Multi-scale image segmentation in solving problems of oil spill detection on the basis of convolutional neural networks

Effective ecological monitoring of water areas of seas and oceans is difficult due to the huge size of the monitored areas. Therefore, machine learning methods are usually used to process the received primary photo and video material when solving problems of oil spill detection. High variability of the water surface of the world ocean and the large volume of each image frame complicate the effective work of neural networks. For this purpose it is proposed to use multiscale image segmentation, in ac-cordance with the K-means algorithm, which will allow to obtain several homogeneous fragments, the recognition of which will provide a solution to the problem of oil spill detection. And for effec-tive processing of the selected recognition objects represented as segments and expanding the capa-bilities of filters, spatial convolution is used, followed by complex integration of spatial pyramids. The desired model is built on a convolutional neural network. The use of the presented solution al-lows to reduce the computational load on the convolutional neural network, as well as to increase the reliability of detection of oil spill results when processing images with high variability of water sur-face parameters. A comparative evaluation of the proposed approach in relation to the model realiz-ing binary semantic segmentation technologies on the basis of complex spatial pyramidal association, which has received wide scientific validation, is presented. The presented solution is proposed to be implemented as a model realized by a series of CNNs. The proposed model allows to reduce the com-putational load of the network approximately 6 times in relation to the canonical approach, while en-suring the coincidence of detection results with the data of visual analysis in 67% of cases (the exper-iment was conducted on 50 images of water areas, 30 of which depicted spills of oil products, and the images did not contain third-party objects), which is almost 1.5 times higher than the capabilities of the tested model.