Abstract:<span>With the rapid development of remote sensing technologies, the frequency of observations of the same location is increasing, and many satellites and sensors produced a large amount of time series images. These images make long-term change detection and dynamic characteristic estimation of ground features possible. However, conventional remote sensing image change detection methods mostly rely on manual visual interpretation and supervised or unsupervised computer-aided classification. Traditional methods always face many bottlenecks when processing big and fast-growing datasets, such as low computational efficiency, low level of automation, and different identification standards and accuracies caused by different operators. With the rapid accumulation of remote sensing data, it has become an important but more challenging task to conduct change detection in a more precise, automated and standardized way. The development of geointelligent computing technologies provides a means of solving these problems and improve the accuracy and efficiency of remote sensing image change detection. In this paper, we presented a novel deep learning model called nest network(NestNet) based on a convolutional neural network to improve the accuracy of the automatic change detection task by using remotely sensed time series images. NestNet extracts the respective features of bi-temporal images using an encoding parallel module and subsequently employs absolute different operations to process the features of two images. Compared with change detection method based on U-Shaped network plus plus (UNet++), the parallel module improves the efficiency of NestNet. Finally, a decoding module is used to generate a predicted change image. This paper compares NestNet to traditional methods and state-of-the-art deep learning models on two datasets. The experimental results demonstrate that the accuracy of NestNet is better than that of state-of-the-art methods. It can be concluded that the NestNet model is a potential approach for change detection using high resolution remote sensing images.</span>

What problem does this paper attempt to address?

The paper primarily addresses the issue of change detection in high-resolution remote sensing images. Specifically, the study proposes a new deep learning model called NestNet, which is a multi-scale change detection method based on convolutional neural networks. The traditional change detection methods mentioned in the paper have the following issues: 1. **Low level of automation**: Traditional methods mostly rely on manual visual interpretation or semi-automatic computer-aided classification, leading to lower computational efficiency and automation levels. 2. **Unstable accuracy**: Due to different operators having different recognition standards and accuracies, traditional methods face bottlenecks when dealing with large-scale and rapidly growing datasets. 3. **Difficulty in multi-scale information fusion**: Existing methods, including both traditional and deep learning methods, struggle to effectively fuse information from different scales. To address the above issues, the research team designed the NestNet model, which has the following features: - **Automatic feature extraction**: NestNet can automatically extract features from images and use encoding modules to process images from two different periods separately. - **Multi-scale information fusion**: By using Absolute Different Operation (ADO) to process the feature differences of images from different time points, and then using decoding modules to fuse these features. - **Parallel modules**: Compared to U-Net++ based change detection methods, NestNet employs parallel modules to process image data from different time periods, improving the model's processing efficiency. The experimental section evaluated the NestNet model using two datasets (OSCD and Google Earth), and the results showed that NestNet outperformed some existing advanced methods in terms of accuracy in change detection tasks.

NestNet: a multiscale convolutional neural network for remote sensing image change detection