The Robustness Evaluation Of The Semantic Segmentation Model In High-resolution Satellite Images

These days, more and more satellite images are available due to the increase in satellites. As a result, deep learning methods are widely applied in satellite image analyses. However, since satellite images are taken from space, they have different properties such as radiometric effect and viewing angle. Therefore, it is crucial to evaluate the robustness of deep learning models. In this study, we collected a large number of high-resolution satellite images in Okinawa, the southernmost part of Japan, and built a semantic segmentation model. For evaluation, we tested the model with three different images:(i)an image included in train data, (ii)an image not included in train data but from same area in Okinawa, (iii)an image in Ehime, 1000 km north from Okinawa. The Okinawa image(ii), which was not included in the train data had nearly the same accuracy as the Okinawa image(i). On the other hand, the Ehime image showed highly lower segmentation accuracy for some classes such as Vegetation, because those classes had different characteristics in this area. However, the major classes such as Trees and Water had very good accuracy and the predicted image looked well qualitatively. Moreover, when we added the Ehime image to the train data, the accuracy drastically increased. In summary, we confirmed enough robustness of the semantic segmentation model so it can be applied to different images.