Satellite and aerial images can play a crucial role in supporting the planning and coordination of global change research by aiding in the development of research strategies and facilitating the implementation of methodologies. Deep learning algorithms have advanced, allowing image sensors to comprehend the scene for the target object with greater accuracy, notably in the domain of segmentation. Due to variable lighting conditions, irregular road geometries, and fuzzy boundaries between the road and other objects, segmenting a road scene from colour photos using a computer vision approach might be difficult. We have used U-Net, Seg-Net, and fully convolutional network (FCN) models to achieve this goal and to clearly separate the road from the non-road component. According to the studies, U-Net outperformed Seg-Net and FCN-32 in terms of mIoU and dice coefficient. Also, the well-known encoder-decoder structure is used by deep convolutional neural networks like SegNet and UNet to segment pictures. These networks' encoders downscale the image gradually while expanding the receptive field in order to encode the features. After recovering the features' spatial dimensions and making the final predictions in full resolution, the decoder utilizes up-sampling techniques to increase the resolution or size of the features.

satellite images, road detection, UNET, deep learning, computer vision