Comparison of Deep Learning Algorithms for Image Segmentation on Satellite Images

Recent advancements in deep learning have significantly contributed to the development of high spatial resolution (HSR) land cover mapping. However, the distinct geographic patterns between urban and rural areas have limited the generalizability of deep learning algorithms across these domains. To address this challenge, separate datasets for rural and urban environments have been proposed in the literature, aiming to achieve more reliable results in real-world applications. In this study, we utilize the publicly available LoveDA HSR dataset for model and parameter comparison. Experiments were conducted on two distinct scenarios: rural and urban areas. The combination of the Adam optimizer, Dice loss function, and UNet++ architecture exhibited the highest performance in both datasets. A weighted average of this combination, based on the number of test samples, was calculated for both groups, yielding a final performance score of 62.14% in terms of mean Intersection over Union (IoU).