The Impact of High and Super Spatial Resolution Remote Sensing Images Datasets Composition on Training and Accuracy of Geofields Semantic Segmentation Neural Networks on Example of a Different Earth’s Surface Classes Recognition

The remote sensing images and masks datasets that are used for different earth’s surface classes recognition tasks can significantly impact neural network models learnability and semantic segmentation results that are received after model training. Datasets composition problematic usually are not investigated from these datasets point of view as it is viewed as neural network data processing problematic in general in each specific remote sensing data semantic segmentation task. In geofields semantic segmentation context general problematic for object semantic segmentation on aerial and satellite data that includes such main problems as images with different spectral characteristics, images with shadows, images with “different object, same spectrum”, images with incorrect annotation and images with class imbalance borderline cases is determined. Mentioned problems are considered as problems of original geofields set representation in datasets. As result of a different earth’s surface classes semantic segmentation experiment with U-Net, STT and MF-CNN it was determined that class imbalance borderline cases and using images with resolution in which class imbalance is higher than using their crops reduce learning ability and recognition accuracy of neural networks and class imbalance borderline cases deletion based data selection in data preprocessing process leads to accuracy increase and models training time decrease.