Sea ice classification using a CNN-Transformer hybrid and AutoIce challenge dataset
Author
Esbensen, Malthe Aaholm
Term
4. term
Publication year
2023
Submitted on
2023-06-01
Pages
58
Abstract
Dette speciale undersøger brugen af TransUNet, en dybdelæringsmodel der kombinerer konvolutionsneurale netværk (CNN) og Transformers for at fange både lokale og globale mønstre, til automatisk at omdanne satellitbilleder til tre haviskort: Sea Ice Concentration (SIC), Stage of Development (SOD) og isflage-størrelser (FLOE). Præcise og løbende haviskort er vigtige for klimaovervågning og for mere sikker og forudsigelig sejlads i arktiske farvande. Der blev trænet flere varianter af TransUNet med forskellige Transformer-konfigurationer. Vi fandt, at antallet af Transformer-lag og størrelsen på de billedudsnit (patches), modellen bruger, havde stor betydning for ydeevnen. Den bedste model blev udvalgt og trænet i 120 epoker (fulde gennemløb af træningsdata). Ydelsen blev målt med passende mål: R^2 for SIC og F1-score for SOD og FLOE. Den bedste valideringsscore nåede 92,15 % samlet, og uafhængige tests gav en samlet score på 86,22 %. For SIC alene opnåede modellen 86,91 % nøjagtighed, en forbedring på 0,57 % i forhold til tidligere arbejde. Resultaterne viser, at TransUNet er en anvendelig metode til semantisk segmentering (at tildele hver pixel en klasse) i fjernmåling til at producere brugbare haviskort.
This thesis explores using TransUNet, a deep learning model that combines convolutional neural networks (CNNs) and Transformers to capture both local and global patterns, to automatically turn satellite images into three sea-ice map products: Sea Ice Concentration (SIC), Stage of Development (SOD), and floe size (FLOE). Accurate, timely sea-ice maps are important for climate monitoring and for safer, more predictable shipping in Arctic waters. Several TransUNet variants were trained with different Transformer configurations. We found that the number of Transformer layers and the image patch size had a strong impact on performance. The best model was selected and trained for 120 epochs (full passes over the training data). Performance was evaluated with task-appropriate metrics: R^2 for SIC and F1 score for SOD and FLOE. The best combined validation score reached 92.15%, and independent tests achieved a combined score of 86.22%. For SIC alone, the model reached 86.91% accuracy, a 0.57% improvement over prior work. These results show that TransUNet is a viable semantic segmentation approach (labeling every pixel) in remote sensing for producing operational sea-ice charts.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents