Improving the U-Net Segmentation Model for Land Cover Classification in Satellite Image Processing

The development of machine learning methods for onboard satellite processing is important in order to facilitate the filtering of collected data samples to maximize the use of the device's limited resources. Land cover classification can be used to focus the collected data on certain terrain types by utilizing classification methods to determine the class probabilities of individual pixels in a collected satellite image. The importance of the accuracy of the segmentation model used for such a task is important in order to avoid the trashing of data samples that offer significant information and the prioritization of data samples which offer less in terms of usable information, which in the case of land cover classification is determined by which terrain features may be prioritized over others. This study focuses on the U-Net segmentation architecture and performs an experimental study on the effects on two aspects on the training of a segmentation model for increased performance. This includes the division of the images in the dataset into smaller patches and the replacement of the CNN encoder of the segmentation architecture. The changes made to the baseline model introduced an increase in the IoU score from 0.68 to 0.7273.