Unsupervised Image Regression for Heterogeneous Change Detection

Abstract

Change detection in heterogeneous multitemporal satellite images is an emerging and challenging topic in remote sensing. In particular, one of the main challenges is to tackle the problem in an unsupervised manner. In this paper we propose an unsupervised framework for bitemporal heterogeneous change detection based on the comparison of affinity matrices and image regression. First, our method quantifies the similarity of affinity matrices computed from co-located image patches in the two images. This is done to automatically identify pixels that are likely to be unchanged. With the identified pixels as pseudo-training data, we learn a transformation to map the first image to the domain of the other image, and vice versa. Four regression methods are selected to carry out the transformation: Gaussian process regression, support vector regression, random forest regression, and a recently proposed kernel regression method called homogeneous pixel transformation. To evaluate the potentials and limitations of our framework, and also the benefits and disadvantages of each regression method, we perform experiments on two real data sets. The results indicate that the comparison of the affinity matrices can already be considered a change detection method by itself. However, image regression is shown to improve the results obtained by the previous step alone and produces accurate change detection maps despite of the heterogeneity of the multitemporal input data. Notably, the random forest regression approach excels by achieving similar accuracy as the other methods, but with a significantly lower computational cost and with fast and robust tuning of hyperparameters.

Heterogeneous Change Detection on Remote Sensing Data with Self-Supervised Deep Canonically Correlated Autoencoders

“…We want the network to learn to individuate changes as abnormal patterns, and not as a rule. Thus, an innovative technique is used to automatically retrieve some training samples located in likely unchanged areas from our data [Luppino et al, 2019], turning our procedure to a completely unsupervised method. This stage is conceived as a method to extract information and to return a probability-like score that expresses the chance that each pixel is changed from one acquisition to the other; this is explained in [Luppino et al, 2020].…”

“…Most current HCD approaches adopt transformations between the input domains, or from these to a common latent domain, to bring data to a space where they can be efficiently compared. Convolutional neural network (CNN) architectures such as autoencoders and generative adversarial networks are flexible and powerful tools that can accomplish these image translation tasks, as reviewed in [Luppino et al, 2019[Luppino et al, , 2020.…”

“…However, this highly efficient algorithm reduces the computational complexity from quadratic to linear in the number of pixels by approximating the random field with a mean-field whose iterative update can be computed using Gaussian filtering in the feature space. The number of iterations and the kernel width of the Gaussian kernels are the only hyper-parameters manually set, and we opted to tune them according to [Luppino et al, 2019]: 5 iterations and a kernel width of 0.1.…”

“…All these SAR channels were log-transformed. The ground truth in Figure 5.1c has been provided by [Luppino et al, 2019] and was manually annotated. Originally these images were of 3500 × 2000 pixels, but they have been resampled to 850 × 500 pixels because of computational time constraints.…”

Snow Avalanche Segmentation in SAR Images With Fully Convolutional Neural Networks

“…Convolutional neural networks (CNNs) have attracted considerable interest for their ability to model complex contextual information in images [3]. Prominent examples in remote sensing are terrain surface classification [4], [5], categorization of aerial scenes [6], detection of changes in the terrain over time from SAR and optical satellite sensors [7], [8], and segmentation of objects from airborne images [9], [10]. Nevertheless, few research efforts have been devoted so far towards detecting avalanche activity from SAR data, which remains an open and challenging endeavour.…”