Nowadays, due to the difficult acquisition of true labels, a semisupervised neural network has shown great potential for change detection (CD) in remote sensing images. However, most of the traditional semisupervised neural network detection frameworks are complex to train and require additional structural analysis, along with a fixed structure, lacking universality. In this article, a semisupervised adaptive ladder network (SSALN) for remote sensing image CD is proposed, which enables dual-input label-incremental architecture searching with a concise and variable structure. First, SSALN is suitable for CD from two remote sensing images of any type with the characteristic of minimal label dependency and automatic network structure adjustment. The network can generate more reliable pseudolabels through continuous iterations to help limited real labels exploit implicit information, identify the most effective network, and form the ascending network structure optimization. Second, the acquisition of pseudolabels is the fusion of semisupervised and unsupervised CD approaches, which ensures the multiperspective information supplement. Multiple CD maps are fused to generate labels for the next iteration, making the predicting more reliable. Finally, both homogenous images and heterogenous images are tested with experiments. Even if the detection object is switched, it can be well adaptive and compatible without manual modification of the network. Experimental results demonstrate that the proposed method can promote the flow of label information through structure searching and self-circulation in the ascending network optimization; thus, it has outstanding performance on tasks of remote sensing image CD.

In this paper, a novel change detection technique is proposed based on multiscale superpixel segmentation and stacked denoising autoencoders (SDAE). This approach is designed to achieve superpixel-based change detection, in which the basic analysis unit is between pixel-based and object-based ones. Given two original images, the difference image (DI) is obtained by conventional DI generation methods. Then, we propose a multiscale superpixel segmentation which is guided by the changing degrees estimated from DI. Different from traditional multiscale superpixel, the proposed multiscale superpixel segmentation is employed in a single map. In the proposed method, SDAE is used to learn the difference representation between bi-temporal superpixels. Bi-temporal superpixels are stacked and fed into SDAE for its pre-training, and then SDAE is fine-tuned according to pseudo labels generated by traditional unsupervised methods. After fine-tuned with back propagation, SDAE can be used to classify all superpixel pairs into changed or unchanged ones. The experimental results on real remote sensing datasets have demonstrated the effectiveness of the proposed approach.

Deep neural networks have evolved significantly in the past decades and are now able to achieve better progression of sensor data. Nonetheless, most of the deep models verify the ruling maxim in deep learning—bigger is better—so they have very complex structures. As the models become more complex, the computational complexity and resource consumption of these deep models are increasing significantly, making them difficult to perform on resource-limited platforms, such as sensor platforms. In this paper, we observe that different layers often have different pruning requirements, and propose a differential evolutionary layer-wise weight pruning method. Firstly, the pruning sensitivity of each layer is analyzed, and then the network is compressed by iterating the weight pruning process. Unlike some other methods that deal with pruning ratio by greedy ways or statistical analysis, we establish an optimization model to find the optimal pruning sensitivity set for each layer. Differential evolution is an effective method based on population optimization which can be used to address this task. Furthermore, we adopt a strategy to recovery some of the removed connections to increase the capacity of the pruned model during the fine-tuning phase. The effectiveness of our method has been demonstrated in experimental studies. Our method compresses the number of weight parameters in LeNet-300-100, LeNet-5, AlexNet and VGG16 by $24\times$, $14\times$, $29\times$ and $12\times$, respectively.

To address the problem of enormous differences in two heterogeneous images, the traditional unsupervised frameworks are most normally realized by converting two images into a common domain with various auxiliary strategies, such as transformation and alignment, which requires extensive calculation and has difficulty balancing the training tasks. To achieve a concise framework, this paper proposes self-guided autoencoders (SGAE) for unsupervised change detection in heterogeneous remote sensing images. Unlike traditional methods that aim to narrow the differences of heterogeneous images to highlight the changed information, SGAE forces the flow of identification in formation generated in unlabeled data through self-guided iterations. First, initial unsupervised networks output an elementary change map that will be screened to obtain reliable pseudolabels. The selected pseudo-labeled samples will be used as the input of a supervised network to obtain another change map. Then, multiple change maps will be fused to refine the confidence of pseudolabels again, obtaining new fused pseudolabeled samples for the self-guided network, which will be trained with pseudo-labeled samples and unlabeled samples. Finally, the above operations will be repeated to continuously optimize the net, which helps itself to extract the discriminative features for classification in self-guided iterations. Experiments compared with several algorithms on four datasets demonstrate the effectiveness and robustness of our method, which can help unsupervised models improve discriminative feature extraction and classification performance with a more flexible learning method.