Deep learning for mapping rainwater drainage networks using Remote Sensing Data

Resumo

Mapping rainwater drainage networks are traditionally lead from image visual interpretation. With the emergence

of automatic extraction algorithms, remote sensing areas began to modernize. To improve the spatial resolution image, sev-
eral works have addressed strategies to generate drainage networks, however, mapping accuracy and computational resources

needed to process this information are still a problem. This work explores the potential of applying different Deep Learning ap-
proaches to the process of extraction information in rainwater drainage networks through digital elevation models obtained by

the Shuttle Radar Topography Mission (SRTM). To perform this, we evaluated three different architectures: U-Net, DeepLab,

and Cycle GANs (Generative Adversarial Networks). The results show that the average intersection over union (IoU) to deter-
mine the drainage networks proved superior in relation to a decision tree used as baseline that proved unsatisfactory to solve the

proposed problem. However, the proposed UNet, DeepLab + FCN and Cycle GAN networks showed averages equal to 93.89%
81.13% and 92.42% respectively. These results indicate that it is possible to perform the geoprocessing of large-scale images
almost in real time, making it an excellent resource to contribute to the mapping of drainage networks.