Comparison of semantic segmentation deep learning models for land use mapping: Site characterization for construction site preparation

Land cover/land use (LCLU) mapping is an important tool in automating site characterization for construction site preparation. By capturing multispectral images from an aerial perspective, LCLU mapping can be performed to characterize the construction site in a safe and efficient manner. The segmented LCLU maps can be used to determine the trafficability of the site and which construction equipment is best suited for clearing the site. While many traditional machine learning methods have been used for image segmentation, convolutional neural networks and deep learning approaches consistently outperform them. Three semantic segmentation models (PSPNet, U-Net, and Segnet) and three base models (VGG, ResNet, and MobileNet) are compared for the task of LCLU mapping. These models are pretrained on the ImageNet dataset and fine-tuned using datasets collected in central Illinois. The models are modified to include an additional channel for near-IR (NIR) images. Seven land cover classes (bare soil, water, roads/pavement, vegetation, trees, built-up, and an unknown category) were determined with an accuracy 82.71% by model ResNet/SegNet. Adding the NIR imagery achieved an accuracy of 74.74% by the semantic segmentation model VGG/PSPNet.