Application of deep learning neural networks for solving the problem of forest fire segmentation on satellite images

The aim of the work is to develop algorithms for the semantic segmentation of forest fire areas on satellite images. Despite the active development of computer vision algorithms, today there are a number of problems in this area that have not been fully solved and do not provide the required accuracy of work. Therefore, today there is a need for the development of algorithms and software that provide high quality image segmentation. The analysis of existing algorithms for image segmentation was carried out and it was revealed that the most suitable algorithms for solving this problem are deep learning neural networks. The machine learning libraries Keras, TensorFlow and PyTorch were reviewed. The library performance was tested on a set of 60,000 images. In the process of research, the PyTorch library showed the best results, so it was decided to use it to develop algorithms. Convolutional neural network consisting of 20 layers has been developed. The neural network was trained using a generated set of 50 images of Earth remote sensing with a resolution of 8000x8000. The set of images was selected from the Landsat 8 satellite database. The main selection criteria concerned the size of the scene, as well as the number of images taken by the satellite during the day. The generated set of images contains data of the following classes: forest fire (red); burnt-out area (black); smoke from a fire (white); reservoirs (blue); forest (green). For a set of images, augmentation was performed, that is, modification of the data for training. Using this method improves the generalizing ability of the neural network, adds new training examples that the neural network has not yet seen and does not provide an opportunity to retrain. As augmentation, the following modifications were performed: image rotation by an arbitrary degree; compression along the axes; stretching along the axes; mirroring along the axes; Gaussian Blur; change in brightness and contrast. The training included 50 epochs, each of which contains 2000 iterations. When choosing an algorithm for learning a neural network, the following algorithms were considered: Adam - adaptive moment estimation; Adagrad - adaptive gradient; RMSProp - gradient descent with momentum. During the research, the best results were obtained using the Adam algorithm. A comparison of the results of the proposed neural network with some analogues is presented. A comparative study of the accuracy of the segmentation algorithms was carried out on a set of reference and test images subjected to noise distortions. To compare the segmentation results, the boundaries of the segmented objects were used, which is a set of points that do not depend on the shading of the segments. To measure the segmentation results, two metrics were used: mean and Hausdorff distance. The study of the quality of work of a number of algorithms showed that they behave unstably when the image is noisy and blurred. Thus, we can conclude that it is advisable to clean the image from noise and increase its clarity before the segmentation procedure. The accuracy of the developed neural network is 94.22%. For the classes of objects, the accuracy was the following: fire - 93.6%; burnt-out area - 95.7; smoke - 87.6; reservoirs - 96.9; forest - 97.3. This result is the best in comparison with the presented analogs. However, the developed system is somewhat inferior to some analogues in terms of such indicators as fire, burnt out area, smoke. However, in such classes as forests, reservoirs, it wins.

Keywords

Authors

References