (mAP at IoU $\geq 0.50$).

Key observations:

• Pure RGB ($\alpha=0.0$) is optimal under high illumination (noon) and low altitude (5 m), but detection rates drop in low light or suboptimal thermal conditions.
• Pure LWIR ($\alpha=1.0$) maintains detection capabilities during post-sunrise and in the colder January environment, but experiences significant resolution loss at altitudes $\geq 15$ m.
• The optimal fusion coefficient $\alpha^*$ for detection is environment-dependent, with $\alpha \approx 0.6$ at January/noon, $\alpha \approx 0.4$ at May/pre-sunset, and so on.
• Peak $\mathrm{mAP}_{0.5}$ is approximately $0.75$–$0.80$ for best-case fusion/altitude scenarios, declining to $\approx 0.50$ in the worst case such as pure RGB at January/post-sunrise (Gallagher et al., 21 Dec 2025).

6. Research Applications and Limitations

AMLID provides 11 fusion intervals, four UAV altitudes, seasonal/diurnal diversity, and broad landmine type coverage, facilitating competitive evaluation of state-of-the-art detectors including YOLO, Faster-R-CNN, and vision transformers. The comprehensive annotation and standardized structure enable reproducibility across a spectrum of multispectral and environmental adaptation strategies.

Known constraints include pronounced class and seasonal imbalance, and domain specificity (single test site with homogeneous soil background). This suggests a requirement for domain adaptation and/or targeted resampling for applications in other contexts. Soil, vegetation, and climatic diversity are not explicitly captured, so cross-domain generalization studies are recommended.

AMLID and all associated annotation files are publicly available (Zenodo DOI: 10.5281/zenodo.18001447). The dataset constitutes a foundational resource for UAS-based, multispectral, humanitarian demining research (Gallagher et al., 21 Dec 2025).