Enhancing Drone Surveillance Reliability Through Structured Hybrid Data Reinforcement under Environmental Variability

The rapid evolution of uncrewed aerial vehicle (unmanned aerial vehicles (UAVs)/Drone) technology has greatly expanded its use in surveillance, logistics, infrastructure inspection, agriculture, and military operations. Despite these advantages, the increasing number of drones flying in controlled and restricted airspace has created significant operational safety and security challenges for modern aerial surveillance systems [1-4]. Unauthorised drone activities can have serious implications for airport operations, critical infrastructure security, and public safety, and can introduce weaknesses in intelligent surveillance environments. This was exemplified in 2018 by the closure of Gatwick Airport in the United Kingdom due to unlawful drone activity, which led to delays for hundreds of aircraft [5, 6]. More recent reports indicate that airspace violation incidents continue to rise worldwide, underscoring the growing need for reliable.

surveillance systems that maintain robust detection performance across heterogeneous environments [7-9]. Accordingly, improving the reliability and operational robustness of aerial surveillance systems has become an important research priority, particularly in safety-critical environments where missed detections may lead to operational disruption and security risks.

In practical aerial surveillance, drone detection remains challenging due to illumination variation, atmospheric disturbance, motion blur, scale diversity, and complex backgrounds, which often reduce localisation accuracy and detection consistency. These limitations are particularly critical in safety-sensitive applications such as airport surveillance, border security, critical infrastructure protection, and urban airspace monitoring, where missed detections may compromise situational awareness [10]. Consequently, improving surveillance robustness under heterogeneous environmental conditions has become a fundamental requirement for intelligent drone monitoring systems. These challenges have motivated the widespread adoption of deep learning-based object detection methods that learn robust visual representations from heterogeneous aerial imagery.

Recent advances in deep learning have improved aerial object detection by enabling models to learn robust visual representations directly from large-scale image datasets. Convolutional Neural Networks (CNNs) and single-stage object detectors, particularly the You Only Look Once (YOLO) family, have become dominant approaches because they balance detection accuracy and computational efficiency, making them well suited for real-time aerial surveillance applications [11-13]. Attention mechanisms such as the Convolutional Block Attention Module (CBAM) have also been incorporated into YOLO-based detectors to enhance feature representation by emphasising informative spatial regions and channel-wise responses. These attention-enhanced architectures have demonstrated improved localisation performance, especially for small aerial targets under cluttered backgrounds and varying illumination conditions [14-16].

Beyond architectural improvements, recent research has increasingly emphasised the role of data-centric AI, in which the quality, diversity, and representativeness of training data are considered equally important for improving detection performance. Data augmentation and synthetic data generation have therefore emerged as effective strategies for enriching aerial datasets and increasing model robustness under diverse environmental conditions. Conventional augmentation techniques, including rotation, scaling, flipping, and photometric transformations, have demonstrated their ability to improve generalisation by introducing moderate geometric and illumination variations. Likewise, synthetic data generation has enabled the simulation of challenging operational scenarios that are difficult, costly, or unsafe to acquire through field observations alone [15, 17-22]. These developments suggest that surveillance reliability depends not only on detector architecture but also on systematic dataset engineering. However, existing studies remain limited by reliance on public or laboratory-controlled datasets, isolated augmentation strategies, and a predominant focus on architectural optimisation rather than on integrated data-centric reinforcement. Consequently, the combined effects of hybrid dataset engineering, illumination harmonisation, adaptive augmentation, and deterministic experimental protocols remain insufficiently explored, motivating the development of a unified surveillance-oriented framework for robust drone detection.

