Achieving carbon peak and carbon neutrality is a major strategic goal for China. With the gradual expansion of China's carbon market, carbon monitoring plays an increasingly significant role in support services. Semantic segmentation of high-resolution remote sensing images can effectively capture the spatial distribution of carbon sources and sinks and their dynamic changes by accurately dividing land cover types. However, remote sensing images often have complex segmentation boundaries and numerous small-scale targets, resulting in segmentation difficulties. This paper proposes a semantic segmentation network for remote sensing images of carbon sources and sinks, SCSNet, aiming to optimize the segmentation effect of remote sensing images by combining the ability of multi-scale feature modeling and local detail capturing. SCSNet uses a Multi-scale Self-attention and Local Aggregation module in the decoder, which introduces a multi-scale self-attention mechanism to capture global contextual information while enhancing the ability to perceive details through local feature aggregation to distinguish similar targets in remote sensing images effectively. The Feature Fusion Module realizes the effective fusion of features at different scales to further enhance the segmentation accuracy of small-scale targets. Additionally, the Feature Refinement Module is introduced, which further refines the boundary information by fusing high-level semantic information and low-level detail features. To validate the effectiveness and superiority of SCSNet, this paper conducts a large number of experiments on public datasets, including LoveDA, Vaihingen, and Potsdam. The experimental results show that SCSNet achieves advanced segmentation performance in remote sensing image segmentation tasks, accurately identifying and differentiating carbon sources and sinks, thus providing powerful support for carbon monitoring and carbon emission assessment. Our source code is available at https://github.com/SCS123-LAB/SCSNet.

High resolution remote sensing image semantic change detection (SCD) helps to accurately capture the spatial distribution and dynamic evolution of carbon sources and sinks by identifying changes in land cover types. However, existing methods suffer from the loss of spatial details and insufficient ability to model global features. Therefore, this letter proposes an SCD model based on spatiotemporal attention perception and multi-scale fusion (SC-SCDNet). The model introduces a multi-scale efficient cross attention block (MCA) in the encoder to bridge the semantic gap, and integrates a feature enhancement module (FEM) to enhance the semantic expression ability of small targets using multi-branch dilated convolution. In addition, a spatiotemporal channel window interaction module (TBCM) is designed to capture global information from both spatial and channel dimensions, enhancing spatial detail expression. The experimental results show that SC-SCDNet achieves the most advanced performance on SECOND and Landsat-SCD datasets, providing a better technical scheme for carbon sources and carbon sinks change detection.

Convolutional neural networks (CNN) have been widely used in the field of remote sensing (RS) scene classification, which have achieved remarkable results. In RS scene classification, local key objects are particularly crucial for classification results. However, most existing CNN methods directly utilize the deep-level global features of CNN, ignoring object-level information in shallow features or leading to redundant and erroneous information when using shallow features. To fully utilize the important information in shallow features, we proposed an end-to-end contextual spatial-channel attention network (CSCANet) to learn multi-layer feature representations and further improve classification performance by employing shallow object-level semantic information. Firstly, ResNet34 is pre-trained to extract different levels of features. Secondly, a contextual spatial-channel attention module is constructed to generate contextual spatial-channel attention features by exploiting features at different levels. Finally, the triple loss function is combined with the central loss function to guide the model training. Experiments on three public RS scene classification datasets (UC-Merced, AID, and NWPU-RESISC45) demonstrate that the proposed method achieves highly competitive results.

Automatic segmentation of the left ventricle (LV) in four-chamber view images is critical for computer-aided cardiac disease diagnosis. The complex structure of the cardiac image and the encoder-decoder networks may cause coarse segmentation results. High-accuracy LV segmentation is still a challenge with existing automatic LV segmentation methods . In this paper, we propose a slide deep reinforcement learning segmentation network for pixelwise LV segmentation. The main architecture of the slide reinforcement learning networks consists of a slider item combined state, a group of morphology transforming actions and an agent network. The specifically designed reinforcement learning state comprises an image item and a slider item, which contains both original image information and network act information. The reinforcement learning actions proposed in this paper enable accurate and fast formulation of the binary segment result for each frame by controlling the length and location of the slider. Additionally, the confidence branch proposed in our experiment provides a continuous frame series environment, and the identification algorithm avoids losing the segmentation target. The segmentation result reveals that the proposed method outperforms FCN, SegNet, U-Net and TransUnet. The IoU improved by 23.01 % , 15.4 % , 11.24 % and 6.9 % . Additionally, we demonstrate how the proposed method can be used as a semisupervised method, which is more convenient for the image annotation process.

Convolutional neural network (CNN) has a complex model structure in hyperspectral image (HSI) classification and the energy consumption during training and inference is high, so it cannot be applied in edge computing devices such as software-defined satellites and unmanned aerial vehicles. In order to solve the classification of HSI in an edge computing environment, inspired by the principle of neuro-dynamics and brain-inspired computing, we use integrate and fire neurons and shuffle squeeze and excitation (SE) module network to construct a spiking neural network (SNN-SSEM). This letter designs an approximate derivative backpropagation algorithm for discontinuous activation function and realizes the training of an SNN. Experiments were conducted on three HSI datasets and the average classification accuracy reached more than 99%. The energy consumption of our model is about 4.5 times that of CNN with the same architecture. This study is an exploration of the application of the scientific theory of brain-inspired computing in hyperspectral remote sensing technology, which can realize real-time classification of HSI in the mobile computing environment.

Recently, deep learning methods have made significant progress in solving hyperspectral images (HSIs) classification problems of high-dimensional features, band redundancy, and spectral mixture. However, the deep neural network is too complex, with a long training time and high energy consumption, making it difficult to deploy on edge computing devices. In order to solve the above problems, this paper proposes a brain-inspired computing framework based on the spiking leaky integrate-and-fire neuron model for HSIs classification. Then we design an approximate derivative algorithm to solve the non-differentiable spike activity of the spiking neuron. The framework uses direct coding to generate spatiotemporal spikes for input HSI and achieves efficient extraction of spatial-spectral features through spiking standard convolution and spiking depthwise separable convolution. Extensive experiments are performed on four benchmark hyperspectral data sets and two public unmanned aerial vehicle-borne hyperspectral data sets. Experiments show that the proposed model has the advantages of high classification accuracy and fewer spiking time steps. The proposed model can save about 10 times computational energy consumption compared with the CNN of the same architecture. This research has great significance for overcoming the technical bottleneck of HSI classification based on brain-inspired computing, solving the critical problems of mobile computing in unmanned autonomous systems, and realizing the engineering application of unmanned aerial vehicles and software-defined satellites. The source code will be made available at https://github.com/Katherine-Cao/HSI_SNN.

Target detection is a-priori conditions for target tracking, classification, recognition, and scene understanding in Remote Sensing Image (RSI) analysis. However, the many traditional algorithms for target detection cannot perform well when the image resolution, especially for high-resolution RSIs, is change. Therefore, in this paper, we introduce a novel target detection algorithm based on the visual saliency of Spiking Neural Networks (SNN), which can efficiently detect the discriminative information from high-resolution RSIs to find targets by a saliency computing. As a result of this, it can provide an efficient and fast calculation method. The proposed visual saliency algorithm was applied to extensive experiments to detect the ship, and experimental results showed the outstanding performance for target detection on the optical RSI and synthetic aperture image.