Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Beyond Clear Skies".
Jane: Here's the gist of the paper. 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 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 Unmanned aerial vehicles (UAVs) have evolved from specialized military systems into versatile aerial platforms, supporting applications in logistics, agriculture, infrastructure inspection, surveillance, and recreation Reliable drone detection is essential for low-altitude airspace surveillance, enabling timely localization and risk mitigation The reliability of DL-based detectors is inherently tied to the diversity and representativeness of their training data Conditions outside this domain can induce substantial performance degradation Existing datasets often provide limited coverage of these scenarios, constraining their ability to support robust generalization under deployment-relevant distribution shifts How the Training Signal is Built SDV-W extends SynDroneVision (SDV) by re-rendering a curated subset of 15 sequences under controlled seasonal and meteorological conditions while preserving scene geometry, camera placement, and drone trajectories Environments are instantiated under three seasonal configurations: spring/summer, autumn, and winter Weather variation is introduced through the Colosseum weather API to render precipitation, fog, and particle effects directly within the simulated environment Each effect is parameterized by an intensity scalar α ∈
0, 1: , controlling its strength (α = 0: no effect; α = 1: full intensity) Illumination settings are defined through the joint calibration of three UE5 components, including the Sun and Sky Actor and the Post Process Volume Frame sampling and re-rendering preserve geometric correspondence between SDV-W frames and their clean-condition counterparts Where the Gains Come From The analysis shows that adverse conditions reduce performance relative to their clear-weather counterparts, with snow and fog being most disruptive, reducing mAP@0.25 by 8.2/6.2 percentage points (pp) and raising FNR by 15.8/12.7 pp The dominant failure mode across conditions is missed detection, as FNR increases substantially more than FDR SDV-W provides the clearest benefits on LRDDv1, where mAP@0.5 increases by +10.4 to +14.1 pp across all seven standard YOLO variants For YOLOFEDER FusionNet on LRDDv2, SDV-W consistently improves mAP while reducing both FNR and FDR across all IoU criteria The largest gains occur for medium objects, with mAP@0.5 increasing by +9.0 pp and FNR decreasing by-10.4 pp Limitations and Future Directions The paper concludes that SDV-W is most effective when used to complement, rather than replace, general-purpose synthetic training data While SDV-W mitigates appearance-driven degradation, scale-induced failures require complementary inference strategies Tiled inference improves sensitivity to extrasmall objects at the expense of large object fidelity and increased image-level FDR A key limitation remains the lack of real-world data capturing genuine and diverse adverse-weather for evaluation The systematic analysis of the tiling design space is beyond the scope of this work Ablation Details for Training Set Size The ablation study indicates that additive inclusion of SDV-W yields the more consistently competitive configuration on public datasets On custom recordings, replacement captures much of the additive-setting benefit despite the fixed training set size This suggests that SDV-W is most effective when added as complementary coverage rather than substituted for existing SDV samples Tiled Inference on LRDDv2 Tiled inference induces a scaledependent trade-off, yielding pronounced gains for extra-small drones, increasing mAP@0.25 from 0.152 to 0.407 and mAP@0.5 from 0.133 to 0.312 Small objects also benefit from tiled inference, exhibiting consistent improvements across the evaluated localization criteria, accompanied by a reduction in FNR However, tiling significantly degrades performance for large objects, as their spatial extent often exceeds the tile support Tiled inference improves performance for small and extrasmall drones, but at the cost of degraded performance for larger drones and a substantially increased image-level FDR SDV-W improves operationaldomain coverage for urban drone detection through controlled seasonal and adverse-weather variation Its main benefit is reflected in more reliable detector behavior under adverse appearance shifts, reducing missed detections and false alarms, which are critical for real-world deployment SDV-W is most effective as a complementary training source, broadening environmental variation without sacrificing cross-domain generalization potential While SDV-W mitigates appearance-driven degradation, scale-induced failures require complementary inference strategies Tiled inference improves sensitivity to extrasmall objects, but shifts the deployment trade-off toward reduced large-object consistency and precision Given the distance-dependent scale variation of drones in real-world surveillance, tiled inference should be deployed cautiously rather than by default A key limitation remains the lack of real-world data capturing genuine and diverse adverse-weather for evaluation The systematic analysis of the tiling design space is beyond the scope of this work The paper is 2608.16191.
