Post-graduate theses
Current Record: 22 of 833
|
Identifier |
000460624 |
Title |
Autonomous fire front detection in satellite images with deep reinforcement learning |
Alternative Title |
Αυτόνομη ανίχνευση πύρινων μετώπων σε δορυφορικές εικόνες με χρήση βαθιάς ενισχυτικής μάθησης |
Author
|
Βελεσιώτη, Μαρία-Ιωάννα Σ.
|
Thesis advisor
|
Τσακαλίδης, Παναγιώτης
|
Reviewer
|
Τσαγκατάκης, Γρηγόριος
Τραχανιάς, Πάνος
|
Abstract |
An enduring concern that has troubled humanity for generations is the out-
break of wildfires in both forests and areas where communities thrive. This phenomenon has catastrophic implications for the ecosystems and wildlife in the impacted regions, as well as for human lives and properties that are lost in the wake
of these fire events. Earth observation satellites play a vital role in the surveil-
lance of land areas, primarily for fire detection. The spatial resolution of these
observations is a matter of great importance for the agencies responsible for gathering and processing space-based data. Obtaining high-resolution geographical
data systematically from private use satellites is a costly endeavor. In contrast,
publicly accessible satellites offer global coverage at considerably lower spatial resolution. Furthermore, there is a trade-off in temporal resolution analysis; public
satellites revisit the same geographical position after intervals of days or even
weeks to capture updates, while private satellites schedule targeted observations
based on the specific requirements of different users, thus not ensuring a systematic monitoring. Despite the potential benefits of satellite-based remote sensing,
current solutions are unable to provide both high spatial and temporal resolution
simultaneously. This presents an opportunity for research aimed at developing
a system that combines lower-resolution observation with the advanced capabilities of private satellites equipped with high-resolution cameras. Such a system
could enable more effective observation of specific areas of interest in the images.
Within a dual-satellite system featuring high and low-resolution cameras respectively, the challenge at hand is the prompt and accurate detection of fire fronts.
The proposed solution involves the utilization of a Deep Reinforcement Learning
framework to train an AI agent. The objective of this agent is to pinpoint the
location of the fire front within the low-resolution image by delineating a discrete
area of interest. This is achieved through the optimization of the agent’s predictive
model using Policy Gradient methods. Guided by the model’s predictions given
the low-resolution observations, the high-resolution camera captures observations
of the selected areas, i.e. those with a higher likelihood of fire. This thesis involves
conducting experiments to assess the algorithm’s performance and the influence of
its hyperparameters. We also investigate how the algorithm behaves across various dataset sizes. Furthermore, we delve into the dataset itself by performing
exploratory analysis to unearth patterns that could improve its handling during
the training process. Based on the insights gained from the data, we refine the
algorithm. Our analysis demonstrates that the proposed method exhibits better
generalization compared to alternative approaches and displays a greater capability to identify the fire front, even in its initial stages.
|
Language |
English |
Subject |
Fire detection |
|
Localization |
|
Ανίχνευση φωτιάς |
|
Εντοπισμός |
Issue date |
2023-12-01 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/4/a/7/metadata-dlib-1701082677-107718-13578.tkl
|
Views |
1042 |