Impact of Spectral and Meteorological Data Fusion on the Accuracy of Woody Plant Identification Using Deep Machine Learning Methods

Identification of woody plant species and their states has an important role in monitoring of forest resources, as well as joint application with remote sensing. For remote monitoring there is still a problem of poor repeatability of the result. Species identification models built from spectral characteristics for a specific time interval or a specific area are often not suitable for other time intervals and areas. This paper discusses a poor repeatability study approach based on analyzing the seasonal dynamics of the annual cycle to accurately identify different maple species. The main research methods are machine learning (ML) algorithms: Random Forest and Gradient Boosting. Including a method to fuse spectral data from a hyperspectral camera with meteorological data such as temperature, precipitation and day length affecting the development of woody plants in the annual cycle. The highest accuracy rate was 58.72%. Data fusion allows the relationship between spectral characteristics and climatic conditions to be considered, resulting in a 13.21% improvement in the quality of species identification by machine learning algorithms. The results allow us to assess the variation and interrelationship of maple species over the annual cycle, as well as the differences in spectral characteristics within each annual cycle. The use of spectral and climatic data fusion made it possible to create a comprehensive model that considers spectral characteristics and their relationship with meteorological conditions, thus increasing the accuracy of species classification, and to evaluate the accuracy of machine learning algorithms in the seasonal dynamics of the annual cycle for all studied species.