Tom: Page 1 --- Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection 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 1.
Jane: Introduction Unmanned aerial vehicles (UAVs) – more commonly known as drones – have evolved from specialized military systems into versatile aerial platforms, supporting applications in logistics, agriculture, infrastructure inspection, surveillance, and recreation
35: However, this increased accessibility also raises security and privacy concerns Consequently, reliable drone detection is essential for low-altitude airspace surveillance Camera-based sensing provides a broadly deployable modality for drone detection The reliability of DL-based detectors is inherently tied to the diversity and representativeness of their training data Conditions outside this domain can induce substantial performance degradation Accordingly, real-world drone detection depends on training data that captures both conventional visual variability – viewpoint, target scale, background structure, and illumination – and operationally relevant distribution shifts induced by seasonality and adverse weather However, acquiring and annotating real-world data across the full ODD remains costly and difficult to control Consequently, existing datasets often provide limited coverage of these scenarios Synthetic data generation offers a scalable means of improving ODD coverage In particular, game-engine-based pipelines expose deployment-relevant factors as controllable parameters, enabling systematic variation of visual conditions while automatically providing pixel-precise annotations SynDroneVision (SDV)
27: demonstrates this potential in drone detection, improving performance and cross-domain robustness (especially when complemented by small amounts of real-world data) Nevertheless, its environmental coverage remains limited to variations in illumination, cloud structure, sky appearance, and clear or overcast conditions Existing weather-aware datasets address this gap only partially 2.
Tom: Related Work This section outlines recent developments in image-based drone detection and examines public datasets in terms of seasonal appearance variation and adverse-weather coverage Image-based drone detection primarily relies on singlestage DL architectures, which offer a favorable compromise between real-time inference and detection accuracy Among single-stage architectures, the YOLO family remains the prevailing choice Architectural adaptations typi cally address domain-specific challenges, including smallscale or long-range drone detection The latter poses a particular challenge in urban surveillance, where structural and vegetative clutter can reduce target-background separability, lowering drone saliency and degrading the performance of generic YOLO-based detectors Transformer-based detectors provide an alternative to YOLO-centric detection architectures Despite these advances, detector evaluation rarely accounts for adverse weather or seasonal appearance shifts Training DL-based drone detection models predominantly relies on application-specific real-world data, whose acquisition remains labor-intensive, difficult to scale, and operationally constrained Consequently, existing real-world drone detection datasets capture, at most, seasonal variation under clear-to-overcast conditions—as in LRDDv1
41: and LRDDv2
42: —while excluding adverse atmospheric effects Adverse-weather variation is therefore introduced almost exclusively through synthetic mechanisms, including image-space augmentation
36: and simulation-based rendering Simulation-based rendering provide greater environmental control, yet existing datasets remain limited in scope 3.
Title and authors: Tom: The paper focuses on this SynDroneVision-Weather dataset, SDV-W, which contains fifty-five thousand one hundred eighty-seven annotated high resolution images from three different urban environments across three seasons and various weather conditions like rain, snow, and fog at multiple levels of intensity.
Jane: That's a lot of data to generate systematically. So what exactly is the main finding they present about this dataset?
Tom: They show that by preserving the scene and trajectory configuration from the original SDV data, you can do matched clean versus adverse comparisons and then quantify exactly how much performance degrades depending on the specific weather condition you're testing against.
Lu: What I find interesting is their systematic approach to building this data; they use a Colosseum weather API to render precipitation or fog directly into the environment, controlling the strength of those effects with an intensity scalar alpha from zero to one.
Meng: Controlling that intensity sounds like it’s a great way to make sure we’re testing degradation in a predictable way, rather than just throwing random bad images at the model.