To address these limitations, this study proposes a structured hybrid data-reinforcement framework for robustness in drone surveillance across heterogeneous environmental conditions. The proposed framework provides four principal contributions. First, a proportional hybrid dataset construction method is introduced to improve environmental representativeness and reduce domain imbalance by integrating diverse image sources with complementary characteristics. Second, illumination harmonisation, Adaptive Multi-Context Augmentation (AMCA), and aeromodelling-based synthetic reinforcement are integrated into a unified data-centric reinforcement pipeline. Third, CBAM-enhanced YOLOv5 is employed to improve feature discrimination and localisation consistency for small aerial targets under complex surveillance scenarios. Finally, a deterministic experimental protocol is established to ensure reproducibility, fairness in benchmarking, and objective evaluation across different model configurations. Collectively, these contributions provide a reproducible, surveillance-oriented framework that enhances operational robustness and supports the development of reliable intelligent aerial surveillance systems capable of operating under heterogeneous environmental conditions with improved adaptability and generalisation capability.

The objective of this study is to develop a systematic hybrid data-reinforcement framework to improve the reliability of drone surveillance systems operating under heterogeneous environmental conditions. As illustrated in Figure 1, the proposed methodology was designed to satisfy four fundamental requirements of engineering research: reproducibility, architectural clarity, experimental validity, and benchmark fairness. All methodological procedures, including dataset preparation, preprocessing, model configuration, experimental settings, and evaluation protocols, are explicitly specified to enable independent replication and objective verification. Dataset partitioning was performed using stratified sampling before any augmentation to preserve class distribution. In contrast, all augmentation and synthetic reinforcement procedures were applied exclusively to the training partition to prevent information leakage into the validation and test sets. The experimental protocol was designed to isolate the contribution of each methodological component by controlling architectural and computational variables across each benchmark comparison, while intentionally varying specific components to meet the experimental objectives. The overall workflow comprises six stages: dataset preparation, preprocessing and illumination harmonisation, model construction, deterministic experimental configuration, performance evaluation, and benchmark validation, as shown in Figure 1.

tu_pian_1.png

Figure 1. Structured hybrid reinforcement framework for robust drone surveillance under environmental variability

3.1 Dataset preparation

The first stage of the proposed framework establishes the DAB-Sky dataset, which serves as the foundation for all experiments. The dataset was constructed by integrating publicly available UAV images with a locally acquired aerial surveillance dataset to improve environmental diversity and operational representativeness. Figure 1 illustrates the overall dataset preparation workflow, while the following subsections describe each stage in detail.

(1) Public dataset collection

Public UAV datasets from previous object detection studies [40, 50] were used to establish the baseline aerial surveillance dataset. These datasets were selected because they contain diverse drone appearances, viewing angles, and environmental conditions. To ensure consistency across heterogeneous data sources, all images were automatically oriented and resized to 640 × 640 pixels before further processing.

(2) Private dataset acquisition

To improve representativeness under real-world aerial surveillance conditions, a private dataset was collected through outdoor image acquisition [51-53]. Data collection was designed to capture realistic operational variability, including illumination changes, atmospheric effects, motion dynamics, object-scale variation, and diverse sky backgrounds. Drone trajectories, viewing distances, flight orientations, and ambient lighting conditions were intentionally varied to represent practical surveillance scenarios.

(3) Proportional hybrid dataset construction

Rather than directly merging heterogeneous datasets, a proportional hybridisation strategy was employed to reduce domain imbalance while preserving environmental diversity. The hybrid dataset is defined as shown in Eq. (1).

$D_{\text {hybrid }}=\alpha D_{\text {public }}+\beta D_{\text {private }}$             (1)

where, $D_{\text {hybrid }}$ denotes the resulting hybrid dataset, $D_{\text {public }}$ and $D_{\text {private }}$ represent the public and private. The proportional weights satisfy, as shown in Eq. (2).

$\alpha+\beta=1$          (2)

thereby maintaining balanced dataset composition and preventing dominance from a single environmental domain.

(4) Dataset composition

The DAB-Sky dataset contains 39,717 images collected from three complementary sources: camera-collected aerial images, aerodynamically generated synthetic images, and publicly available UAV datasets enhanced using AMCA.

This hybrid composition was designed to increase environmental diversity, reduce domain bias, and improve model generalisation under heterogeneous surveillance conditions. Table 2 summarises the distribution of object classes within the DAB-Sky dataset.

