Monitoring deforestation in Jordan using deep semantic segmentation with satellite imagery

Jordan is witnessing major transformations in its environmental and topographic features, where the desert dominates a large part of the territory, with very limited forest areas. Over the last three decades, Jordan has lost about a third of its natural forests at an annual rate of 1.6%. In this study, we develop a deep learning model to automatically monitor deforestation in Jordan based on the semantic segmentation of multitemporal Landsat-8 satellite images. Very few studies and datasets are devoted to monitoring forest cover changes using semantic image segmentation with deep neural networks. Therefore, we have collected a new benchmarking dataset of Jordanian forests between 2010 and 2020. The proposed model includes an efficient encoder-decoder architecture, with which we can extract a set of discriminating features that semantically assign every key image pixel into either forest or non-forest classes. The deep architecture is first initialised by a CNN-based pre-trained model and then it is fine-tuned on the forest images through an effective transfer learning procedure to improve its generalisation ability. Then, a set of key features are extracted by encoding each forest image into low-dimensional semantic maps to formulate the generic descriptors used in image segmentation. Finally, the output of the segmentation process is used to detect any dissimilarity in the forest area or boundaries using an absolute pixel-pixel similarity check over many years. The experimental results proved the effectiveness of the proposed model in segmenting the forest images and predicting any loss (deforestation) or gain (reforestation), and the model achieved an accuracy of 94.8% and MIoU of 82.1%. Moreover, the deep semantic features are discriminative enough to efficiently identify and estimate the amount of deforestation change in terms of accuracy and computational resources.