Lalam: They also define illumination settings by jointly calibrating three UE5 components, including the Sun and Sky Actor and the Post Process Volume frame sampling and re-rendering to keep the geometric correspondence between those SDV-W frames and their clean condition counterparts.
The paper's summary: Tom: So, what does this mean for our drone detection models? Essentially, they test these adverse conditions on several representative YOLO models and real world datasets and show that SDV-W improves the detector reliability when there are shifts in appearance.
Jane: When it comes to the results, how much of a difference are we talking about in terms of performance metrics? Are we looking at misses or false alarms specifically?
Tom: They find that adverse conditions reduce performance compared to clear weather. For example, snow and fog are the most disruptive conditions they tested, reducing the mean average precision at twenty five percent mAP at zero by eight point two and six point two percentage points, while raising the false negative rate by fifteen point eight and twelve point seven percentage points.
Lu: The dominant failure mode they see across all these conditions is actually missed detection; it increases much more than the false alarm rate does. That’s a key insight about what goes wrong first in these scenarios.
Meng: So, if we're looking at deployment on an edge device, focusing on reducing those misses sounds like a high priority for us.
Lalam: They also found that SDV-W provides the clearest benefits when applied to LRDDv1, where the mean average precision at fifty percent mAP increases by ten point four to fourteen point one percentage points across all seven standard YOLO variants.
The paper's improvements: Tom: One major improvement they highlight is that for the YOLOFEDER FusionNet on LRDDv2, SDV-W consistently improves the mean average precision while simultaneously reducing both the false negative rate and the false discovery rate across all intersection over union criteria.
Jane: That sounds like a win because it tackles both types of errors at once. What about object size? Does this improvement apply to all sizes equally?
Tom: They found that the largest gains happen for medium objects specifically; the mean average precision at fifty percent mAP increases by nine point zero percentage points and the false negative rate decreases by ten point four percentage points in that category.
Lu: That specificity is important; it tells us exactly where the model is getting help—it's not a blanket improvement across every object type. It’s targeting medium sized drones.
Meng: If we are deploying, knowing which object sizes benefit most helps us decide where to focus our training efforts or where we need to build extra robustness into the model architecture itself.
Lalam: They also looked at a tiling inference strategy on LRDDv2 and found it improves detection performance for extrasmall objects by increasing the mean average precision at twenty five percent mAP from zero point one five two to zero point four zero seven, and the mean average precision at fifty percent mAP from zero point one three three to zero point three one two.
Conclusion: Tom: To wrap this up, the authors conclude that SDV-W is best used as a complement to general purpose synthetic training data rather than replacing it entirely because it successfully mitigates appearance driven degradation.
Jane: So, if I'm just listening on the road, what’s the big picture here? It sounds like we can use this controlled synthetic variation to broaden our environmental coverage without sacrificing the ability for our models to generalize to real-world deployment scenarios.
Lu: The paper also points out that while SDV-W handles appearance shifts well, scale induced failures still require complementary inference strategies, which is where tiling inference comes in.
Meng: From a production standpoint, that means we need a multi-layered strategy: use SDV-W for broad environmental variation during training, and then apply something like tiled inference during actual deployment to handle those extra small objects that benefit from it.
Lalam: The authors also mention a limitation they flag: the lack of real world data capturing genuine and diverse adverse weather conditions for evaluation, and they state that the systematic analysis of the tiling design space is beyond what this work covers.
Tom: So, to summarize, SynDroneVision-Weather provides reliable detector behavior under appearance shifts by quantifying condition specific degradation across different YOLO models.
Jane: It’s a solid piece of work because it gives us a systematic way to generate data that helps our models survive the unpredictable weather we'll encounter when we actually put them out there.
Lu: I think it opens up a path for better synthetic data creation, which is key because acquiring and annotating real world data for these specific scenarios is so hard.
Meng: I’m cautiously optimistic about the scale gains seen on medium objects, but we’ll need to see how stable that performance remains when we introduce those tiling inference trade offs.
