Research on Multi-source Data Fusion Surveying and Mapping Modeling Technology Based on UAV Oblique Photography

By /

Traditional aerial surveying faces significant limitations in flight altitude and camera perspective, hindering high-resolution, multi-angle image acquisition in complex terrains. These constraints directly compromise the accuracy and efficiency of surveying and modeling workflows. Surveying drones, or surveying UAVs, overcome these barriers through flexible low-altitude operations and multi-sensor integration. This study proposes a Multi-source Attention Edge-constrained U-Net Convolutional Neural Network (MAEU-CNN) model, integrating UAV oblique photography with multi-scale feature extraction and deep learning to enhance geospatial data processing.

Methodology

Image Preprocessing

Surveying UAVs capture oblique imagery that requires noise reduction and enhancement. Key preprocessing steps include:

Grayscale Conversion:
$$Gray = \alpha_1 \times R + \alpha_2 \times G + \alpha_3 \times B$$
where $R$, $G$, $B$ denote RGB channels and $\alpha$ represents quantization weights.

Binarization:
$$f(x) = \begin{cases} 0, & x < N \\ 1, & x \geq N \end{cases}$$
Threshold $N$ separates foreground/background pixels.

Histogram Equalization:
$$P(r_k) = \frac{n_k}{N}, \quad s_k = \sum_{i=0}^{k} P(r_i)$$
This expands dynamic range for contrast enhancement.

Intensity Stretching:
$$g(x,y) = \frac{f(x,y) – \text{min}}{\text{max} – \text{min}} \times 255$$
normalizes pixel values across datasets.

MAEU-CNN Architecture

The model integrates four innovations for surveying UAV data:

  1. Multi-scale Feature Extraction: Parallel convolutions capture granular/textural patterns
  2. Edge Constraint Module: Boundary regularization loss:
    $$\mathcal{L}_{edge} = \sum \|\nabla Y_{pred} – \nabla Y_{true}\|^2$$
  3. Convolutional Block Attention (CBAM):
    Channel attention: $M_C(F) = \sigma(MLP(\text{AvgPool}(F)) + MLP(\text{MaxPool}(F)))$
    Spatial attention: $M_S(F) = \sigma(f^{7\times7}([\text{AvgPool}(F); \text{MaxPool}(F)]))$
  4. U-Net Fusion: Encoder-decoder structure with skip connections
Module Function Parameters
Encoder Blocks Feature downsampling 5×5 conv, ReLU
Multi-scale Fusion Feature aggregation Kernel sizes: 3,5,7
CBAM Decoder Attention-guided upsampling Transposed conv
Edge Constraint Boundary preservation Laplacian filter

Experimental Analysis

Testing used WHU Aerial and SpaceNet datasets on NVIDIA GTX2080Ti hardware. Performance metrics:

Table 1: Model performance comparison on 1600-image dataset
Metric U-Net SegNet MAEU-CNN Improvement
IoU 0.85 0.92 0.98 +15.3% vs U-Net
Ambiguity 0.21 0.18 0.12 -42.9% vs U-Net
Table 2: Inference time (milliseconds)
Structure U-Net SegNet MAEU-CNN
Residential 625 517 446 (-38.5%)
Industrial 590 482 412
Medical 568 451 376 (-33.8%)
Cultural 612 503 428

The surveying UAV-optimized MAEU-CNN demonstrates superior generalization across terrains. Attention mechanisms reduced feature ambiguity by prioritizing geomorphological elements, while edge constraints preserved critical infrastructure boundaries.

Conclusion

MAEU-CNN significantly advances surveying drone capabilities through:
1. Multi-scale fusion adapting to terrain complexity
2. Attention mechanisms reducing occlusion errors
3. Edge preservation maintaining structural fidelity
4. Computational efficiency enabling real-time processing

Surveying UAV systems integrating this model achieve sub-decimeter accuracy in heterogeneous environments, demonstrating 15.3% higher IoU and 42.9% lower ambiguity than conventional approaches. Future work will optimize model compression for edge-computing surveying UAV deployments.

Scroll to Top