AerialCSP: Synthetic Dataset for CSP Inspection

• AerialCSP is a hybrid synthetic dataset that blends CAD-derived 3D renders with real UAV imagery to represent parabolic trough CSP plants accurately.
• The dataset pipeline integrates 3D rendering in Blender, advanced inpainting using Kandinsky and Stable Diffusion, and meticulous compositing to ensure pixel-level annotation.
• Pretraining on AerialCSP significantly enhances detection and segmentation performance, enabling robust fault detection in real CSP inspection tasks with limited annotated data.

AerialCSP is a synthetic or virtual dataset for aerial RGB imagery of Parabolic Trough Concentrated Solar Power plants, introduced to reduce the gap between generic computer-vision pretraining data and the visually unusual inspection regime of CSP solar fields. It combines CAD-derived 3D rendering of plant components with inpainted real UAV backgrounds, provides annotations for object detection and instance segmentation, and is intended primarily as a pretraining corpus for downstream CSP inspection tasks in which real labeled data are scarce, expensive, and often constrained by privacy and proprietary restrictions (Pérez-Cutiño et al., 1 Aug 2025).

1. Problem setting and domain specificity

Parabolic Trough CSP imagery differs sharply from the scene statistics that dominate ImageNet, COCO, or Cityscapes. A PT field consists of long, highly reflective parabolic mirrors, Heat Collector Elements running along the focal line, torque tubes, supports, and other repeated structural components. In aerial RGB imagery from UAVs, these structures are highly specular, geometrically repetitive, and dominated by sky and ground reflections rather than stable texture. At the same time, inspection targets such as broken tubes, chipped mirror corners, or missing glass envelopes are rare and occupy only a small number of pixels. The result is a severe domain shift: off-the-shelf detectors and segmenters trained on generic corpora treat CSP content as out-of-distribution clutter, segment foreground and background poorly, and miss rare classes badly when only limited real annotation is available (Pérez-Cutiño et al., 1 Aug 2025).

This problem is compounded by the practical difficulty of building large real datasets. UAV operations over active plants, storage of large image volumes, and manual annotation of thousands of instances are expensive, and data collection is often restricted by privacy and proprietary constraints from plant operators. AerialCSP was proposed as a direct response to this regime: small-data, domain-specific, visually atypical, and operationally costly to label (Pérez-Cutiño et al., 1 Aug 2025).

A common misconception is that AerialCSP is merely a conventional synthetic dataset generated in a fully virtual world. In fact, its central premise is hybridization: rendered CSP objects are composited onto real CSP environments from UAV imagery after digital removal of the original structures. This makes the dataset neither purely rendered nor merely augmented real imagery.

2. Dataset construction pipeline

The AerialCSP pipeline has three stages: 3D modeling and rendering in Blender, background generation by inpainting real UAV imagery, and compositing of rendered plant components onto those backgrounds (Pérez-Cutiño et al., 1 Aug 2025).

The first stage starts from engineering CAD plans of a Solar Collector Element, exported to STL and imported into Blender. Blender scripting is used to define materials for mirrors, HCE, HCE supports, torque tube, and supporting structure; place a fixed light source perpendicular to the XY plane to emulate direct sunlight; place a camera between the light source and SCE to mimic a nadir-looking UAV camera; and rotate the SCE around the Y-axis to simulate solar tracking with a $1^\circ$ step size. For each tracking angle $\theta$ from $-90^\circ$ to $+90^\circ$, yielding 181 positions, the pipeline renders a photorealistic RGB image and a segmentation mask in which each class has a unique flat color. These masks provide exact pixel-level segmentation.

The second stage begins from 800 real UAV images collected over several CSP plants using a DJI Matrice 300 RTK, flown at 20–23 m altitude with the camera pointing straight down and the gimbal at $-90^\circ$. In each image, a rectangular region containing the CSP structures is marked. Image inpainting is then applied to remove collectors and reconstruct plausible terrain and context. The models used are Kandinsky and Stable Diffusion in latent-diffusion inpainting mode, driven by a fixed generic prompt tuned to produce realistic terrain or ground texture. A coarse rectangular mask is sufficient because the inpainting fills the entire region.

The third stage composites rendered SCEs onto the inpainted backgrounds. For each rendered SCE and tracking angle, a background is randomly chosen from the Kandinsky-based and Stable Diffusion-based sets. Random 2D transformations—translation, rotation, and scaling—are applied to the rendered object and its mask before pasting. Images with objects too close to image borders are filtered out, removing 42 samples. The final dataset contains 18,058 labeled images (Pérez-Cutiño et al., 1 Aug 2025).

