Earth Observation is the gathering of information about planet Earth's system via Remote Sensing technologies for monitoring land cover types and their changes. One of the most attractive use cases is the monitoring of polar regions, that recently observed some dramatic changes due to global warming. Indeed drifting ice caps and icebergs represent threats to ship activities and navigation in polar areas, and the risk of collision with land-derived ice highlights the need to design a robust and automatic Sea-Ice classifier for delivering up-to-date and accurate information. To achieve this goal, satellite data such as Sentinel-1 Synthetic Aperture Radar images from the European Union's Copernicus program can be given in input to a Deep Learning classifier based on Convolutional Neural Networks capable of giving the content categorization of such images as output. For the task at hand, the availability of labeled data is generally scarce, therefore the problem of learning with partially labeled data must be faced. Therefore, this work aims at leveraging the broader pool of unlabeled satellite data available to open up new classification solutions. This thesis proposes a Semi-Supervised Learning approach based on Generative Adversarial Networks, that takes in input both labeled and unlabeled data and outputs the classification results exploiting the knowledge retrieved from both the data sources. Its classification performance is evaluated and it is later compared with the Supervised Learning approach and the Transfer Learning approach based on pre-trained networks. This work empirically proves that the Semi-Supervised Generative Adversarial Networks approach outperforms the Supervised Learning method, improving its Overall Accuracy by at least 5% in configurations with less than 100 training labeled samples available in the use cases under evaluation, achieving performance comparable to the Transfer Learning approach and even overcoming it under specific experimental configurations. Further analyses are then executed to highlight the effectiveness of the proposed solution.
L'Osservazione della Terra consiste nel collezionare informazioni attraverso il Telerilevamento al fine di monitorare le variazioni della superficie terrestre. Uno dei casi più interessanti è il monitoraggio delle zone polari, poiché iceberg e calotte di ghiaccio alla deriva comportano una minaccia per le attività di navigazione nelle zone polari. Pertanto, il rischio di collisione con esse rende necessario un sistema automatico e robusto di classificazione Mare-Ghiaccio che produca informazioni accurate. A tale scopo, dati satellitari come immagini Radar ad Apertura Sintetica raccolte da Sentinel-1 per il programma Europeo Copernicus possono essere fornite in input ad un classificatore ad Apprendimento Profondo basato su Reti Neurali Convoluzionali capace di produrre in output la classificazione di tali immagini. In tale contesto la disponibilità di dati etichettati è limitata ed è pertanto necessario apprendere con dati parzialmente etichettati. Questo lavoro punta a sfruttare il numeroso insieme di dati satellitari non etichettati per proporre nuove soluzioni di classificazione. Questa tesi propone un approccio di Apprendimento Semi-Supervisionato basato su Reti Generative Antagoniste, che riceve in input sia dati etichettati che non etichettati e produce in output i risultati di classificazione facendo leva dulla conoscenza acquisita da entrambe le sorgenti. Le performance di classificazione raggiunte da tale approccio sono valutate e poi comparate con l'Apprendimento Supervisionato e l'Apprendimento per Trasferimento basato su reti pre-addestrate. Questo lavoro dimostra empiricamente che l'approccio Semi-Supervisionato basato su Reti Generative Antagoniste supera in prestazioni l'Apprendimento Supervisionato, migliorandone l'Accuratezza Complessiva di almeno il 5% in configurazioni con meno di 100 dati etichettati disponibili durante la fase di apprendimento nei casi d'uso valutati, raggiungendo performance comparabili a quelle dell'Apprendimento per Trasferimento e superandole in configurazioni sperimentali specifiche. Infine, ulteriori analisi sono condotte al fine di evidenziare l'efficacia della soluzione proposta.
Deep learning for sea-ice classification on synthetic aperture radar (SAR) images in earth observation. Classification using semi-supervised generative adversarial networks on partially labeled data
Staccone, Francesco
2019/2020
Abstract
Earth Observation is the gathering of information about planet Earth's system via Remote Sensing technologies for monitoring land cover types and their changes. One of the most attractive use cases is the monitoring of polar regions, that recently observed some dramatic changes due to global warming. Indeed drifting ice caps and icebergs represent threats to ship activities and navigation in polar areas, and the risk of collision with land-derived ice highlights the need to design a robust and automatic Sea-Ice classifier for delivering up-to-date and accurate information. To achieve this goal, satellite data such as Sentinel-1 Synthetic Aperture Radar images from the European Union's Copernicus program can be given in input to a Deep Learning classifier based on Convolutional Neural Networks capable of giving the content categorization of such images as output. For the task at hand, the availability of labeled data is generally scarce, therefore the problem of learning with partially labeled data must be faced. Therefore, this work aims at leveraging the broader pool of unlabeled satellite data available to open up new classification solutions. This thesis proposes a Semi-Supervised Learning approach based on Generative Adversarial Networks, that takes in input both labeled and unlabeled data and outputs the classification results exploiting the knowledge retrieved from both the data sources. Its classification performance is evaluated and it is later compared with the Supervised Learning approach and the Transfer Learning approach based on pre-trained networks. This work empirically proves that the Semi-Supervised Generative Adversarial Networks approach outperforms the Supervised Learning method, improving its Overall Accuracy by at least 5% in configurations with less than 100 training labeled samples available in the use cases under evaluation, achieving performance comparable to the Transfer Learning approach and even overcoming it under specific experimental configurations. Further analyses are then executed to highlight the effectiveness of the proposed solution.| File | Dimensione | Formato | |
|---|---|---|---|
|
Francesco_Staccone_Politecnico_di_Milano_Thesis.pdf
accessibile in internet per tutti
Descrizione: Tesi di Laurea Magistrale in Computer Science di Francesco Staccone
Dimensione
11.52 MB
Formato
Adobe PDF
|
11.52 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/165343