Deep Learning-Based Segmentation for Terrain Classification in Aerial Imagery

It is important in agriculture the role of remote sensing applied to the classification of LandCover. Recently deep convolutional neural networks (CNN) have become increasingly and widely popular for their application to the study of monitoring and mapping of the land. In this work, we study existing semantic networks when applying to public datasets such as LandCover.ai. A comparison of fifteen neural networks is made and we find out that, in spite of they all have good performances, there are differences in the state of the outliers so we carry on a sistematical study of them. Our outcomes show that the most promising models achieve an accuracy of 99.11%, with a 71.5% of intersection over union (IoU) and 89.29% of recall, based on test set. We also conduct a study of the outliers dividing the misclassifications for tipology and find out that ANN, BiSeNetV2 and SETR-Naïve are the most effective models for handling the outliers. The dataset on which this research was carried out is publicly available at https://landcover.ai.linuxpolska.com/.