Stage	Inputs and tools	Output
3D rendering	CAD, STL, Blender, solar-tracking sweep	RGB renders and class masks
Background generation	800 UAV images, Kandinsky, Stable Diffusion	Inpainted CSP backgrounds
Compositing	Rendered SCEs plus inpainted backgrounds	18,058 labeled aerial scenes

This design preserves realistic ground texture, road and soil context, approximate sensor characteristics, and CSP-specific geometry simultaneously. The paper does not report quantitative domain-gap metrics such as FID or t-SNE statistics, but the construction is explicitly aimed at reducing the gap through real backgrounds, correct 3D geometry, reflective materials, two inpainting models, and random 2D transformations (Pérez-Cutiño et al., 1 Aug 2025).

3. Content, annotations, and supported tasks

AerialCSP focuses on PT structural components rather than explicit fault imagery. The main synthetic classes are mirrors, HCE, HCE support, torque tube, and supporting structure. These are modeled per SCE and vary with tracking angle and viewpoint. The dataset does not explicitly model broken HCEs or broken mirrors; those appear only in the real downstream evaluation set. The intended mechanism is that pretraining on normal structure and context improves later fine-tuning for anomalous or damaged instances (Pérez-Cutiño et al., 1 Aug 2025).

The dataset is distributed with a standard 75%/15%/10% train/validation/test split. Given 18,058 images, the test set contains 1,807 images. The paper does not specify a fixed image resolution for AerialCSP, but it states that real UAV imagery is high-resolution at 20–23 m altitude and that rendered SCE crops are composited to fill a central area with random scale and rotation. The viewpoint therefore mimics nadir or near-nadir inspection imagery.

AerialCSP supports both object detection and instance segmentation. Bounding boxes are provided in YOLO format: $$\langle \text{class} \rangle\ \langle x_\text{center}\rangle\ \langle y_\text{center}\rangle\ \langle w\rangle\ \langle h\rangle$$ with coordinates normalized to $[0,1]$ and 0-based class indices. Instance masks are provided as normalized polygons: $$\langle \text{class-index}\rangle\ \langle x_1\rangle\ \langle y_1\rangle\ \dots\ \langle x_n\rangle\ \langle y_n\rangle$$ where $(x_i,y_i)$ are polygon vertices derived directly from Blender masks. Because the simulation has access to exact geometry and object IDs, the annotations are described as pixel-perfect, without annotation noise or ambiguity, including for small and thin components (Pérez-Cutiño et al., 1 Aug 2025).

The paper also states that the dataset is well suited as a pretraining resource for downstream tasks such as defect detection, pose or tracking-angle estimation, and optical characterization by attaching task-specific heads. This suggests that its value lies not only in the benchmark tasks it directly supports, but in its role as a semantic backbone for CSP-specific representation learning.

4. Benchmarking on the synthetic dataset

The paper benchmarks AerialCSP with Ultralytics YOLOv11 on bounding-box detection and instance segmentation, using three model scales: [yolo11n](https://www.emergentmind.com/topics/yolo11n), yolo11m, and yolo11x, with -seg variants for segmentation. All models are trained for 200 epochs on AerialCSP, using standard YOLO augmentations such as flips and cropping, and the best checkpoint is selected by validation performance (Pérez-Cutiño et al., 1 Aug 2025).

Evaluation follows the standard detection metrics: $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}$$ with classwise AP derived from the precision–recall curve, $\theta$0 at IoU $\theta$1, and $\theta$2 averaged over IoU thresholds from 0.5 to 0.95 in 0.05 increments. Mean AP is

$\theta$3

and IoU is

$\theta$4

On the AerialCSP test set of 1,807 images, all detection models achieve more than 90% mAP50 for every class, and mAP50–95 is at least 61% for all classes. The paper gives representative YOLOv11x detection results: mirrors with $\theta$5, $\theta$6, $\theta$7, and $\theta$8–95 $\theta$9; HCE with $-90^\circ$0 and $-90^\circ$1–95 $-90^\circ$2; HCE support with $-90^\circ$3 and $-90^\circ$4–95 $-90^\circ$5; supporting structure with $-90^\circ$6 and $-90^\circ$7–95 $-90^\circ$8; and torque tube with $-90^\circ$9 and $+90^\circ$0–95 $+90^\circ$1 (Pérez-Cutiño et al., 1 Aug 2025).

Instance segmentation is harder. Mirrors obtain mAP50 around 84–89 and mAP50–95 around 65–70; HCE achieves mAP50 around 86–92 and mAP50–95 around 37–40; smaller parts such as HCE support, supporting structure, and torque tube fall to mAP50–95 of approximately 22–29. The paper attributes this degradation primarily to small object sizes, strong foreshortening, and occlusions when SCEs are rotated. The central interpretation is not that segmentation is solved, but that AerialCSP is internally coherent and learnable: high detection performance indicates that the rendered images and exact annotations define a consistent training distribution (Pérez-Cutiño et al., 1 Aug 2025).

