Work

Deep Learning-Based Classification of UAV Orthophotos Using MIDNet Architecture

Year: 2026
Type: article
Abstract: Photogrammetric methods have advanced significantly, enabling progress in cartography, construction, agriculture, and natural disaster monitoring. The integration of Structure from Motion (SfM) and orthophoto mapping has facilitated the generation of high-resolution, error-corrected images for various geospatial analyses. However, traditional deep learning-based Convolutional Neural Networks (CNNs) for orthophoto classification face challenges such as high computational costs, limited multiscale feature extraction, and suboptimal accuracy in complex landscapes. To address these limitations, this study introduces Multiscale Inception Depthwise Network (MIDNet), a novel CNN-based architecture designed for efficient and precise classification of UAV-derived high-resolution orthophotos. MIDNet leverages inception modules for multiscale feature extraction and depthwise separable convolutions to enhance computational efficiency without sacrificing performance. Experimental validation conducted on the generated reference dataset demonstrates that MIDNet outperforms the compared deep learning models, achieving an overall accuracy of 96.97%, an average accuracy of 95.96% and a kappa coefficient of 96.29%, surpassing DenseNet121 (OA: 96.32%, AA: 95.47%, Kappa: 95.50%) and InceptionV3 (OA: 96.60%, AA: 94.05%, Kappa: 95.85%), while maintaining the smallest model size (4.05 MB) and fastest testing time (8 seconds). These results underscore MIDNet’s superior classification accuracy, lightweight design, and suitability for resource-constrained environments, making it a compelling advancement in orthophoto classification techniques.
Source: DergiPark
Language: English
Open Access status: green
OpenAlex ID: W7148217513
FWCI:
Cites: 25
Cited by:
Related to: 10