Results of training of convolutional neural network for satellite four-channel image segmentation are performed. Input images contain blue, green, red and near-infrared channels. The algorithm was trained to detect buildings and other urban areas. Modification of the U-Net neural network with two encoders was used. The values of Sorensen coefficient and Jaccard index were calculated for 16 different urban regions.
The goal of given investigation is to develop deep learning and convolutional neural network methods for automatically extracting the locations of objects such as water resource, forest and urban areas from given aerial images. We show how deep neural networks implemented on modern GPUs can be used to efficiently learn highly discriminative image features. For deep learning on supercomputer NVIDIA DGX-1 we used the marked image database UrbanAtlas, which contains images of 21 classes. Images obtained from the Landsat-8 satellites are used for estimation of automatic object detection quality. Object detection on aerial images has found application at urban planning, forest management, climate modelling, etc.
Investigation about using convolutional neural networks for detection geo-objects on the satellite images from Landsat-8 was presented. U-NET convolutional neural network architecture for implementing the recognition algorithm was used. The neural network was trained by the marked image base “Urban Atlas”. Urban Atlas contains images of 21 classes, but in current research was used 3 classes: “Forest”, “Agriculture” and “Water”. Images obtained from the Landsat-8 satellites are used for estimation of automatic object detection quality. To analyze the accuracy of the object detection algorithm, the selected regions were compared with the areas by previously marked by experts. Two modification of object detector were proposed and analyzing.
The aim of this research is to create a deep learning algorithm for automated forest areas segmentation on high-resolution aerial images. Loss values and Dice coefficient, which compares results of algorithms with real masks, was used to measure the quality of developed model. On the whole this paper demonstrates how convolutional neural network implemented on modern GPUs can be applied for the detection of forests on satellite images.