5. Transfer to real CSP fault detection

The principal claim of AerialCSP is not synthetic benchmark performance but transfer to real inspection data. For this purpose, the paper constructs a real UAV dataset of 800 aerial images from multiple PT plants and annotates four detection classes: mirrors, HCEs, broken HCEs, and broken mirrors. The split is 400 images for testing and 400 for training, with five training subsets of 10, 50, 100, 200, and 400 images to measure data efficiency (Pérez-Cutiño et al., 1 Aug 2025).

Two pretraining scenarios are compared for YOLOv11n, YOLOv11m, and YOLOv11x. In the “Not pretrained” condition, models are initialized from COCO-pretrained weights. In the “Pretrained” condition, models are first trained on AerialCSP and then fine-tuned on the real CSP dataset. Downstream hyperparameters are kept identical, and real-data augmentation uses random flips, cropping, and mosaic.

For the largest real training set of 400 images, the reported overall detection results are:

Model	COCO-only	AerialCSP-pretrained
yolo11n	26.0 mAP50 / 21.1 mAP50–95	63.2 / 44.1
yolo11m	38.1 / 26.7	77.0 / 56.4
yolo11x	61.2 / 41.4	86.1 / 65.8

Across all training set sizes, AerialCSP-pretrained models dominate the COCO-only models. The paper highlights one especially consequential result: the nano model pretrained on AerialCSP reaches 63.2% mAP50, outperforming the extra-large model trained only on COCO at 61.2% mAP50. This is directly relevant for UAV deployment, where smaller models are often preferred (Pérez-Cutiño et al., 1 Aug 2025).

The rare-defect results are even more pronounced. With 400 real training images, broken HCE mAP50 changes from 0.00 to 43.7 for yolo11n, from 3.99 to 77.0 for yolo11m, and from 35.3 to 82.0 for yolo11x. Broken-mirror mAP50 changes from 6.90 to 30.2 for yolo11n, from 23.6 to 56.4 for yolo11m, and from 37.7 to 70.7 for yolo11x. The noteworthy point is that synthetic AerialCSP images never contained broken elements explicitly. The paper interprets the transfer gain as evidence that models learn the visual grammar of normal CSP components and context, which then improves recognition of deviations during fine-tuning (Pérez-Cutiño et al., 1 Aug 2025).

This directly counters another common misunderstanding: AerialCSP is not valuable only if it contains explicit synthetic defects. In the reported experiments, pretraining on healthy component structure alone substantially improves fault detection, including for rare and small defects. The paper further states that, in practice, operators can train effective fault detectors with about half or less real annotations while still achieving higher performance.

6. Realism, limitations, extensions, and scope of the term

AerialCSP reduces domain gap through several explicit design choices: real background textures from UAV flights, physically plausible CSP geometry from CAD and Blender, reflective materials that cast shadows, multiple inpainting models for mild domain randomization, and random translation, rotation, and scaling to mimic flight irregularities and off-nadir views. The downstream evaluation set comes from multiple CSP plants rather than only the plants used to produce backgrounds, and the pretrained models generalize across those plants despite the absence of explicit synthetic defect classes. The paper characterizes AerialCSP as a domain-specific pretraining corpus analogous to ImageNet for generic vision tasks (Pérez-Cutiño et al., 1 Aug 2025).

Its limitations are equally explicit. Synthetic mirrors may not perfectly reproduce the complex specular patterns seen in practice. The simulator does not model detailed sensor noise, compression artifacts, extreme weather, or atmospheric effects. Broken components are absent from the synthetic dataset. Background inpainting quality varies and may leave artifacts. The dataset is RGB-only, whereas real CSP inspection often uses thermal or IR imagery. Future directions proposed in the paper include explicit synthetic faults such as broken or missing mirrors and tubes, misalignment, structural sag, soiling, and leaks; thermal or IR simulation; better physically based materials and illumination; more sophisticated background generation; and the use of AerialCSP as a source domain in domain-adaptation or active-learning pipelines (Pérez-Cutiño et al., 1 Aug 2025).

A final point of scope is terminological. In the CSP-vision literature, AerialCSP denotes the synthetic aerial dataset for PT plants introduced in (Pérez-Cutiño et al., 1 Aug 2025). Elsewhere, however, the same acronym or label has been used more loosely. In UAV-aided 5G research, it has been associated with an “Aerial Communication Service Platform” built around a flying DU+RU with integrated access and backhaul (Mundlamuri et al., 2023). In low-altitude cooperative sensing, it has also been used informally for an AI-based aerial cooperative sensing or processing problem over FWA infrastructure (Zhang et al., 8 May 2026). Within concentrated solar power computer vision, the term is therefore best understood in the narrower and specific sense established by the 2025 dataset paper: a publicly released, annotated synthetic resource for CSP aerial inspection, available at https://mpcutino.github.io/aerialcsp/ (Pérez-Cutiño et al., 1 Aug 2025).