Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection

Tamara R. Lenhard; Andreas Weinmann; Tobias Koch

cs.CVAug 17, 2026

Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection

Authors: Tamara R. Lenhard, Andreas Weinmann, Tobias Koch

Organizations: Institute for the Protection of Terrestrial Infrastructures, German Aerospace Center (DLR), Sankt Augustin, Germany · ACIDA Lab, Technical University of Applied Sciences Würzburg-Schweinfurt, Schweinfurt, Germany

Abstract

Reliable drone detection under real-world deployment conditions requires training data that spans the full operational design domain, including adverse weather and seasonal appearance variation. However, acquiring and annotating such data at scale remains highly resource-intensive, as adverse-weather conditions are inherently difficult to control, reproduce, and sample systematically. Existing datasets therefore typically provide only limited coverage of such conditions. Conversely, synthetic data offers a scalable alternative: environmental variation becomes controllable, while modern game-engine-based pipelines provide realistic rendering and automatic annotations. Leveraging this potential, we introduce SynDroneVision-Weather (SDV-W), an systematic extension of SynDroneVision (SDV) targeting adverse-weather and seasonal domain shifts in urban drone detection. SDV-W comprises 55,187 annotated high-resolution images from three urban environments, rendered across three seasonal configurations and diverse weather conditions, including rain, snow, and fog at multiple severity levels. By preserving SDV's scene and trajectory configuration, SDV-W enables matched clean-adverse comparisons and quantification of condition-specific detector degradation. Across representative YOLO models and real-world datasets, we show that SDV-W improves detector reliability under adverse appearance shifts, reduces missed detections and false alarms, and is most effective as a complement to general-purpose synthetic drone-detection data. SDV-W will be publicly released upon paper acceptance.

Figures & tables

Figure 1: Example of synthetic rain augmentation. Left: clear reference frame from DUT Anti-UAV [ 50 ] (Apache-2.0). Right: the same frame from the rainy test set of [ 37 , 36 ] , with salient rain-streak artifacts and chromatic shifts.

Figure 2: Seasonal appearance variation in SDV-W. Identical scenes are re-rendered with season-specific configurations, altering vegetation, illumination, and atmospheric appearance while preserving scene geometry.

Figure 3: Weather-condition variation in SDV-W. Representative autumn (top) and winter (bottom) scenes are shown under sunny, rain, fog, rain-fog, and season-specific leaves/snow conditions. Non-sunny examples are shown at heavy severity.

Weather Cond.	Season			Total
	spring/summer	autumn	winter
rain	5,598	5,200	5,199	15,997
fog	4,400	5,200	5,197	14,797
rain $+$ fog	2,396	2,600	2,600	7,596
sunny	200	2,000	2,600	4,800
snow	–	–	6,000	6,000

Table 1: Composition of SDV-W across seasons and weather.

Size Category	Object Area ( $\text{pixels}^{2}$ )	Description
x-small (XS)	$A<16^{2}$	extremely small objects
small (S)	$16^{2}\leq A<32^{2}$	small objects
medium (M)	$32^{2}\leq A<96^{2}$	medium-sized objects
large (L)	$A\geq 96^{2}$	large objects

Table 2: Definition of object size categories based on bounding-box area

A=w\times h

. Categories follow the COCO size taxonomy, extended with an x-small class to better capture tiny drones [ 32 ] .

Figure 4: Distribution of object size categories in SDV-W. Share of x-small, small, medium, and large objects are shown w.r.t. its total number of annotated instances (N). Size categories follow Tab. 2 for the default YOLO input size of 640 $\times$ 640.

Figure 5: Weather-induced degradation of YOLO detection performance. Absolute deviations in mAP@0.25, FNR, and FDR are reported in percentage points (pp) relative to clear-weather performance. Larger values indicate greater degradation. Orange markers denote the mean across YOLO models. Horizontal intervals and $\pm$ annotations represent the corresponding standard deviation. — Figure 5: Weather-induced degradation of YOLO detection performance. Absolute deviations in [email protected], FNR, and FDR are reported in percentage points (pp) relative to clear-weather performance. Larger values indicate greater degradation. Orange markers denote the mean across YOLO models. Horizontal intervals and $\pm$ annotations represent the corresponding standard deviation.

