Detection of rural roads from planet images using convolutional neural networks

• Published in: Boletim de Ciências Geodésicas Vol. 31; pp. 1 - 15
• Main Authors: Do Crato, Jorgiana Kamila Teixeira, Jijón-Palma, Mário Ernesto, Centeno, Jorge António Silva
• Format: Journal Article
• Language: English
• Published: Curitiba Universidade Federal do Paraná, Centro Politécnico 01.01.2025
• Subjects: Artificial intelligence; Artificial neural networks; Deep learning; Feature maps; Image resolution; Mapping; Metropolitan areas; Neural networks; Optimization; Parameter identification; Recognition; Remote sensing; Roads & highways; Rural areas; Rural roads; Satellite constellations; Satellite imagery; Semantics; Target detection; Training; Unmanned aerial vehicles; Brazil
• ISSN: 1413-4853; 1982-2170; 1982-2170
• DOI: 10.1590/s1982-21702025000100004

Updating maps in Brazil is hindered by considerable obstacles, primarily to the high costs associated with it and the difficulty of accessing the regions. Moreover, the accelerated rate of environmental transformation, particularly in rural settings, represents an additional challenge. This study proposes to use high-resolution satellite images from Planet constellation, in conjunction with artificial intelligence, specifically UNet, to automatically identify rural roads in the metropolitan region of Curitiba, Paraná, Brazil. The objective is to identify the optimal parameters for automating the detection of rural roads. The UNet, with its distinctive U-shaped architecture, is highly effective in segmenting and detecting targets while simultaneously preserving the feature maps in each convolution. In this study, the network was trained on satellite images containing rural roads, resulting in segmented maps with an encouraging 91.95% accuracy in road detection. Nevertheless, further improvement is possible, as evidenced by the method's precision of 75.83% and F1-Score of 69.07%. These outcomes indicate the possibility of enhancement through the expansion of the training dataset, thereby better addressing the network's recognition constraints. One potential avenue for optimizing detection using the methodology would be the incorporation of supplementary training samples, which could potentially mitigate the network's recognition limitations.