Land objects size approximation informational technology

Abstract

Despite rapid development of Big Data and data processing technologies, it is still quite a challenge to implement efficient pipelines for obtaining reliable and reasonable results. In re-mote sensing, thousands of gigabytes of data is acquired by satellite and aircraft vehicles, but only the modest amount of data is processed. Among the data processed, even less results re-mains reliable overtime due to the nature of satellite imagery. Nowadays, machine learning area affects nearly every domain of knowledge including remote sensing. A toolset of machine learning suggests ways to automate data processing which makes data pipelines constructure much more effective. Deep neural networks demonstrate the best results so far in image pro-cessing field. However, applying deep learning achievements in Remote Sensing still remain challenging in regard to multi-channel satellite imagery of a very-high resolution. Despite astonishing results deep learning field demonstrates in remote sensing competitions such as Open Street Map, these are applicable on regular RGB images of popular image formats such JPEG and similar. In current paper, a technology is suggested to apply deep neural network toolset in multispectral images of a very high resolution acquired from a commercial WorldView-3 satellite vehicle. The suggested approach covers topics of dataset preparation to fully replicate the results of the suggested neural network training. The proposed neural network solution output are segmentation maps. In addition to replicatable neural network structure, a particular use case is considered in detail which implements a fully automated technology of highly precise approximation of physical sizes of the land objects. The neural network performance is measured and compared to modern similar neural network solutions.