Detection of Changes in Low Frequency SAR Images Using Two CNNs

In October 2020, I and three other authors published an article at the IEEE Geoscience and Remote Sensing Letters Journal presenting an algorithm for the detection of changes in SAR images captured by a special class of SAR systems---the so-called wavelength-resolution SAR systems. These systems produce images at the low-VHF band with resolutions close to the wavelength, which is in the order of one meter.

SAR systems working around these frequencies are useful to detect large objects concealed under dense forests, since they can penetrate small objects, such as leaves and branches from trees. With this in mind, we used the CARABAS-II dataset to train and evaluate the proposed algorithm. This dataset is composed of 24 wavelength-resolution SAR images captured at northern Sweden in 2002. Each image has 25 military vehicles concealed under foliage.

Difference images, which are consisted of the subtraction of two CARABAS-II images of the same scene captured at different instants of time, were analyzed by the algorithm. The performed steps are shown in the diagram above. In short, a difference image is scanned by a low complexity FCNN which semantically segments the image into pixels that may be parts of changes---positive pixels---and pixels that likely are not parts of changes---negative pixels, creating a binary segmentation map. Following, the DBSCAN clustering algorithm is used to find clusters of positive pixels in the segmentation map. Then, patches centralized at the central points of the found clusters are extracted from the original difference image to be further analyzed by a classification CNN. This network, which has around 30 times more parameters than the segmentation network, classifies each patch as containing or not containing a change.

The main goal of using two networks for this problem instead of using a single segmentation network, for instance, is to keep the receptive field of the algorithm as low as possible. Since the CARABAS-II dataset has very few images, when compared to datasets composed of optical images, having a small receptive field is important to not overfit the change detection algorithm to the spatial distribution of the vehicles contained in these images.

The proposed algorithm achieved excellent performance in the executed tests, beating every other comparable algorithm at the time of publication. The Receiver Operating Characteristic (ROC) curve of the algorithm is presented in the plot below and compared with other systems. The articles referenced in the plot, as well as other information about training, testing, cross-validation, etc. can be found in the article CNN-Based Change Detection Algorithm for Wavelength-Resolution SAR Images.?

The algorithm has been developed in Python using multiple known libraries, such as Keras/Tensorflow, Scikit-Learn, NumPy, etc. The Python files and the Jupyter notebook where the algorithm is executed are available at GitHub. Feel free to check it out!

OBS: Embraer R-99 Airplane image obtained at https://www.flickr.com/photos/24874528@N04/2370466415/ . It uses a CC BY-SA 2.0 license, which is available at https://creativecommons.org/licenses/by-sa/2.0/. No modifications were made to the image.