		LRDDv1					DUT Anti-UAV
Model	w/	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	@0.25	@0.5	@0.5-0.95			@0.25	@0.5	@0.5-0.95
YOLOv5l	–	0.580	0.426	0.131	0.489	0.077	0.957	0.937	0.674	0.102	0.035
YOLOv5l	✓	0.700	0.566	0.204	0.486	0.030	0.936	0.928	0.723	0.108	0.031
YOLOv8m	–	0.606	0.442	0.135	0.480	0.059	0.954	0.932	0.680	0.106	0.033
YOLOv8m	✓	0.711	0.583	0.214	0.474	0.023	0.963	0.945	0.691	0.053	0.076

Table 3: Comparison of training with and without weather-condition synthetic data (SDV-W) on LRDDv1 and DUT Anti-UAV across multiple YOLO architectures and model scales. Best-performing values are highlighted in bold .

w/	Size	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
SDV-W	Cat.	@0.25	@0.5	@0.5-0.95
–	all	0.540	0.507	0.254	0.515	0.270
✓	all	0.561	0.530	0.258	0.491	0.235
–	XS	0.144	0.125	0.052	0.743	–
✓	XS	0.152	0.133	0.055	0.735	–
–	S	0.683	0.668	0.314	0.253	–

Table 4: Impact of weather-conditioned synthetic data (SDV-W) on the performance of YOLO-FEDER FusionNet on LRDDv2. Results are reported overall and for each size category defined in Tab. 2 . Best-performing values are highlighted in bold .

		Summer					Winter
Dataset	w/	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	@0.25	@0.5	@0.5-0.95			@0.25	@0.5	@0.5-0.95
R1-POS3	–	0.764	0.743	0.331	0.313	0.083	0.931	0.827	0.404	0.208	0.006
R1-POS3	✓	0.821	0.794	0.354	0.326	0.004	0.965	0.836	0.393	0.164	0.004
R2-POS7	–	0.985	0.909	0.450	0.076	0.010	0.983	0.749	0.294	0.025	0.054
R2-POS7	✓	0.982	0.901	0.422	0.061	0.008	0.988	0.675	0.223	0.028	0.006

Table 5: Performance of YOLO-FEDER FusionNet with and without weather-conditioned synthetic training data (SDV-W) under seasonal conditions (summer vs. winter) on the datasets R1-POS3 and R2-POS7. Results are aggregated across all size categories. Size-specific performance indicators are provided in the supplementary material. Best-performing values are highlighted in bold .

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Environment
University Site	Venetian City [ 8 ]	Urban Downtown [ 39 ]
seq0027	seq0009	seq0001
seq0062	seq0011	seq0047
seq0063	seq0014	seq0048
seq0070	seq0016	seq0051
seq0072	seq0017

Appendix

Table I : SDV source sequences underlying the generation of SDV-W, grouped by their corresponding environments.

Figure I : Paired SDV/SDV-W rendering example. The original SDV frame is shown above, and the corresponding SDV-W rendering under rainy autumn conditions is shown below.

Material	Albedo_1_ColorOverlay				Albedo_Brightness	Albedo_Add
	R	G	B	A
MI_Foliage_Tree_Pine_Leaves	0.617	0.255	0.000	9.000	1.756	0.198
MI_Tree_Narrowleaf_Leaf_A	1.000	0.785	0.000	$-$ 0.091	1.522	$-$ 0.581
MI_Foliage_Trre_Leaves	0.976	0.718	0.062	0.000	–	–

Appendix

Table II : Material-level foliage recoloring parameters for the Urban Downtown [ 39 ] autumn configuration.

Season	Weather	Solar Time		Direct. Light Intensity		Rayleigh Scattering (Channel Values)
	Cond.			(lux)		R		G		B
		from	to	min	max	min	max	min	max	min	max
spring/summer	sunny	11.44	12.19	6.4	10.0	0.175	0.175	0.410	0.410	1.000	1.000
	rain	9.57	20.13	0.5	9.0	0.000	0.806	0.192	0.861	0.212	1.000
	fog	6.59	11.64	1.0	8.0	0.028	1.000	0.047	0.834	0.553	1.000
	rain+fog	9.57	20.13	0.5	9.0	0.175	0.806	0.192	0.861	0.212	1.000

Appendix

Table III : Sun and Sky Actor [ 16 ] parameter ranges for SDV-W lighting variation across University Site and Venetian City [ 8 ] , aggregated over weather severity levels.

