2024 , Volume 29, № 5, p.100-112

The problem of determining the species composition for forest stands using remote sensing data has been attracting significant attention for decades. The relevance of this problem is due to the constant improvement and development of remote sensing tools and technologies including the intensive spread of unmanned imagery technologies in recent years.

Remote sensing based on small unmanned aerial vehicles (UAVs) is a rapidly developing technology. Compared to manned aircraft, UAVs are an easy-to-use and low-cost tool for remote sensing of forests. Survey cameras mounted on UAVs allow data collection even in cloudy conditions. UAVs can produce flexible temporal resolution and extremely high (up to several centimeters) spatial resolution of images, where tree features can be seen at the level of branches and even leaves. Therefore, not only spectral features but also spatial (textural and geometric) features play a significant role in tree species recognition in ultra-high spatial resolution images.

Currently, the most effective approach to analyse high spatial resolution aerospace images is to apply deep learning methods based on convolutional neural networks (CNNs). This is due to the fact that CNNs are specifically designed to analyse spatial patterns and they don’t require “manual” extraction of spatial features. Unlike traditional image segmentation and recognition algorithms, CNNs provide the ability to analyse spectral and spatial features of objects in an image jointly.

The paper investigates the possibility of automatic identification and classification of aspen and birch stands in RGB images of ultra-high spatial resolution obtained from unmanned aerial vehicles. To solve the problem of semantic segmentation we compared the performance of convolutional neural networks based on different architectures: U-Net, FPN, PSPNet, Linknet, DeepLabV3, DeepLabV3+. The results showed that DeepLabV3+ architecture and modifications of U-Net with Inception-blocks allow to achieve the best results of semantic segmentation, reaching the highest values of IoU (∼0.83) and F-score (∼0.91) quality metrics.