Satellite Image Segmentation Using Deep Learning for Deforestation Detection

Abstract

The problem of automatic monitoring the deforestation process is considered for efficient prevention of illegal deforestation. Image segmentation model on the basis of U-Net family of deep neural networks (DNNs) was created. The forest/deforestation dataset was collected by parsing areas of Ukrainian forestries, where satellite images of 512×512 pixels contain areas with forest, deforestation, and other areas. The dataset with satellite imagery and segmented masks was uploaded at GitHub repository where it is available with the correspondent code for distributed training on tensor processing units (TPU). To overcome the imbalance of created dataset the hybrid loss function was created and tested in the training environment. K-fold cross validation and numerous runs for different random seeds were conducted to prove the model and dataset usefulness and stability during the training and validation process. The following asymptotic values of intersection over union (IOU) mean $(IOU_{mean})$ and standard deviation $(IOU_{std})$ were obtained after more than 100 epochs: $IOU_{mean}^{kfold}=0.52, IOU_{std}^{kfold}=0.03$ for cross-validation, and $IOU_{mean}^{random}=0.51, IOU_{std}^{random}=0.03$ for various random seed initialization. These results demonstrate that variation of images in the dataset and randomness of initialization have no significant effect on model performance, but the future research will be needed in the view of the possible increase of datasets where performance could be improved by the larger data representation, but some decrease of performance could be observed due to possible wider data variability. It is especially important for deployment of U-Net-like DNNS on devices with the limited computational resources for Edge Computing layer.