Convolutional Neural Networks for the Segmentation of Multispectral Earth Remote Sensing Images

This article describes a modernized approach to the segmentation of multispectral satellite images of Earth remote sensing using convolutional neural networks (CNN). Various modern algorithms for the segmentation of Earth remote sensing images are considered, including their disadvantages. The proposed approach is an improved algorithm developed by the authors and described in article 1. The previously proposed method for using CNN took into account some of the errors that can occur when processing CNN images using a sliding window. The current modification also excludes the appearance of these inaccuracies. Also, the method proposed in the article uses the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI), which have a high correlation coefficient with real objects present in the images. This paper provides examples that illustrate the work of the proposed method. For this approach, the number of CNN calls was modelled depending on the results obtained after building maps of the NDVI and NDWI indexes. The work also displays a comparison of time costs in the new and previously proposed approaches and also presented a comparison of segmentation results. © 2021, The Author(s), under exclusive license to Springer Nature Switzerland AG.