U-Net Based Deep Learning Approach for Land Classification in Aerial Imagery

With the invention of aerial imagery and satellite photography, reading the content of the images and having a better and easier understanding of their analysis has been challenging. Although high-quality aerial images include all the details in the image, how can scientists be sure of what they observe within the image? Also, in some cases where the images have low quality, the weather of the region is cloudy, or even the region is covered (lakes covered by plants, roads covered by trees, and so on), and in some other similar challenging cases, we have difficulty understanding the content of the images easily. Considering this problem, we are looking for an approach that uses neural networks to analyse and read the content of aerial images and more precisely detect the type of land within the images. By utilizing convolutional neural networks with the U-Net model, we propose an automated and reliable solution for perceiving land types in aerial imagery. We evaluated the performance of our system using performance metrics such as accuracy, precision, and F1 scores for each land type. The results showed that our system achieved high accuracy and precision for specific land types. We believe that our system can help those who need to analyse aerial imagery better understand the content of the images and make more informed decisions. Keywords: land type classification, unet classification, land type detection