2025 , Volume 30, № 3, p.111-126

High resolution satellite images allow creating large-scale tree species maps for remote and extended territories. Machine learning methods can automate this process. Currently, convolutional neural networks based on the U-Net architecture is widely used, however, it requires careful tuning of the model and parameters.

In this paper the problem of tree species semantic segmentation on high spatial resolution multispectral satellite images using U-Net-like models is studied. For this purpose, several datasets were prepared consisting of multispectral and panchromatic bands combinations of the WorldView-2 satellite image. Segmentation was carried out using four CNN models based on the U-Net architecture: U-Net, U-Net++, Attention U-Net and Trans U-Net.

The influence of various parameters on the semantic segmentation quality was examined. It was determined that the most stable training is achieved with the percentage linear stretching as a method for spectral bands data normalization. A comparison of the segmentation results using various combinations of spectral bands was carried out. In most cases, the use of blue, green, red and NIR_1 bands combination led to higher IoU-metric compared to the use of all 8 bands. According to the results of the experimental studies, it can be noted that the Attention U-Net model provides the highest accuracy in recognizing tree species in forest stands.

Also, two approaches for utilizing the panchromatic band in the segmentation process were presented, It leads to a significant improvement of the quality of the obtained results, which indicates the importance of the panchromatic band for tree species identification.

The results of this work can be directly applied in forest inventory tasks and forest condition monitoring byreducing the time spent on production of tree species maps.