A neural network model for space image segmentation in monitoring of deforestation factors

The article deals with the important problem of applying neural network models in the tasks of monitoring the condition of forest areas using optical satellite images. Objective. The objective of this research was to develop a neural network model of forest felling suitable for automation of decoding multispectral Sentinel-2 satellite images in forest monitoring tasks in the example of the Khanty-Mansi Autonomous Okrug. Materials and Methods. The basis of the developed model was the forest harvesting satellite images segmentation procedure based on the convolutional neural network of deep learning. Sentinel-2 data interpretation was done using the modern QGIS geoinformation system. Over 990 pairs of multi-temporal space images of the forest territories of Khanty-Mansi Autonomous Okrug in winter (snow) period of 2018-2022 were processed to prepare the training data set. More than 70 000 images of the training data set and corresponding masks of forest cutting contours were generated with these images. Results. As a result of this work, a neural network model of forest felling was created, which implements efficient segmentation of space images, which allowed automating the labor-intensive procedure of interpretation of multispectral Sentinel-2 images in order to identify forest logging contours. The set of training data totally 70 000 frames was divided into the training, validation (test) and test (control) samples, the amount of which for the development of the neural network model of forest felling was 58 000, 10 000 and 3600 frames respectively. The novelty of the model is determined, on the one hand, by the use of winter (snow) satellite images for training the neural network and, on the other hand, by the use of pairs of images acquired before and after forest felling. Recall, Precision and F-measure metrics were used as the criteria to evaluate the accuracy of the trained neural network model, and the values were calculated from the test sample. The calculated accuracy of felling detection using the proposed model on the test sample of data on different metrics reaches 85-87%. Conclusion. The developed neural network model of forest felling can be used for monitoring and mapping of forest resources of the northern forestry territories of Russia using multispectral Sentinel-2 satellite images.