Lalam: Ultimately, the paper "Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection" shows how controlled synthetic variation can be a practical means of mitigating the simulation reality gap when combining it with real data.
Institute for the Protection of Terrestrial Infrastructures, German Aerospace Center (DLR) · Technical University of Applied Sciences Wurzburg-Schweinfurt
cs.CV
Submitted: 2026-08-17
Updated: 2026-11-15
Comments: Accepted at 2027 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), Workshop "Real World Surveillance: Applications and Challenges, 7th"
Code: https://github.com/ultralytics/yolov5
Project page: https://codexlabsllc.github.io/Colosseum
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Reliable drone detection under real-world deployment conditions requires training data that spans the full operational design domain…
Key concepts
- Operational Design Domain (ODD)
- The ODD is the complete set of real-world conditions—like different seasons, weather, and lighting—under which a drone detection system must reliably perform. Training data must cover this entire domain to ensure the detector works well when deployed in unpredictable, real-world scenarios.
- SDV-W
- This is a synthetic dataset created by re-rendering existing drone sequences under controlled seasonal and meteorological variations. It systematically introduces weather effects like rain or fog at different intensities, allowing researchers to test how detectors handle these specific adverse conditions.
- Tiled Inference
- This is a technique applied during the detection process where an image is broken into smaller tiles for analysis. It helps improve the system's sensitivity to very small objects, leading to better detection for extrasmall drones, although it can negatively impact the accuracy of detecting large objects.
- Matched Clean-Adverse Comparisons
- This involves comparing a drone detection model's performance on an image taken in clear weather directly against its performance on the same scene rendered under adverse conditions. This comparison quantifies exactly how much a detector degrades when facing specific environmental shifts.
Terminology
Summary
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 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
Introduction to the Problem
Unmanned aerial vehicles (UAVs) have evolved from specialized military systems into versatile aerial platforms, supporting applications in logistics, agriculture, infrastructure inspection, surveillance, and recreation Reliable drone detection is essential for low-altitude airspace surveillance, enabling timely localization and risk mitigation The reliability of DL-based detectors is inherently tied to the diversity and representativeness of their training data Conditions outside this domain can induce substantial performance degradation Existing datasets often provide limited coverage of these scenarios, constraining their ability to support robust generalization under deployment-relevant distribution shifts
How the Training Signal is Built
SDV-W extends SynDroneVision (SDV) by re-rendering a curated subset of 15 sequences under controlled seasonal and meteorological conditions while preserving scene geometry, camera placement, and drone trajectories Environments are instantiated under three seasonal configurations: spring/summer, autumn, and winter Weather variation is introduced through the Colosseum weather API to render precipitation, fog, and particle effects directly within the simulated environment Each effect is parameterized by an intensity scalar α ∈ [0, 1], controlling its strength (α = 0: no effect; α = 1: full intensity) Illumination settings are defined through the joint calibration of three UE5 components, including the Sun and Sky Actor and the Post Process Volume Frame sampling and re-rendering preserve geometric correspondence between SDV-W frames and their clean-condition counterparts
Where the Gains Come From
The analysis shows that adverse conditions reduce performance relative to their clear-weather counterparts, with snow and fog being most disruptive, reducing mAP@0.25 by 8.2/6.2 percentage points (pp) and raising FNR by 15. p. 5 Snow and fog are most disruptive, reducing mAP@0. 2 percentage points (pp) and
raising FNR by 15. Snow and fog are most disruptive, reducing mAP@0. 2 percentage points (pp) and
raising FNR by 15. 034
✓ = applies
