Mapping urban tree fragments with orbital images and deep learning

Abstract

Vegetation is one of the main components of the environment, as it has a great positive influence on environmental quality, making it a very relevant object of study. Currently, remote sensing is one of the best options for studying the spatial behavior of vegetation, capable of generating a large volume of data. However, this large volume of data creates an additional demand, requiring efficient processing techniques to transform it into useful information. The objective of this work was to perform the segmentation of vegetation fragments in an urban context using two deep learning networks and a machine learning algorithm. The U-Net, DeepLabV3+, and Random Forest networks were used in this work. For the training of these networks, a dataset composed of Planet Scope images of urban areas from four different dates was used. As a result, it was observed that the DeepLabV3+ model had superior performance to the others when trained with a multitemporal dataset, thus obtaining Accuracy (0.936), Precision (0.847), Recall (0.804), and F1-Score (0.799). The results indicate that training the networks with a multitemporal dataset promoted an increase in metrics as well as homogeneous results among the different segmented images.