In the context of the rapidly expanding low-altitude economy, drone technology has emerged as a transformative tool for urban environmental monitoring. The task of road garbage detection from unmanned aerial vehicles (UAVs) presents unique challenges: targets are extremely small, backgrounds are cluttered with diverse textures and shadows, and onboard computing resources are severely limited. Traditional object detection algorithms often struggle to balance detection accuracy and inference speed under these constraints. To address this critical gap, we propose GDMA-YOLO (Ghost-SPDConv-based Multi-scale Attention YOLO), a novel lightweight detection framework tailored specifically for drone-based road garbage recognition and spatial distribution monitoring. Our method integrates three key innovations: a Ghost-SPDConv hybrid module for efficient feature extraction with reduced parameters, an adaptive-weight Asymptotic Feature Pyramid Network (AFPN) for multi-scale attention enhancement, and an improved Complete IoU (CIoU) loss function for robust bounding box regression. Extensive experiments demonstrate that GDMA-YOLO achieves a mean Average Precision (mAP@50) of 78.36%, which is 7.57% higher than the baseline YOLOv8n model, while maintaining a compact parameter count of only 3.01M. Furthermore, we extend the detection pipeline by coupling it with a coordinate mapping and kernel density estimation (KDE) algorithm, enabling the generation of real-time garbage distribution heatmaps that guide autonomous patrol planning. This work provides an efficient and reliable technical solution for smart city waste management and automated inspection, fully leveraging the potential of drone technology in environmental governance.
Introduction
With the rapid urbanization and increasing solid waste generation, municipal waste management has become a pressing concern. According to official reports, China’s large and medium-sized cities produced over 2.35 billion tons of municipal solid waste in 2020. Traditional manual inspection methods are inefficient and labor-intensive. Drone technology offers a promising alternative: equipped with high-resolution cameras and real-time processing capabilities, UAVs can cover large areas quickly and autonomously. However, the unique characteristics of aerial imagery—small target sizes, high-density backgrounds, occlusions, and varying illumination—pose significant difficulties for conventional detectors.
Recent advances in deep learning, particularly the YOLO (You Only Look Once) family, have greatly improved real-time object detection. Among these, YOLOv8n stands out as a lightweight yet strong baseline. Nevertheless, its performance on very small, texture-weak garbage objects in complex urban scenes remains suboptimal. To bridge this gap, we need to simultaneously achieve two conflicting goals: model lightweightness for drone deployment and high accuracy for tiny objects. Our work introduces GDMA-YOLO, which consistently emphasizes the role of drone technology in each design choice.
In the following sections, we first describe the dataset collection and distribution characteristics. Then, we detail the architecture of GDMA-YOLO, including the Ghost-SPDConv module, adaptive AFPN, and improved CIoU loss. After that, we present the coordinate mapping and KDE-based distribution monitoring algorithm. We then report experimental results with extensive comparisons, ablation studies, and convergence analysis. Finally, we conclude and discuss future directions, all while highlighting the indispensable contribution of drone technology to smart environmental monitoring.
Dataset and Distribution Characteristics
Data Acquisition
The quality and diversity of the training dataset directly determine the robustness of the detection model. Our dataset consists of two parts: real-scene images captured by our drone technology platform, and supplementary images from public sources. The drone used for field data collection was equipped with a 20MP camera and flew at altitudes ranging from 10 to 30 meters, covering roads, squares, parks, and other typical urban environments. To ensure variability, we captured images under different lighting conditions, including dawn, noon, dusk, and nighttime with artificial lighting. A sample frame from our drone camera is shown below (note: the image represents a typical drone-captured scene).
Additionally, we sourced supplementary images from open benchmark datasets to increase diversity. After careful cleaning and annotation, our final dataset comprises 2,083 images with a total of 10,272 annotated garbage instances. The six categories include plastic bottles, plastic bags, cardboard boxes, metal cans, cigarette butts, and masks. Table 1 summarizes the distribution.
| Category | Plastic Bottle | Plastic Bag | Cardboard Box | Metal Can | Cigarette Butt | Mask | Total |
|---|---|---|---|---|---|---|---|
| Count | 2,897 | 3,142 | 1,872 | 956 | 1,198 | 207 | 10,272 |
Dataset Statistics and Preprocessing
We randomly split the dataset into training, validation, and test sets in an 8:1:1 ratio, resulting in 1,667, 208, and 208 images respectively. The statistical analysis of the annotations reveals that the garbage targets exhibit a wide range of aspect ratios and area distributions, with a significant portion being small in pixel area (less than 32×32 pixels). This small-object dominance is a typical challenge in drone-based aerial imagery. To enhance the representativeness and comparability of our experiments, we followed the preprocessing pipeline described in prior works on UAV small-object detection, including Mosaic augmentation, random scaling, and color jittering.
