January 23, 2025

• Interesting article Ten deep learning techniques to address small data problems with remote sensing (Dec 2023) - 78x cited.
• Figure 1 shows that only 5% of the papers are about water areas:
  
• Also looked into list of reviewed papers. Only one paper (Jan 2021) was about wetlands (Flevoland data - radar based).
• From the 80 reviewed papers, more than half was about classification, only 24% was on segmentation. On average, the papers had 242 samples per class (mean).
• Only 13/80 of the papers used IoU as metric. Only one used Dice (Deep Learning Based Land Cover and Crop Type Classification: A Comparative Study (May 2021)
• Safonova gives only two examples of biodiversity tracking.
• Quote from section 3.2: "The small data problem becomes more pronounced when analysing high-resolution data".
• Figure 2 only shows 4 self-supervised papers. Could only find one in the list: Self-Supervised Feature Learning for Multimodal Remote Sensing Image Land Cover Classification. It was validated on 4 datasets. Only saw some sparse water in the Houston 2018 dataset. Fig 11 shows the classification accuracy increase with the number of labeled samples increases.
• In section 4.2 the other three papers are introduced. All four are related with the hyperspectral image classification problem.
• The paper AeroRIT: A New Scene for Hyperspectral Image Analysis (April 2020) compares different U-Nets, is concerned with low resolution and uses both mIoU and mDice as measure.
• In the comparison with state-of-the-art, the proposed method shows suspectable high accuracies above 99%. To my suprise semi-supervised outperformed supervised methods.
• According to Fig 3, I would say that the labeled status is moderate, which leads to semi-supervised or weakly supervised learning. Instead self-supervised learning is used, which in the branch of rarely labeled, for non-rare instances, where no expert knowledge is available:
  

Previous Labbooks

• Labbook 2024.
• Labbook 2023.
• Labbook 2022.
• Labbook 2021.