Drone images constitute the largest proportion of the dataset to ensure adequate representation of the primary surveillance target. Bird images were intentionally included because of their visual similarity to drones, thereby reducing false-positive detections. Meanwhile, the sky, sun, and tree classes provide realistic environmental contexts commonly encountered in aerial surveillance, thereby improving background diversity and feature discrimination. The composition of data sources is summarised in Table 3.

The integration of real, synthetic, and AMCA-enhanced images provides complementary environmental characteristics that improve surveillance-oriented feature learning while reducing domain imbalance during model training.

Table 2. Distribution of object classes in the DAB-Sky dataset

Table 3. Composition of the DAB-Sky hybrid dataset

(5) Dataset partitioning

To ensure reproducible and unbiased performance evaluation, the DAB-Sky dataset was partitioned into independent training, validation, and test subsets using a stratified sampling strategy that preserved both object-class distribution and environmental diversity across all partitions, as shown in Table 4.

Table 4. Dataset partitioning strategy

The training partition was intentionally allocated the largest proportion of the dataset to maximise exposure to diverse drone poses, illumination conditions, and environmental variations during model optimisation. The validation partition was used for hyperparameter tuning and convergence monitoring, whereas the test partition was reserved exclusively for final performance evaluation.

Although the test partition represents only 4% of the complete dataset, stratified sampling ensured that all object classes and environmental conditions remained proportionally represented, providing an unbiased assessment while maximising the amount of data available for model training. Dataset partitioning was completed before any augmentation or synthetic reinforcement was applied. Subsequently, AMCA and aeromodelling-based synthetic reinforcement were applied exclusively to the training partition, whereas the validation and test partitions remained unchanged across all experiments. This protocol prevented information leakage and ensured fair benchmark comparison across all evaluated model configurations. As presented in Table 4, 89% of the dataset was allocated for training to maximise exposure to rare drone poses, illumination variations, and synthetic perturbations.

In comparison, 7% and 4% were reserved for validation and independent testing, respectively. Although the test subset comprises only 794 images, it was generated using stratified sampling to preserve the original class proportions and environmental diversity of the complete DAB-Sky dataset. Consequently, all object categories (drone, bird, sky, sun, and tree) remained adequately represented in the independent test partition, providing a reliable basis for unbiased performance evaluation while preventing data leakage between the training and test partitions.

3.2 Preprocessing and illumination harmonisation

To improve photometric consistency across heterogeneous aerial surveillance conditions, illumination harmonisation was applied before geometric augmentation. As described in Section 3.1.5, this preprocessing step was performed exclusively on the training partition after dataset splitting to prevent information leakage for each input image, as shown in Eq. (3).

$\mathrm{I} \in D_{\text {hybrid }}$          (3)

where, I denote an input image and $D_{\text {hybrid }}$ represents the proportionally constructed hybrid dataset; photometric normalisation was applied to reduce illumination inconsistency while preserving discriminative visual information.

The harmonisation procedure consisted of three sequential operations: Global histogram equalisation, Adaptive brightness scaling (±20%), and Contrast normalisation. Histogram equalisation redistributed image intensity to improve global contrast, adaptive brightness scaling simulated moderate illumination variability commonly encountered during aerial surveillance, and contrast normalisation enhanced feature consistency across different acquisition conditions. The harmonised images served as input to AMCA, ensuring that subsequent geometric transformations were applied to photometrically normalised data.

3.3 Adaptive Multi-Context Augmentation

To improve model robustness under heterogeneous aerial surveillance conditions, AMCA was developed to generate synchronised geometric and photometric perturbations while preserving the integrity of annotations. Consistent with the dataset partitioning strategy described in Section 3.1.5, AMCA was applied exclusively to the training partition after dataset splitting. Unlike conventional augmentation methods that apply transformations independently, AMCA applies coordinated transformations to images and annotations within predefined parameter ranges. Each training sample is represented as an image–annotation pair $(I, B)$, where, I denotes the input image and B represents the corresponding bounding-box annotations, including the object centre coordinates, width, height, and class label. Image transformations and annotation updates were performed synchronously to maintain spatial consistency throughout the augmentation process.

