Enhancing Earth Observation through AI: Semantic Segmentation of woodland areas with limited dataset

This thesis focuses on the application of advanced semantic segmentation techniques to multispectral satellite images, primarily for the analysis and classification of woodlands. A key characteristic of this work is the use of limited datasets, both in terms of the number of images and the accuracy of the masks. The thesis begins with an introduction defining the problem of semantic segmentation and outlining the research objectives, which involve comparing various methodologies to evaluate their effectiveness in classification tasks. Following this, a literature review explains the principles of neural networks, deep learning models, and the various techniques currently in use, including the concepts of transfer learning and fine-tuning, which are crucial for understanding the work conducted. Subsequently, the methodology used is detailed: the models employed, their implementations and parameters, the processing of satellite data to create ad-hoc datasets, and the training and testing pipeline, along with an explanation of the metrics used. Finally, the experiments are presented in their entirety, highlighting the characteristics of the dataset, the optimal parameters selected for the model, and culminating in a discussion of the results and conclusions. The findings indicate potential areas for improvement in segmentation accuracy, providing insights that could inform future research in the field of semantic segmentation in remote sensing.