At the error-type level, the dominant failure mode is
missed detection. Across conditions, FNR increases substantially more than FDR (Fig. Under stricter
IoU criteria, this degradation becomes more pronounced
(cf. ): for snow, the drop increases from 8. Under stricter
IoU criteria, this degradation becomes more pronounced
(cf. ): for snow, the drop increases from 8. Across all IoU criteria,
SDV-W improves mAP while reducing both FNR and FDR. The largest gains occur for medium objects (mAP@0. Across all IoU criteria,
SDV-W improves mAP while reducing both FNR and FDR. The largest gains occur for medium objects (mAP@0. Details are provided in Sec. Tiled Inference for Small-Object Recovery
Global resizing to 640×640 can induce a pronounced compression of the object-scale distribution, increasing the
prevalence of small and extra-small targets for which detector sensitivity is most limited (cf. Details are provided in Sec. Tiled Inference for Small-Object Recovery
Global resizing to 640×640 can induce a pronounced compression of the object-scale distribution, increasing the
prevalence of small and extra-small targets for which detector sensitivity is most limited (cf. 25 to account for annotation-induced localization uncertainty [28]. To capture operational failure modes, we additionally report the false negative rate (FNR) and false discovery rate (FDR), reflecting missed detections and false
alarms, respectively. 25 to account for annotation-induced localization uncertainty [28]. To capture operational failure modes, we additionally report the false negative rate (FNR) and false discovery rate (FDR), reflecting missed detections and false
alarms, respectively.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations and Future Directions
The paper concludes that SDV-W is most effective when used to complement, rather than replace, general-purpose synthetic training data While SDV-W mitigates appearance-driven degradation, scale-induced failures require complementary inference strategies Tiled inference improves sensitivity to extrasmall objects at the expense of large object fidelity and increased image-level FDR A key limitation remains the lack of real-world data capturing genuine and diverse adverse-weather for evaluation The systematic analysis of the tiling design space is beyond the scope of this work
Ablation Details for Training Set Size
The ablation study indicates that additive inclusion of SDV-W yields the more consistently competitive configuration on public datasets On custom recordings, replacement captures much of the additive-setting benefit despite the fixed training set size This suggests that SDV-W is most effective when added as complementary coverage rather than substituted for existing SDV samples
Tiled Inference on LRDDv2
Tiled inference induces a scaledependent trade-off, yielding pronounced gains for extra-small drones, increasing mAP@0.25 from 0.152 to 0.407 and mAP@0. p. 5 Evaluation Procedure
Evaluation is conducted on the real-world datasets described above, comprising the public datasets DUT AntiUAV, LRDDv1, and LRDDv2, as well as the custom recordings R1-POS3 and R2-POS7. Detection performance is
quantified using standard object detection metrics: mean
average precision (mAP) at an intersection over union
(IoU) threshold of 0.Where the paper says it
Conclusion
SDV-W improves operationaldomain coverage for urban drone detection through controlled seasonal and adverse-weather variation Its main benefit is reflected in more reliable detector behavior under adverse appearance shifts, reducing missed detections and false alarms, which are critical for real-world deployment SDV-W is most effective as a complementary training source, broadening environmental variation without sacrificing cross-domain generalization potential While SDV-W mitigates appearance-driven degradation, scale-induced failures require complementary inference strategies Tiled inference improves sensitivity to extrasmall objects, but shifts the deployment trade-off toward reduced large-object consistency and precision Given the distance-dependent scale variation of drones in real-world surveillance, tiled inference should be deployed cautiously rather than by default A key limitation remains the lack of real-world data capturing genuine and diverse adverse-weather for evaluation The systematic analysis of the tiling design space is beyond the scope of this work The paper is 2608.16191.