Geometric transformations consisted of rotation (±15°), horizontal flipping, and vertical flipping. Because these operations modify the object's spatial orientation, bounding-box coordinates were automatically recalculated using geometric transformation matrices, followed by boundary clipping, removal of degenerate bounding boxes, and minimum-area verification to preserve annotation consistency. Photometric transformations consisted of brightness adjustment (±20%), exposure variation (−10%), and saturation variation (+25%), thereby simulating realistic illumination variability encountered in aerial surveillance environments. Unlike geometric transformations, these operations modify only the pixel intensity and colour distribution without altering object geometry; therefore, the original bounding-box annotations remain unchanged. Representative AMCA operations are illustrated in Figure 2. As shown in Figure 2(a), geometric transformations require automatic bounding-box recalculation to preserve annotation consistency after spatial modification. In contrast, Figure 2(b) demonstrates that photometric transformations alter only the image appearance while preserving the original bounding-box annotations, as object geometry is unaffected.

tu_pian_2.png

(a) Geometric transformation

tu_pian_3.jpg

(b) Photometric transformation

Figure 2. Representative Adaptive Multi-Context Augmentation (AMCA) operations. (a) Geometric transformations with automatic bounding-box recalculation to preserve annotation consistency, and (b) photometric transformations with unchanged bounding-box annotations, as object geometry remains unaffected

3.4 Aeromodelling-based synthetic reinforcement

To improve model robustness under heterogeneous aerial surveillance conditions, aeromodelling-based synthetic reinforcement was employed to generate realistic operational perturbations that are difficult to capture through field data acquisition. Controlled environmental and motion perturbations were applied to the training data to produce a synthetic dataset, denoted by $D_{S Y N}$. The generation process was governed by a predefined transformation parameter set, $\theta_{S Y N}$, which controls illumination variation, viewpointsynthetic dataset, denoted by $D_{S Y N}$. The generation process was governed by a predefined transformation parameter set, $\theta_{S Y N}$, which controls illumination variation, viewpoint changes, motion blur, atmospheric effects, and geometric perturbations. Consistent with the dataset-partitioning strategy described in Section 3.1, synthetic reinforcement was applied exclusively to the training partition after dataset splitting, while the validation and test partitions remained unchanged to prevent information leakage and ensure unbiased performance evaluation.

Five surveillance-oriented perturbation models were incorporated: Motion blur, simulating drone and camera movement; Camera gain noise, modelling sensor noise and electronic gain variation; Hue variation, representing atmospheric colour shifts; Saturation fluctuation, simulating weather- and illumination-induced colour variation; Exposure variability, modelling over- and under-exposure caused by changing illumination conditions. The final reinforced training dataset is defined by Eq. (4).

$D_{\text {final }}=D_{\text {hybrid }} \cup \gamma D_{S Y N}$           (4)

where, $D_{\text {final }}$ denotes the reinforced training dataset, $D_{\text {hybrid }}$ represents the proportionally constructed hybrid dataset, $D_{S Y N}$ denotes the synthetic sample set, and $\gamma \in[0,1]$ controls the reinforcement intensity. Representative synthetic perturbations are illustrated in Figure 3. Each perturbation model employs predefined parameter ranges to simulate realistic surveillance conditions while preserving object identity and annotation consistency, thereby improving environmental diversity and feature generalisation during training.

tu_pian_4.png

Figure 3. Representative aeromodelling-based synthetic reinforcement is applied during training, together with the corresponding transformation parameters used to simulate realistic aerial surveillance conditions

3.5 Convolutional Block Attention Module-enhanced YOLOv5 architecture

To evaluate the contribution of attention-guided feature learning, the CBAM was incorporated into the selection of YOLOv5 configurations within the proposed framework. The baseline YOLOv5 architecture was retained, while CBAM was integrated into the backbone feature extraction network without modifying the overall detection pipeline. CBAM sequentially applies the channel attention followed by spatial attention to improve feature representation. The channel attention mechanism emphasises informative feature channels associated with drone characteristics, whereas the spatial attention mechanism highlights discriminative object regions while suppressing background interference. This sequential attention process improves feature discrimination for small aerial targets observed under heterogeneous surveillance conditions. Because CBAM was introduced as a lightweight attention module, the overall computational structure of YOLOv5 remained unchanged, thereby preserving inference efficiency while enhancing surveillance-oriented feature representation. The effectiveness of CBAM was subsequently evaluated through benchmark experiments comparing it against the baseline YOLOv5 architecture under identical conditions.

