Vegetation zone segmentation in multispectral imagery

In the paper showing A U-Net-type convolutional neural network proposed for vegetation segmentation in multispectral imagery. The architecture of this network has modified and expanded to achieve better results with a smaller training dataset. The use of the network has made it possible to improve the solution to the problem of accurately identifying objects from the background based on existing information without the use of additional means of obtaining data. The processing of images of a corresponding nature is shown for 4-channel multispectral satellite images using a U-Net-shaped network trained on aerial images. Preprocessing showed the need to take into account the image formation model and perform post-training of the network based on the obtained data. A comparison conducted between the results of vegetation zone delineation in multispectral images using a convolutional neural network with a modified U-Net-like architecture and the Normalized Difference Vegetation Index (NDVI). The NDVI vegetation index uses data from various spectral channels and calculated as the ratio between intensities in the red and near-infrared channels of multispectral imagery, displayed as a number from -1 to +1, where a higher index value signifies denser green vegetation. Shown that for acceptable results, it is enough to use 4-channel images (NIR+RGB). The simulation performed in the MATLAB system.