UC San Diego

Mangrove forests are rich ecosystems that support our planet and the mankind in many unique ways. Unfortunately, these mangroves are declining at a rapid rate due to deforestation and other activities of the mankind. Monitoring and tracking of these mangrove trees is essential for their conservation. Machine Learning can be used for this purpose but to take advantage of the power of machine learning, image data needs to be captured for these mangrove ecosystems. This data collection is done using Unmanned Aerial Vehicles like drones in this research. Manually labeling the acquired image data for machine learning applications is a tedious and time-consuming task. This called for the development of architectures which could learn from limited labeled data and take advantage of the large amounts of unlabelled data. Such architectures are the semi-supervised semantic segmentation architectures and are studied in this research. We have shown how different semi-supervised models like the self-learning based Pseudo-Labelling architecture, the graph-based label propagation architecture and the deep leaning UNet-Autoencoder architecture perform on the task of mangrove segmentation in the aerial imagery. In order to evaluate different models we mainly look at the Intersection over Union because of it’s popularity in segmentation tasks. Overall, we see that the deep learning UNet-Autoencoder architecture performs the best with an average IoU of 0.78. Conceivably, the performance of each of the models improves as more labelled data is provided for training. The highest IoU obtained in this research is 0.9 with the UNet-Autoencoder when as much as 75% of the data provided is labelled.