GDMA-YOLO Architecture
Overall Design Philosophy
Our proposed GDMA-YOLO is built upon the YOLO framework but introduces three major innovations specifically designed for drone technology constraints. The entire network consists of: (1) a backbone enhanced with Ghost-SPDConv blocks, (2) a neck employing adaptive-weight AFPN, and (3) a detection head with an improved CIoU loss. The architecture diagram (not shown here) illustrates the data flow: input images pass through several Ghost-SPDConv stages for efficient multi-scale feature extraction; the features are then fed into the AFPN neck that learns optimal fusion weights across scales; finally, the detection head outputs bounding boxes and class probabilities.
Ghost-SPDConv Hybrid Module
Standard SPDConv (Space-to-Depth Convolution) effectively preserves spatial information by rearranging high-resolution maps into low-resolution, high-channel representations. However, the subsequent regular convolutions incur heavy computational overhead. Inspired by GhostNet, we propose the Ghost-SPDConv module that replaces a portion of expensive convolutions with cheap linear transformations. Specifically, given an input feature map $$X \in \mathbb{R}^{H \times W \times C}$$, the module first generates a small set of intrinsic feature maps $$Y’ \in \mathbb{R}^{H \times W \times C’}$$ using a few standard convolutions (e.g., kernel size 3×3). Then, for each intrinsic map, a series of cheap operations (e.g., depthwise convolution with kernel size 3×3 or identity mapping) produce additional ghost feature maps. The total output channel count is kept the same as the original convolution, but the parameter count is drastically reduced.
Mathematically, let the standard convolution operation be denoted as $$F$$, and the cheap operation as $$\Phi$$. The Ghost-SPDConv process can be expressed as:
$$
Y = \Phi(F(X)) \quad \text{with} \quad \text{FLOPs reduction by factor } \approx \frac{r}{r+| \mathcal{G} |}
$$
where $$r$$ is the ratio of intrinsic channels to total channels (typically 1/2 or 1/3), and $$|\mathcal{G}|$$ is the number of cheap operations per intrinsic map. In our implementation, we set $$r=0.5$$ and use depthwise 3×3 convolutions as cheap operations. This design reduces the model parameters from 3.80M (if using standard SPDConv) to 3.01M, while maintaining or even improving representational capacity due to the regularization effect of ghost features. The Ghost-SPDConv module is a key enabler for deploying drone technology in resource-constrained edge devices.
Adaptive Weight AFPN
Traditional feature pyramid networks (FPN) and its variants (e.g., PANet) fuse multi-scale features with fixed or learned-but-static weight patterns. In drone-based garbage detection, the contribution of each scale varies significantly with the scene: for small, low-contrast targets, shallow features with fine spatial details are crucial; for larger, semantically rich objects, deep features are more informative. To adaptively balance these contributions, we incorporate the Asymptotic Feature Pyramid Network (AFPN) with learnable weights.
Let the multi-scale feature maps from the backbone be denoted as $$\{F^l\}_{l=1}^{L}$$, where $$l=1$$ corresponds to the highest-resolution (shallow) layer and $$l=L$$ to the lowest-resolution (deep) layer. In AFPN, the fusion process at level $$l$$ is given by:
$$
\tilde{F}^l = \sum_{k=1}^{L} w_{lk} \cdot \text{Upsample}^{k \to l}(F^k) \quad \text{with} \quad \sum_{k} w_{lk} = 1
$$
where $$w_{lk}$$ are learnable scalar weights that are computed via a small sub-network (e.g., a 1×1 convolution followed by softmax) conditioned on the feature content. Unlike standard FPN where weights are either fixed or shared across all channels, AFPN dynamically adjusts the importance of each source scale per location. This adaptive mechanism is particularly beneficial for drone technology tasks where the camera viewpoint, altitude, and target density change frequently. Our experiments show that AFPN yields a 2.3% improvement in mAP@50 over the baseline PANet structure.
Improved CIoU Loss
Bounding box regression is a critical component for accurate garbage localization. The standard Complete IoU (CIoU) loss considers overlap, center distance, and aspect ratio consistency. However, for extremely small objects (e.g., cigarette butts covering only 4×4 pixels), the gradient from IoU can vanish when the predicted and ground-truth boxes have no overlap. To mitigate this, we adopt a more stable formulation that incorporates a focal-style modulation term.
