Development of a Computer Vision System Using Machine Learning to Access the Overgrowing of Higher Aquatic Vegetation in the Tsimlyansk Reservoir

Modeling the process of overgrowing of shallow waters with higher aquatic vegetation is of great practical importance for the fisheries industry of our country and is an integral part of monitoring water bodies. The paper presents the results of developing a computer vision system using machine learning based on the SegNet and U-Net architectures to assess the overgrowing of the Tsimlyansk Reservoir with higher aquatic vegetation. A dataset consisting of 200 pairs of Landsat images covering 24 different sections of the Tsimlyansk Reservoir for different years, as well as the corresponding overgrowing marks, was used to train and test the models. The SegNet training process lasted for 50 epochs, U-Net was trained for 30 epochs. Each training epoch included iterations on the training data, calculating the loss function, backpropagating gradients, and updating weights using an optimizer. After each epoch, the model was validated on a validation sample to assess its performance. The accuracy of the SegNet model was 0,869, U-Net – 0,881. To assess the quality of overgrowth segmentation, the Jaccard coefficients (IoU) were measured on the test sample. The U-Net model showed an IoU of 0,665, while SegNet showed an IoU of 0,633.