Using deep learning for land cover feature recognition in remote sensing (RS) images is a widely adopted and effective approach. However, most existing methods face challenges in maintaining consistent performance across datasets with diverse characteristics. Consequently, the development of models capable of robust adaptation to heterogeneous datasets has become an essential area of research. Inspired by the growing prominence of attention mechanisms, we propose the multiattention integration (MAI) network, which incorporates three distinct attention mechanisms into the encoder. We design a decoder to adaptively fuse composite features and enhance representation learning. In addition, we introduce a remote sensing dataset, Changfeng–Feidong dataset (CF-FD), and perform experiments on five datasets with varying characteristics for model comparison. The results demonstrate that our model exhibits strong and stable performance across different datasets, highlighting its robustness and adaptability to diverse remote sensing scenarios.

The forest ecosystem possesses the ability to sequester carbon, which is essential for maintaining the balance of the global carbon cycle. Canopy height is a directly related growth parameter of forest that is crucial for accurately estimating forest carbon stocks. Particularly in extensively planted forests, an accurate estimation of canopy height can effectively evaluate their ecological value. However, obtaining a high-resolution canopy height model (CHM) through remote sensing technology remains a challenge, and surveying numerous field values is also difficult. We propose a solution based on unmanned aerial vehicle (UAV) and GaoFen-7 (GF7) stereo satellite data. Through a two-stage upscaling approach, a scale-matched canopy height dataset derived from CHMUAV was utilized to optimize and generate a wide-ranging satellite-scale CHM. The model (LASSO-CHM-Random Forest, LA-CHM-RF) proposed in this study performs well in optimizing the original CHMGF7 (R2=0.73, RMSE = 3.16 m, rRMSE = 25%). The results indicate that the two-stage-upscaling method enhanced the quality of CHMGF7, reducing errors by 19% at the UAV scale and by 33% in large areas. We demonstrate that forest density, rather than forest height, is the primary factor influencing the accuracy of the optimized model.

Accurate estimation of the noise present in hyperspectral images is crucial for the application of various processing algorithms. Works on noise statistics in hyperspectral data have focused on the estimation of noise variances in the spectral bands. These methods implicitly assume that all spatial pixels have the same noise distributions. However, this is not true for pushbroom sensors, which have two-dimensional detector arrays where the electro-optical characteristics of a detector vary from its neighbors in either dimension. We present a method to estimate the noise variances in all the individual detectors. The modified residual method provides a structure for analyzing the noise statistics in hyperspectral images but assumes uniform noise statistics across all spatial pixels. Starting with that methodology, we develop a method to estimate the noise statistics in individual spatial as well as spectral detector pixels, even when the noise is signal dependent. Results on simulated data show that the proposed method estimates the individual noise variances more accurately than both the classic methods and state-of-the-art deep learning methods. The application of the method to noise removal indicates a visible improvement in noisy data.

Hyperspectral image super-resolution aims to reconstruct high-resolution images from their low-resolution counterparts, addressing the inherent limitations of imaging hardware. Recent hyperspectral image super-resolution methods based on convolutional neural networks (CNNs) and transformer have achieved impressive results, but face significant challenges, including CNN’s limited receptive field and transformer’s high computational complexity. To address these issues, we propose the multiple attention-enhanced CNN and transformer network, which combines CNN’s strength in local feature extraction with the transformer’s ability to model global dependencies. The proposed model comprises three main components: the shallow feature embedding module for initial spatial–spectral feature extraction; the multiple attention-enhanced CNN and transformer module for hierarchical feature refinement; and the upsampling module for final image reconstruction. Specifically, the multiple attention-enhanced CNN and transformer module integrates the spectral–spatial attention-enhanced convolutional block to strengthen local spectral–spatial dependencies and the multiscale attention transformer block to capture multiscale global information. Furthermore, residual connections across layers preserve hierarchical features, enabling the model to reconstruct fine spatial and spectral details. Extensive experiments on benchmark hyperspectral datasets demonstrate that the proposed model outperforms state-of-the-art methods in both quantitative metrics and qualitative visual assessments, delivering superior performance in spatial detail restoration and spectral fidelity.

Monitoring small celestial bodies, space debris, and satellites is crucial for space situational awareness. Satellites, with their altitude and angular advantages, serve as ideal platforms for long-range and wide-area monitoring of space targets. However, detecting space objects remains challenging due to weak signals and low signal-to-noise ratios caused by interstellar clutter. Most existing methods rely on target appearance, which can lead to false alarms due to the misidentification of stars. To address this issue, we focus on the intensity temporal profiles (ITPs) of individual pixels. When a target passes through a pixel, it generates a transient disturbance in the ITP, which can be used to detect weak space targets. We extract both deep and statistical features from the ITP and use a decision process to combine these features for effective ITP classification, enabling precise detection. Specifically, we design a network, ConvLSTM-1D, to effectively extract transient disturbance temporal features. These deep features are fused with statistical features to form a feature vector, and AdaBoost is used as a classifier to achieve high performance and reliable detection. We conducted experiments on both synthetic and real datasets, comparing our method with baseline approaches. The experimental results demonstrate that our proposed detection method outperforms others in both qualitative and quantitative assessments, showing significant potential for practical applications.