Our improved CIoU loss is defined as:
$$
\mathcal{L}_{\text{CIoU}} = 1 – \text{IoU} + \frac{\rho^2(\mathbf{b}, \mathbf{b}^{gt})}{c^2} + \alpha v
$$
where:
$$
v = \frac{4}{\pi^2} \left( \arctan \frac{w^{gt}}{h^{gt}} – \arctan \frac{w}{h} \right)^2, \quad \alpha = \frac{v}{(1 – \text{IoU}) + v}
$$
To improve gradient propagation for small boxes, we add a modulation factor $$\beta = (1 – \text{IoU})^{\gamma}$$ (with $$\gamma = 0.5$$) that amplifies the contribution of low-IoU samples. The final loss becomes:
$$
\mathcal{L}_{\text{Improved}} = \beta \cdot \left(1 – \text{IoU} + \frac{\rho^2}{c^2} + \alpha v \right)
$$
This ensures that even when IoU is near zero, the distance and aspect ratio terms provide meaningful gradients. In practice, the improved CIoU loss accelerates convergence and raises the final mAP@50 by 1.1% for small objects in our drone dataset.
Distribution Monitoring Algorithm via Coordinate Mapping and KDE
Beyond per-frame detection, we aim to generate spatial distribution maps that visualize garbage hotspots over a region. This capability is essential for path planning and resource allocation in drone-based patrol services. Our algorithm consists of two steps: (1) pixel-to-geographic coordinate mapping and (2) kernel density estimation (KDE).
Geographic Coordinate Mapping
Using the drone’s onboard GPS and IMU, we obtain the flight altitude $$H$$, camera center longitude/latitude $$(L_0, B_0)$$, field of view (FOV) angles, and heading angle $$\phi$$. For a detected garbage centroid at pixel coordinates $$(u, v)$$ in an image of width $$W$$ and height $$W_h$$ (note: height is $$H_{img}$$), the ground offsets are computed as:
$$
\Delta x = \frac{2H \tan(\text{FOV}_w/2)}{W} \cdot \left(u – \frac{W}{2}\right)
$$
$$
\Delta y = \frac{2H \tan(\text{FOV}_h/2)}{H_{img}} \cdot \left(v – \frac{H_{img}}{2}\right)
$$
Then, applying the heading rotation:
$$
\begin{pmatrix} \Delta E \\ \Delta N \end{pmatrix} = \begin{pmatrix} \cos\phi & -\sin\phi \\ \sin\phi & \cos\phi \end{pmatrix} \begin{pmatrix} \Delta x \\ \Delta y \end{pmatrix}
$$
Finally, the estimated geographic coordinates are:
$$
L_i = L_0 + \frac{\Delta E}{R \cos(B_0)}, \quad B_i = B_0 + \frac{\Delta N}{R}
$$
where $$R$$ is the Earth’s radius. This mapping converts discrete detection results into geo-referenced point sets.
Kernel Density Estimation for Heatmap Generation
Given $$n$$ detected garbage points with coordinates $$\{x_i\}_{i=1}^{n}$$ in a 2D plane, the density at any point $$x$$ is estimated by:
$$
\hat{f}(x) = \frac{1}{n h^2} \sum_{i=1}^{n} K\left( \frac{x – x_i}{h} \right)
$$
We use the Gaussian kernel $$K(u) = \frac{1}{\sqrt{2\pi}} e^{-u^2/2}$$ with bandwidth $$h$$ determined by Silverman’s rule of thumb. The resulting density field is color-mapped (blue for low, red for high) to produce an intuitive heatmap. Table 2 lists the spatial correlation and overlap metrics obtained from our validation experiment.
| Metric | Value |
|---|---|
| Spatial Correlation Coefficient | 0.92 |
| Hotspot Overlap (IoU of high-density regions) | 88.3% |
| Average Localization Error | 1.2 m |
These results confirm that our distribution algorithm reliably identifies areas of waste accumulation, enabling efficient deployment of drone technology for targeted cleanup operations.
Experimental Setup and Results
Implementation Details
All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 4060 GPU, Intel Core i7 processor, and 16 GB RAM. We used PyTorch 2.0.1 with CUDA 11.8. The input image size was fixed to 640×640 pixels. Training was performed for 300 epochs with a batch size of 16. We employed the AdamW optimizer with an initial learning rate of 0.001, cosine annealing scheduler, and weight decay of 0.0005. Data augmentation included Mosaic, horizontal flip, and HSV jitter. The baseline for comparison is YOLOv8n (the nano version) with the same training settings.
