Hybrid Variability Aware Network (HVANet): A Self-Supervised Deep Framework for Label-Free SAR Image Change Detection

Abstract

Synthetic aperture radar (SAR) image change detection (CD) aims to automatically recognize changes over the same geographic region by comparing prechange and postchange SAR images. However, the detection performance is usually subject to several restrictions and problems, including the absence of labeled SAR samples, inherent multiplicative speckle noise, and class imbalance. More importantly, for bitemporal SAR images, changed regions tend to present highly variable sizes, irregular shapes, and different textures, typically referred to as hybrid variabilities, further bringing great difficulties to CD. In this paper, we argue that these internal hybrid variabilities can also be used for learning stronger feature representation, and we propose a hybrid variability aware network (HVANet) for completely unsupervised label-free SAR image CD by taking inspiration from recent developments in deep self-supervised learning. First, since different changed regions may exhibit hybrid variabilities, it is necessary to enrich distinguishable information within the input features. To this end, in shallow feature extraction, we generalize the traditional spatial patch (SP) feature to allow for each pixel in bitemporal images to be represented at diverse scales and resolutions, called extended SP (ESP). Second, with the carefully customized ESP features, HVANet performs local spatial structure information extraction and multiscale–multiresolution (MS-MR) information encoding simultaneously through a local spatial stream and a scale-resolution stream, respectively. Intrinsically, HVANet projects the ESP features into a new high-level feature space, where the change identification becomes easier. Third, to train the framework effectively, a self-supervision layer is attached to the top of the HVANet to enable the two-stream feature learning and recognition of changed pixels in the corresponding feature space, in a self-supervised manner. Experimental results on three low/medium-resolution SAR datasets demonstrate the effectiveness and superiority of the proposed framework in unsupervised SAR CD tasks.

Change detection in SAR image based on weighted difference image generation and optimized random forest

“…In order to quantitatively analyse the results of change detection and evaluate the proposed method, the following evaluation metrics are employed: FP , that is, pixels with no change in the ground detected as the number of pixels with a change, FN , that is, pixels with a change in the ground detected as the number of pixels without a change, the total number of errors OE is the sum of FP and FN , the percentage of correct classification PCC is the ratio of the number of correctly detected pixels to the total pixels ratio, and KC is an indicator of the similarity between the detection result image and the real change image on the ground [41, 42]. The above evaluation indexes are calculated as follows: $$$$\begin{equation}OE = FP + FN\end{equation}$$$$ $$$$\begin{equation}PCC = \frac{{{N}_u + {N}_c - FP - FN}}{{{N}_u + {N}_c}} \times 100\% \end{equation}$$$$ $$$$\begin{equation}KC = \frac{{PCC - PRE}}{{1 - PRE}}\end{equation}$$$$ $$$$\begin{equation}PRE = \frac{{\left( {{N}_c - FN + FP} \right) \cdot {N}_c + \left( {{N}_u - FP + FN} \right) \cdot {N}_u}}{{\left( {{N}_c + {N}_u} \right) \cdot \left( {{N}_c + {N}_u} \right)}}\end{equation}$$$$where $${N}_c$$ and $${N}_u$$ are the pixels that changed and did not change in the real surface image, respe...…”

“…In order to quantitatively analyse the results of change detection and evaluate the proposed method, the following evaluation metrics are employed: FP, that is, pixels with no change in the ground detected as the number of pixels with a change, FN, that is, pixels with a change in the ground detected as the number of pixels without a change, the total number of errors OE is the sum of FP and FN, the percentage of correct classification PCC is the ratio of the number of correctly detected pixels to the total pixels ratio, and KC is an indicator of the similarity between the detection result image and the real change image on the ground [41,42]. The above evaluation indexes are calculated as follows:…”