Season	Weather	Direct. Light
	Cond.	Int. (lux)		Rot. Z (deg)
		min	max	min	max
spring/summer	rain	2.0	4.0	20	300
	fog	1.0	20.0	20	189
	rain+fog	2.0	4.0	20	320
autumn	sunny	1.0	3.0	83	200

Appendix

Table IV : Directional Light [ 12 ] parameter ranges for Urban Downtown [ 39 ] . Intensity (Int.) and rotation (Rot.) are reported for each season-weather condition and aggregated over weather severity levels. Rotation varies only along the Z axis, while X and Y are fixed at their default settings of 0 ∘ and

-

46 ∘ , respectively.

Figure II : Statistical characterization of bounding box annotations in SDV-W. From left to right: spatial distribution of bounding box centers, joint distribution of normalized width and height, distribution of aspect ratios, and distribution of relative bounding box areas. Here, the object area ratio denotes the bounding-box area relative to the image area.

Season	Weather	Temp		Tint
	Cond.	min	max	min	max
spring/summer	sunny	4,599	5,877	0.00	0.00
	rain	4,300	11,240	$-$ 0.09	0.00
	fog	4,300	11,439	0.00	0.00
	rain $+$ fog	4,300	11,240	$-$ 0.09	0.00
autumn	sunny	3,900	8,768	0.00	0.00

Appendix

Table V : Season- and weather-specific Post Process Volume [ 13 ] settings. Temp and Tint denote min-max ranges across all environments and weather severity levels.

Dataset	Size	Image Scale		$\Delta$ Share
	Cat.	original	640 $\times$ 640	(pp)
DUT Anti-UAV [ 50 ]	XS	87	569	$+$ 21.91
	S	764	674	$-$ 4.09
	M	808	482	$-$ 14.82
	L	541	475	$-$ 3.00
LRDDv1 [ 42 ]	XS	2,055	6,607	$+$ 20.37

Appendix

Table VI : Effect of input resizing on object-size category distribution across real-world datasets. Counts are reported before and after resizing to an input resolution of 640

\times

640 pixels.

\Delta

Share denotes the resizing-induced change in each category’s relative frequency expressed in percentage points (pp). “–” indicates that no objects of the corresponding size category are present.

Figure III : Distribution of object size categories across public and custom drone detection datasets. Top: Public datasets DUT Anti-UAV [ 50 ] , LRDDv1 [ 42 ] , and LRDDv2 [ 41 ] . Bottom: Custom datasets R1-POS3 and R2-POS7 [ 31 , 30 ] under summer and winter conditions. For each dataset, the proportion of x-small, small, medium, and large objects is shown relative to the total number of annotated instances ( $N$ ). Size categories are defined according to Tab. 2 (main paper) w.r.t. an input image size of 640 $\times$ 640.

Figure IV : Weather-induced degradation of YOLO detection performance. Absolute reductions in mAP@0.5 (top) and mAP@0.5-0.95 (bottom) are reported in percentage points (pp) relative to clear-weather performance for each weather condition. Larger values indicate greater degradation. Orange markers denote the mean degradation across YOLO model configurations. Horizontal intervals and $\pm$ annotations represent the corresponding standard deviation. — Figure IV : Weather-induced degradation of YOLO detection performance. Absolute reductions in [email protected] (top) and [email protected] (bottom) are reported in percentage points (pp) relative to clear-weather performance for each weather condition. Larger values indicate greater degradation. Orange markers denote the mean degradation across YOLO model configurations. Horizontal intervals and $\pm$ annotations represent the corresponding standard deviation.

			LRDDv1					DUT Anti-UAV
Model	w/	Size	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	Cat.	@0.25	@0.5	@0.5-0.95			@0.25	@0.5	@0.5-0.95
YOLO-	–	all	0.772	0.578	0.192	0.310	0.056	0.979	0.967	0.721	0.045	0.032
FEDER	✓	all	0.768	0.571	0.190	0.314	0.054	0.981	0.967	0.716	0.040	0.032
	–	XS	0.213	0.175	0.060	0.553	–	0.876	0.857	0.500	0.086	–
	✓	XS	0.198	0.169	0.057	0.583	–	0.886	0.866	0.498	0.077	–

Appendix

Table VII : Impact of weather-conditioned synthetic data (SDV-W) on the performance of YOLO-FEDER FusionNet on LRDDv1 [ 42 ] and DUT Anti-UAV [ 50 ] . Results are reported overall and for each size category defined in Tab. 2 (main paper). Best-performing values are highlighted in bold ; tied best values are underlined .