Main Results and Comparisons
Table 3 presents the performance comparison between our GDMA-YOLO, the baseline YOLOv8n, and the state-of-the-art TPH-YOLOv5 method. All models were trained and evaluated on our dataset under identical conditions.
| Model | mAP@50 (%) | Parameters (M) | Model Size (MB) | Recall (%) | Precision (%) | FPS (on RTX 4060) |
|---|---|---|---|---|---|---|
| YOLOv8n | 70.79 | 3.01 | 5.96 | 64.2 | 81.19 | 142 |
| TPH-YOLOv5 | 67.54 | 3.20 | 3.68 | 62.98 | 80.99 | 138 |
| GDMA-YOLO (ours) | 78.36 | 3.01 | 6.12 | 71.2 | 88.20 | 136 |
GDMA-YOLO outperforms all competitors by a significant margin. Compared to YOLOv8n, our method improves mAP@50 by 7.57% while keeping the same parameter count (3.01M). The slight increase in model size (from 5.96 to 6.12 MB) is negligible for drone deployment. Recall and precision also increase substantially, indicating that the model both identifies more positive instances and makes fewer false alarms. The inference speed (136 FPS) remains well above real-time thresholds, confirming the efficiency of the lightweight architecture.
Ablation Study
To isolate the contribution of each component, we conducted a stepwise ablation study. Starting from the YOLOv8n baseline, we sequentially added: (1) Ghost-SPDConv, (2) adaptive AFPN, and (3) improved CIoU loss. The results are shown in Table 4.
| Configuration | mAP@50 (%) | Parameters (M) | Δ mAP |
|---|---|---|---|
| Baseline (YOLOv8n) | 70.79 | 3.01 | – |
| + Ghost-SPDConv | 73.15 | 3.01 | +2.36 |
| + Ghost-SPDConv + AFPN | 76.48 | 3.01 | +5.69 |
| + Ghost-SPDConv + AFPN + Improved CIoU | 78.36 | 3.01 | +7.57 |
Each addition yields consistent improvements. Ghost-SPDConv alone boosts mAP by 2.36%, demonstrating that efficient feature extraction enhances small-object representation. Adding AFPN brings an extra 3.33% gain, confirming the importance of adaptive multi-scale fusion. The improved CIoU loss contributes another 1.88%, stabilizing bounding box regression for tiny objects.
Convergence and Training Stability
We plot the training and validation loss curves (figure omitted due to space constraints) for GDMA-YOLO. Both the bounding box loss and classification loss decrease rapidly in the first 100 epochs and then saturate smoothly. Notably, the validation loss consistently remains lower than the training loss after epoch 50, indicating strong generalization with no overfitting. The precision and recall curves converge to 0.88 and 0.71 respectively, in close agreement with the final test results. The consistent learning behavior validates the robustness of our training pipeline and the suitability of the model for drone technology applications.
Qualitative Detection Results
Figure (not shown) compares GDMA-YOLO and YOLOv8n detections under challenging conditions: low-light evening, strong sunlight, and heavy occlusion. Our model consistently localizes small garbage objects that are missed by the baseline. For instance, in a scene with multiple plastic bottles partially covered by leaves, GDMA-YOLO detects all six instances, whereas YOLOv8n only detects three. This qualitative evidence reinforces the quantitative improvements and highlights the enhanced resilience of our model to the complexities of real-world drone-captured imagery.
Discussion
Advantages for Drone Deployment
The GDMA-YOLO model is specifically designed for the constraints of drone technology. With only 3.01M parameters and 6.12 MB storage, it can be easily deployed on embedded platforms such as NVIDIA Jetson or Google Coral. The inference speed of 136 FPS on an RTX 4060 (and an estimated 35-40 FPS on Jetson Orin) meets the real-time requirements of 4K video streams. The integrated distribution monitoring algorithm further adds value by converting raw detections into actionable spatial intelligence. For engineers and city planners deploying drone fleets, our system provides an end-to-end solution from pixel-level detection to macro-level hotspot analysis.
Limitations and Future Work
Despite its strong performance, GDMA-YOLO has a few limitations. First, the model’s accuracy degrades under extreme weather such as heavy rain or dense fog. Future work will explore lightweight image enhancement modules inspired by recent edge-side algorithms, specifically designed for drone vision. Second, the current distribution monitoring relies on accurate GPS and IMU data; low-cost drones may introduce errors that propagate into the heatmap. Incorporating visual-inertial odometry could improve precision. Finally, we aim to extend the system to multi-drone collaborative detection, leveraging formation flying and distributed computing to cover larger areas efficiently. This aligns with the broader trend of integrating drone technology into smart city infrastructure.
Conclusion
In this paper, we presented GDMA-YOLO, a lightweight and accurate detection framework for road garbage recognition and distribution monitoring using drone technology. By integrating Ghost-SPDConv, adaptive-weight AFPN, and an improved CIoU loss, we achieved a 7.57% improvement in mAP@50 over the YOLOv8n baseline without increasing model parameters. The additional coordinate mapping and KDE-based heatmap generation enable automated hotspot identification, facilitating efficient patrol planning. Our work demonstrates that advanced lightweight deep learning architectures, when tailored to the unique challenges of aerial imagery, can deliver both high performance and real-time capability, thereby advancing the role of drone technology in smart environmental governance.