--- Page 1 ---
Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection
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
- Introduction
Unmanned aerial vehicles (UAVs) – more commonly known as drones – have evolved from specialized military systems into versatile aerial platforms, supporting applications in logistics, agriculture, infrastructure inspection, surveillance, and recreation [35] However, this increased accessibility also raises security and privacy concerns Consequently, reliable drone detection is essential for low-altitude airspace surveillance Camera-based sensing provides a broadly deployable modality for drone detection The reliability of DL-based detectors is inherently tied to the diversity and representativeness of their training data Conditions outside this domain can induce substantial performance degradation Accordingly, real-world drone detection depends on training data that captures both conventional visual variability – viewpoint, target scale, background structure, and illumination – and operationally relevant distribution shifts induced by seasonality and adverse weather However, acquiring and annotating real-world data across the full ODD remains costly and difficult to control Consequently, existing datasets often provide limited coverage of these scenarios Synthetic data generation offers a scalable means of improving ODD coverage In particular, game-engine-based pipelines expose deployment-relevant factors as controllable parameters, enabling systematic variation of visual conditions while automatically providing pixel-precise annotations SynDroneVision (SDV) [27] demonstrates this potential in drone detection, improving performance and cross-domain robustness (especially when complemented by small amounts of real-world data) Nevertheless, its environmental coverage remains limited to variations in illumination, cloud structure, sky appearance, and clear or overcast conditions Existing weather-aware datasets address this gap only partially
- Related Work
This section outlines recent developments in image-based drone detection and examines public datasets in terms of seasonal appearance variation and adverse-weather coverage Image-based drone detection primarily relies on singlestage DL architectures, which offer a favorable compromise between real-time inference and detection accuracy Among single-stage architectures, the YOLO family remains the prevailing choice Architectural adaptations typi cally address domain-specific challenges, including smallscale or long-range drone detection The latter poses a particular challenge in urban surveillance, where structural and vegetative clutter can reduce target-background separability, lowering drone saliency and degrading the performance of generic YOLO-based detectors Transformer-based detectors provide an alternative to YOLO-centric detection architectures Despite these advances, detector evaluation rarely accounts for adverse weather or seasonal appearance shifts Training DL-based drone detection models predominantly relies on application-specific real-world data, whose acquisition remains labor-intensive, difficult to scale, and operationally constrained Consequently, existing real-world drone detection datasets capture, at most, seasonal variation under clear-to-overcast conditions—as in LRDDv1 [41] and LRDDv2 [42]—while excluding adverse atmospheric effects Adverse-weather variation is therefore introduced almost exclusively through synthetic mechanisms, including image-space augmentation [36] and simulation-based rendering Simulation-based rendering provide greater environmental control, yet existing datasets remain limited in scope
- SynDroneVision-Weather
This section introduces the proposed SynDroneVisionWeather (SDV-W) dataset, detailing its generation methodology, composition, and key characteristics SDV-W extends SDV by re-rendering a curated subset of 15 sequences under controlled seasonal and meteorological conditions Camera placement, drone trajectories, rendering configuration, and pixel-accurate annotations are preserved by reusing the original SDV data generation pipeline Environments are instantiated under three seasonal configurations: spring/summer, autumn, and winter Weather variation is introduced through the Colosseum weather API to render precipitation, fog, and particle effects directly within the simulated environment Each effect is parameterized by an intensity scalar α ∈ [0, 1], controlling its strength (α = 0: no effect; α = 1: full intensity) Illumination settings are defined through the joint calibration of three UE5 components, including the Sun and Sky Actor and the Post Process Volume
- Experimental Setup
This section describes the detectors, data sources, training protocol, and evaluation procedure used to assess the effect of weather-conditioned synthetic data on real-world drone detection The selected detectors cover representative drone detection architectures, enabling a separation of training-data effects from architecture-dependent performance variation The experimental setup builds on a heterogeneous collection of synthetic and real-world datasets for drone detection SDV [27] and its weather-conditioned extension SDV-W (Sec. 3) constitute the synthetic data basis, providing largescale visual data under controlled seasonal and adverseweather variation The real-world data basis complements this controlled synthetic setting with publicly available datasets and custom recordings
- Results
