Instruments and Systems: Monitoring, Control, and Diagnostics Annotation << Back Semantic Segmentation of Seasonal Multi-Temporal Multispectral Images Using Fuzzy Logic Models and U-Net I.N. ISHCHUK, YU.YU. GROMOV, A.A. SOLODOVNIKOV, V.V. VOLOGIN, YA.I. LVOVICH The article considers the problem of semantic segmentation of infrared and visible images of vegetation, road surfaces, vehicles and engineering structures obtained from multispectral means of unmanned aerial vehicles (UAVs) for summer, autumn and winter periods of the year. Determining the change in the radiation temperature of the earth's surface (RTS) for different periods of the year using multispectral survey data of high spatiotemporal resolution is necessary for modeling the microthermal environment and identifying stable features for distinguishing man-made objects. The factors affecting the RTS were studied by correlating the RTS with meteorological data. Recognition of man-made objects was carried out using the U-Net deep learning network. The results showed that on clear and cloudy days there were significant differences in the RTS and the release of sensible heat for typical man-made surfaces, such as asphalt and concrete roads, metal ceilings, vehicles and engineering structures. On clear days, the daily fluctuations of different surfaces varied significantly, which was not observed on cloudy days. Artificial surfaces always acted as the main source of abnormal heat. Solar radiation and air temperature among meteorological factors were positively correlated with surface temperature. Air humidity was negatively correlated with the RTS and played a role in cooling the earth's surface. The experimental studies were carried out on the territory of an engineering town with an area of 11.4 hectares. Using a multispectral camera installed on a UAV, images with a resolution of 18 cm per pixel were obtained for different time intervals. The process of surveying the earth's surface was carried out six times during the day, starting at 6:00 and ending at 24:00, with intervals of 4 hours. Remote monitoring of the Earth was carried out in August, October, November and December 2025. Keywords: multi-temporal multispectral imaging, unmanned aerial vehicle, deep learning, object parameter estimation, semantic segmentation. DOI: 10.25791/pribor.8.2026.1697 Pp. 29-47.