YOLO	w/	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	@0.25	@0.5	@0.5-0.95
v5l	–	0.517	0.495	0.255	0.659	0.091
v5l	✓	0.530	0.509	0.266	0.653	0.083
v8m	–	0.534	0.514	0.266	0.630	0.108
v8m	✓	0.519	0.504	0.266	0.642	0.084
v8l	–	0.516	0.498	0.252	0.646	0.098

Appendix

Table VIII : Comparison of training with and without weather-condition synthetic data (SDV-W) on LRDDv2 [ 41 ] across multiple YOLO architectures and model scales. Best-performing values are highlighted in bold , while tied best values are underlined .

			Summer					Winter
Dataset	w/	Size	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	Cat.	@0.25	@0.5	@0.5-0.95			@0.25	@0.5	@0.5-0.95
R1-POS3 [ 31 , 30 ]	–	all	0.764	0.743	0.331	0.313	0.083	0.931	0.827	0.404	0.208	0.006
	✓	all	0.821	0.794	0.354	0.326	0.004	0.965	0.836	0.393	0.164	0.004
	–	XS	0.364	0.358	0.175	0.555	–	0.596	0.572	0.250	0.206	–
	✓	XS	0.373	0.348	0.160	0.575	–	0.644	0.634	0.271	0.191	–

Appendix

Table IX : Performance of YOLO-FEDER FusionNet with and without weather-conditioned synthetic training data (SDV-W) under seasonal conditions (summer vs. winter) on the datasets R1-POS3 and R2-POS7 [ 31 , 30 ] . Results are reported overall and for each size category defined in Tab. 2 (main paper). Best-performing values are highlighted in bold ; tied best values are underlined . “–” indicates that no objects of the corresponding size category are present.

Dataset	w/	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	SDV-W	@0.25	@0.5	@0.5-0.95
LRDDv1 [ 42 ]	–	0.772	0.578	0.192	0.310	0.056
	✓ R	$0.774\pm 0.008$	$0.577\pm 0.010$	$0.192\pm 0.005$	$0.315\pm 0.006$	$0.053\pm 0.011$
	✓ A	0.768	0.571	0.190	0.314	0.054
LRDDv2 [ 41 ]	–	0.540	0.507	0.254	0.515	0.270
	✓ R	$0.551\pm 0.002$	$0.516\pm 0.004$	$0.253\pm 0.002$	$0.510\pm 0.006$	$0.239\pm 0.022$

Appendix

Table X : Ablation study disentangling the effects of weather diversity and training set size for YOLO-FEDER FusionNet. Weather-conditioned synthetic data (SDV-W) either substitute a subset of SDV samples or extend the original training set, consisting of SDV and DUT Anti-UAV [ 50 ] .

Tiling	Size	mAP $\uparrow$			FNR $\downarrow$	FDR $\downarrow$
	Cat.	@0.25	@0.5	@0.5-0.95
–	all	0.561	0.530	0.258	0.491	0.235
✓	all	0.669	0.580	0.245	0.305	0.426
–	XS	0.152	0.133	0.055	0.735	–
✓	XS	0.407	0.312	0.108	0.434	–
–	S	0.705	0.690	0.311	0.227	–

Appendix

Table XI : Impact of tiled inference relative to resize-based inference for YOLO-FEDER FusionNet on LRDDv2 [ 41 ] . Best-performing values are highlighted in bold .

