Complex Approach of High-Resolution Multispectral Data Engineering for Deep Neural Network Processing

A lot of terabytes of complex geospatial data are acquired every day, and it is used in almost every field of science and solves such problems as vegetation health monitoring, disaster management, surveillance, etc. In order to solve mentioned problems this data usually requires multiple steps of pre-processing before inferencing via machine learning algorithms. These steps may include such families of algorithms as image tiling or data augmentation. However, various studies focused on the basic concepts and research on techniques for remote sensing very high-resolution data pre-processing is in scarce. The current article proposes an approach for data engineering to improve results of processing via the deep learning techniques. The algorithm and dataset are developed, they combine image-tiling techniques and satellite imagery properties. A suggested solution is tested on featured deep convolutional neural networks, such as FuseNet and region-based Mask R-CNN. Described approach for data engineering demonstrates segmentation quality increase for 6%, which is a notable improvement, considering a number of objects of interest in modern high-resolution satellite imagery.