Advanced method of land cover classification based on high spatial resolution data and convolutional neural network

Abstract

Based on modern satellite products Planet with high spatial resolution 3 meters, authors of this paper improved the neural network methodology for constructing land cover classification maps based on satellite data of high spatial resolution using the latest architectures of convolutional neural networks. The process of information features formation for types of land cover is described and the method of land cover type classification on the basis of satellite data of high spatial resolution is improved. A method for filtering artificial objects and other types of land cover using a probabilistic channel is proposed, and a convolutional neural network architecture to classify high-resolution spatial satellite data is developed. The problem of building density maps for the quarters of the city atlas construction is solved and the metrics for estimating the accuracy of classification map construction methods are analyzed. This will make it possible to obtain high-precision building maps to calculate the building area by functional segments of the Urban Atlas and monitor the development of the city in time. This will make it possible to create the first geospatial analogue of the product Copernicus Urban Atlas for Kyiv using high spatial resolution data. This Urban Atlas will be the first such product in Ukraine, which can be further extended to other cities in Ukraine. As a further development, the authors plan to create a methodology for combining satellite and in-situ air quality monitoring data in the city based on the developed Urban Atlas, which will provide high-precision layers of PM10 and PM2.5 concentrations with high spatial and temporal resolution of Ukraine.