Assessment of Deep Learning Based Image Segmentation for Identifying Floating Net Cages from Very High-Resolution Capella Synthetic Aperture Radar (SAR) Data

The availability of a constellation remote sensing satellite system using very high resolution (VHR) synthetic aperture radar (SAR) is beneficial to obtain information from the earth surface in a detail and timely manner. However, the ability to perform digital image classification using SAR data using conventional per-pixel based or object-based methods is hindered by limited variables (in forms of polarizations), where only single-, dual-, or quad-polarizations are available. Here, we aim to assess the potential application of deep learning-based image segmentation using UNet and Deep Residual UNet (ResUNet) to identify floating net cages using single polarization from Capella X-Band SAR data. The study area took place in Saguling Dam, West Java, Indonesia. Around 1086 tiles each for image patches and labels were collected and augmented to generate 7602 total pair of images and labels used in the training data. The results suggested that both UNet and ResUNet with data augmentation, produced the Intersection over Union (IoU) values of 88.3 to 89.7 (training), 89.0 to 90.3 (validation), and 88.7 to 90.1 (test). In addition, the model generalization performance when applied to the larger scene yielded the overall accuracy of 96.23 to 97.41, and IoU values of 66.255 to 79.253, for UNet and ResUNet with probability being set to 0.5, respectively. Although there was still misclassification produced from using deep learning-based image segmentation, the results suggested the potential from Deep Learning, to be used for classifying single-pol SAR data and to support rapid and automated monitoring using VHR Capella SAR. Additional fine-tuning conducted in this study proven the capability to reduce the false-positive from the initial model.