The evaluation first exploits SDV-W’s controlled design to quantify weather-induced detector degradation under paired synthetic conditions It then focuses on SDV-W’s primary role as a weather-conditioned training resource, assessing real-world robustness gains while disentangling the effects of data diversity, dataset scale, and target size Performance under Synthetic Weather & Seasonal Variation shows that snow and fog are most disruptive, reducing mAP@0.25 by 8.2/6.2 percentage points (pp) and raising FNR by 15.8/12.7 pp SDV-W provides the clearest benefits on LRDDv1, where mAP@0. p. 6 On public real-world
datasets, SDV-W provides the clearest benefits in settings
where reference performance is not already saturated. The
most consistent gains are observed on LRDDv1: across
all seven standard YOLO variants, mAP@0.Where the paper says it
--- Page 2 ---
Beyond Clear Skies: Synthetic Seasonal and Weather Variations for Real-World Drone Detection
- SynDroneVision-Weather
This section introduces the proposed SynDroneVisionWeather (SDV-W) dataset, detailing its generation methodology, composition, and key characteristics SDV-W extends SDV by re-rendering a curated subset of 15 sequences under controlled seasonal and meteorological conditions while preserving scene geometry, camera placement, and drone trajectories Environments are instantiated under three seasonal configurations: spring/summer, autumn, and winter Weather variation is introduced through the Colosseum weather API to render precipitation, fog, and particle effects directly within the simulated environment Each effect is parameterized by an intensity scalar α ∈ [0, 1], controlling its strength (α = 0: no effect; α = 1: full intensity) Illumination settings are defined through the joint calibration of three UE5 components, including the Sun and Sky Actor and the Post Process Volume
--- Page 3 ---
SynDroneVision-Weather
This section introduces the proposed SynDroneVisionWeather (SDV-W) dataset, detailing its generation methodology, composition, and key characteristics SDV-W extends SDV by re-rendering a curated subset of 15 sequences under controlled seasonal and meteorological conditions while preserving scene geometry, camera placement, and drone trajectories Environments are instantiated under three seasonal configurations: spring/summer, autumn, and winter Weather variation is introduced through the Colosseum weather API to render precipitation, fog, and particle effects directly within the simulated environment Each effect is parameterized by an intensity scalar α ∈ [0, 1], controlling its strength (α = 0: no effect; α = 1: full intensity) Illumination settings are defined through the joint calibration of three UE5 components, including the Sun and Sky Actor and the Post Process Volume
--- Page 4 ---
Table 1. Composition of SDV-W across seasons and weather
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 (rain+fog)
--- Page 5 ---
Table II. Composition of SDV-W across seasons and weath… p. 3 Composition of SDV-W across seasons and weather. Weather Cond.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
- Weather Robustness through Controlled Synthetic Variation
The system can be improved by integrating SynDroneVision-Weather (SDV-W) into training pipelines to enhance reliability under adverse conditions. SDV-W enables matched clean-adverse comparisons and quantification of condition-specific detector degradation
across different YOLO models, specifically showing that across representative YOLO models and real-world datasets, SDV-W improves detector reliability under adverse appearance shifts.
- Targeted Performance Gains on Specific Object Scales
The AI system can be enhanced by employing tiled inference when processing images, which improves detection performance for small and extrasmall objects
by increasing mAP@0.5 from 0.133 to 0.312 on LRDDv2, while reducing FNR by 30.1 pp for these categories, although this comes with a trade-off where localization of large objects deteriorates.
- Architecture-Specific Mitigation of False Detections
For models like YOLOv8l and YOLOv9c on datasets like LRDDv2, the system can leverage SDV-W to reduce false detections, as the consistent pattern is instead a reduction in false detections, with FDR decreasing across all variants.
This suggests that SDV-W is most effective when added as complementary coverage rather than substituted for existing SDV samples
on public datasets.
- Scale-Specific Training Strategy for Custom Data
When training on custom recordings like R1-POS3, the system can use SDV-W to achieve scale-selective gains, where mAP@0.5 gains [are] concentrated in small and medium object categories.
This demonstrates that SDV-W is most effective when used to complement, rather than replace, general-purpose synthetic training data
for custom scenes.
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), a 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 of 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 controlled, condition-specific assessment of detector degradation through paired clear/adverse frames. Evaluations across representative YOLO models and real-world datasets demonstrate that SDV-W improves detection performance by reducing both 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.
Sources
- LRDDv2: Enhanced Long-Range Drone Detection Dataset with Range Information and Comprehensive Real-World Challenges
- Performance Optimization of YOLO-FEDER FusionNet for Robust Drone Detection in Visually Complex Environments
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs