Land-Use/Land-Cover change detection based on a Siamese global learning framework for high spatial resolution remote sensing imagery

ForkNet: Strong Semantic Feature Representation and Subregion Supervision for Accurate Remote Sensing Change Detection

“…It is also known as mid-fusion methods in other literature [17]. FDCNN [21] and Siam-GL [22] belong to this category. The main difference between FDCNN and Siam-GL is the merging of image feature representations, with the former generating difference feature representations and the latter generating joint feature representations.…”

“…First, change is semantic information, and weak semantic image feature representations can impair change detection. In these methods, the image feature extraction stage (involving DFEN only) generally computes the feature representation layer by layer and involves subsampling layers, which results in a hierarchical multiscale image feature representation [6], [21], [22]. There is a large semantic gap between low-level feature maps and high-level feature maps in such image feature representations due to the difference in depth [12], [25].…”

“…1. Illustration of the difference between other siamese change detection networks (a) [6], [22] and our ForkNet (b). Other siamese change detection networks generally only have a downsample DFEN, while, in ForkNet, the feature extraction consists of three parts: DFEN, CRAM, and FPN.…”

Hyperspectral Image Classification of Brain-Inspired Spiking Neural Network Based on Approximate Derivative Algorithm

“…A S an essential part of earth observing technology, remote sensing technology has been widely used in forest management [1], soil salinity estimation [2], mineral exploration [3], anomaly detection [4], military reconnaissance [5], urban planning [6], road extraction [7], and land-user change detection [8]. In recent years, with the development of remote sensing technology and hyperspectral imaging technology, hyperspectral images (HSIs) contain richer spectral information and higher spatial resolution.…”

Change Detection in VHR Imagery With Severe Co-Registration Errors Using Deep Learning: A Comparative Study

“…The authors also proposed a combination of the weighted binary cross-entropy loss (WBCE) and the dice coefficient loss to improve the training of imbalanced samples. Finally, in [50], focus was put on semantic CD and a Siamese framework with a global hierarchical (G-H) sampling mechanism was trained on three datasets with semantic annotated changes [51,52]. The purpose of the G-H sampling mechanism is the mitigation of the imbalance problem.…”