Explore similar work

Aug 18, 2026cs.CV

Training with synthetic data for drone detection in thermal imagery

Ground-to-Air (G2A) drone detection in medium- and long-wave infrared (MWIR/LWIR) imagery is challenging due to reduced texture information, sensor noise, weak thermal contrast, and the scarcity of annotated data. This work investigates a synthetic-first training strategy that combines synthetic scene generation with fine-tuning on real data. We show that synthetic data provides an effective basis for learning initial object representations, while real in-domain thermal imagery is still essential for reliable deployment. Even small amounts of real IR data substantially reduce domain gaps. Our experiments indicate that dataset alignment has a stronger impact on performance than model scale. Finally, our analysis of the dataset suggests that semantic alignment in feature space is the strongest predictor of model performance, while radiometric properties such as entropy and dynamic range also contribute to detection robustness. This work provides a foundation for combining synthetic and real IR data for effective G2A drone detection.

Tanel Liiv, Sander Soodla, Nzamba Bignoumba +2

Marduk Technologies OÜ, Tallinn, Estonia · Intelligent Imaging, TNO Defence, Safety and Security, The Hague, The Netherlands

Infrared Object Detection Synthetic Data Pretraining Synthetic-to-Real Domain Adaptation Synthetic Data Generation

Ask PDF

Sep 7, 2026cs.CV

DroneGround: Open-Vocabulary Drone Payload Characterization Using Synthetic Data and Grounded Vision-Language Models

Automated drone surveillance has become increasingly important for public safety, critical infrastructure protection,and restricted airspace monitoring. While existing vision-based systems achieve strong performance for drone detection and tracking, reliable payload characterization remains highly challenging under long-range imaging conditions due to limited availability of annotated real-world datasets, and substantial distribution shifts encountered during deployment. Existing approaches formulate payload characterization as a closed-set object detection problem, limiting their ability to recognize previously unseen payloads and generalize beyond the training distribution. To address these challenges, we generate a photorealistic synthetic drone-payload dataset using Unreal Engine 5 and Cosys-AirSim and propose DroneGround: Grounded Vision-Language Payload Characterization, a two-stage framework for robust open-vocabulary payload analysis. DroneGround first employs a YOLO26s detector to localize drones and extract drone-centric image crops, followed by a LoRA-fine-tuned PaliGemma vision-language model that generates seman- tic descriptions of the detected drones and their attached payloads, enabling open-vocabulary payload characterization beyond predefined categories. An occlusion-based grounding module further provides interpretable payload localization by identifying image regions responsible for the generated descriptions. Extensive experiments on both synthetic and real-world drone imagery demonstrate that DroneGround substantially improves robustness under synthetic-to-real distribution shifts, outperforming a conventional closed-set payload detector by improving the F1-score from 82.5% to 96.3%, while achieving significantly better generalization to previously unseen payload categories (80.4%versus 42.7% F1). Dataset and code will be released upon request.

Ami Pandat, Rajasekhar Punna, Gopika Vinod +1

Homi Bhabha National Institute, Mumbai, Maharashtra, India · Bhabha Atomic Research Centre, Mumbai, Maharashtra, India

Open-Vocabulary Object Detection Vision-Language Grounding

Ask PDF

Jul 2, 2026cs.CV

Diagnosing Aerial-View Object Detectors with Foundational Image Generative Models

Recent advances in large-scale image generative models enable photorealistic scene synthesis with controllable attributes. Beyond data augmentation, their potential as diagnostic tools for trained vision systems remains unexplored in the aerial and remote sensing domains. We introduce a synthetic diagnostic framework for aerial-view vehicle detection that combines text-guided generation, attribute-controlled editing, and automated attribute verification to construct a controllable synthetic testbed. This enables fine-grained evaluation of pretrained detectors under diverse scene types and environmental conditions that are difficult to isolate in real datasets. Across three detection architectures and three real aerial datasets, synthetic scene-wise performance trends closely match real-world weaknesses. Guided by these diagnostics, targeted supplementation with small real datasets from the identified weak categories yields improvements of up to 13% AP50 while requiring substantially fewer additional samples than non-targeted augmentation. Our results show that controlled synthetic probing can predict real-domain performance gaps and guide efficient data collection. The proposed diagnostic framework is modular and can incorporate alternative generative or vision-language models as capabilities evolve. Our code and datasets are available here: https://humansensinglab.github.io/AVODDiag/

Stanislav Panev, Minhyek Jeon, Vaishnavi Khindkar +5

Carnegie Mellon University, USA · DEVCOM Army Research Laboratory, USA · Florida State University, USA

Synthetic Data Augmentation Synthetic Data Generation Aerial Object Detection

Ask PDF