Transfer Learning With CNNs for Segmentation of PALSAR-2 Power Decomposition Components

Abstract

Water/ice/land region segmentation is an important task for remote sensing, as it analyses the occurrence of water or ice on the earth's surface. Many previous deep learning researches effectively utilized multispectral satellite images for highly accurate water/ice/land region segmentation. However, the deep-learning-based segmentation of synthetic aperture radar (SAR) images still remains a challenging task due to the unavailability of enough labeled data. In order to overcome this issue, we designed a two-step deep-learning-based transfer learning model that needs a very limited number of labeled samples. The proposed approach consists of two models. The first model is a deep encoder-decoder 6SD to Landsat-8 multispectral translation model (DTF) that translates fully polarimetric PALSAR-2 6SD data to 6 new features. As for the second model (transfer learning), it utilizes the DTF features to fine-tune the model using the Landsat-8 multispectral pre-trained model for water/ice/land segmentation. Hereinafter, the proposed two-step model is referred to as DTF-TL. Also, a qualitative and quantitative analysis was carried out to evaluate the performance of the proposed model (DTF-TL) and compare it with various transfer learning methods. Overall, the DTF-TL model outperformed the other models with consistent and reliable water/ice/land segmentation results in terms of the recall (0.980), precision (0.981), F1-score (0.981), MIoU (0.962), and Accuracy (0.989).

Explainable, Physics-Aware, Trustworthy Artificial Intelligence: A paradigm shift for synthetic aperture radar

“…The introduced physical layer can be inserted in a deep neural network for substitution, extracting explainable and meaningful features, either as the input of a DNN or fused with DNN features in intermediate layers. A common way is to insert a physical layer into the input layer to obtain the polarimetric features for PolSAR image classification, including the elements of coherency matrix, Pauli decomposition features, etc [31], [32]. Similarly, the sub-aperture images are generated as the input for target detection [33].…”

ACTL: Asymmetric Convolutional Transfer Learning for Tree Species Identification Based on Deep Neural Network

“…Transfer learning, which is a technology using existing source domain data to solve relevant target domain problems [26]. In recent years, transfer learning has been widely used in land use classification, object detection, and semantic segmentation [27]- [30], which has greatly reduced the network's requirement for labeled data [31]. Lin et al adopted the method of instance transfer unsupervised classification of hyperspectral images [32] and projected source and target domain samples to a new feature subspace.…”