3.6 Experimental setup and reproducibility protocol

All experiments were conducted under deterministic conditions to ensure reproducibility, objective evaluation, and fair benchmarking. The experimental protocol was designed to isolate the contribution of individual methodological components while maintaining identical computational conditions wherever applicable. Across all benchmark configurations, the YOLOv5s backbone, pretrained weights, hardware platform, dataset partition, image resolution, and evaluation metrics were kept constant. In contrast, the incorporation of CBAM and the optimisation strategy (SGD, AdamW, or AdamW with cosine learning rate scheduling) were intentionally varied in selected experiments to evaluate their individual contributions to detection performance.

1. Hardware Configuration: All experiments were performed on an NVIDIA Tesla T4 GPU with CUDA acceleration and a single CPU core. Maintaining an identical hardware platform ensured computational consistency across all benchmark configurations.
2. Software Environment: The implementation was based on the Ultralytics YOLOv5 framework, using Python 3.12.12 and the pretrained yolov5s model.pt weights.

The input image resolution, anchor configuration, training epochs, and evaluation metrics were kept identical throughout all experiments, whereas the optimiser was varied only in the designated benchmark scenarios.

3. Deterministic Training Protocol: To ensure experimental reproducibility, all benchmark configurations employed fixed random seed initialisation, identical dataset partitioning, consistent preprocessing procedures, and standardised hyperparameter settings whenever applicable. Only the methodological components under investigation, namely dataset reinforcement, CBAM integration, and optimisation strategy, were intentionally modified according to the benchmark design. This controlled protocol enabled performance differences to be attributed to the evaluated methodological factors rather than to uncontrolled experimental variation.

3.7 Evaluation metrics

Detection performance was evaluated using four widely adopted object detection metrics: Precision, Recall, mean Average Precision at an Intersection-over-Union threshold of 0.5 (mAP@0.5), and mean Average Precision averaged over IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95). Precision measures the proportion of correctly identified drone detections among all predicted detections, whereas Recall quantifies the proportion of detected drone targets relative to all ground-truth objects. The mAP@0.5 metric evaluates detection accuracy using a fixed IoU threshold of 0.5, while mAP@0.5:0.95 provides a more stringent assessment by averaging localisation performance across multiple IoU thresholds. Together, these complementary metrics provide a comprehensive evaluation of detection accuracy, localisation quality, and operational robustness under heterogeneous aerial surveillance conditions, enabling fair comparison across all benchmark configurations.

3.8 Benchmark design and fair experimental comparison

To systematically evaluate the contribution of each methodological component, six benchmark configurations were designed to represent the progressive development of the proposed surveillance framework. Each configuration introduces a specific combination of hybrid dataset engineering, illumination harmonisation, AMCA, aeromodelling-based synthetic reinforcement, CBAM integration, and optimisation strategy, enabling incremental assessment of their individual and combined effects on drone detection performance. To ensure benchmark fairness, several experimental factors were maintained constant across all benchmark configurations, including the YOLOv5s backbone, pretrained weights, input image resolution, dataset partition, hardware platform, evaluation metrics, and deterministic training protocol. In contrast, CBAM integration and optimiser selection (SGD, AdamW, and AdamW with cosine learning-rate scheduling) were intentionally varied across the corresponding benchmark configurations to quantify their individual contributions. At the same time, dataset reinforcement was progressively introduced in accordance with the experimental design. This controlled benchmark design enabled performance differences to be attributed to the evaluated methodological components rather than to uncontrolled variations in computational settings. Consequently, the proposed framework provides an objective and reproducible basis for comparing the influence of data-centric reinforcement, attention-guided feature learning, and training optimisation on surveillance performance under heterogeneous aerial environmental conditions.