Yingjun Zhao

Monocular depth estimation plays a crucial role in applications such as autonomous driving and mobile 3D reconstruction. However, existing lightweight methods are often constrained by limited computational resources and rely on shallow feature representations for direct depth regression. As a result, cross-scale residual information is insufficiently modeled, which limits their ability to preserve structural consistency and recover fine-grained details in complex scenes. To address these challenges, we propose LapCR-Net, a lightweight monocular depth estimation network based on Laplacian residual reconstruction. Specifically, we formulate a progressive Laplacian residual framework that decomposes depth prediction into a coarse-to-fine multi-scale refinement process. To enhance feature representation in the decoder, we introduce a Structure-aware Feature Recalibration (SFR) module and a Depth-guided Convolution Module (DCM), which strengthen spatial semantic correlations and improve residual prediction across scales. Furthermore, we design an uncertainty-driven collaborative refinement strategy to adaptively adjust residual correction strength. By estimating prediction uncertainty, the proposed strategy sharpens object boundaries while suppressing texture artifacts. Extensive experiments on the NYU-Depth V2 and KITTI benchmarks demonstrate that LapCR-Net achieves competitive performance with only 5.4 M parameters. In particular, it shows clear advantages in structural preservation and detail reconstruction, achieving a favorable trade-off between accuracy and computational efficiency.

Automatic detection of electric pylons in optical remote sensing imagery is important for large-scale powerline monitoring, but remains challenging due to complex backgrounds, small target appearances, and large variations in pylon-shadow structures. This paper proposes a Refined Deformable-DETR framework with a Spatial Context-aware Query Modulation (SCQM) module to enhance object query representations. SCQM aggregates image-level contextual information from encoder memory and generates channel-wise modulation vectors to recalibrate object queries before deformable cross-attention, thereby providing image-conditioned channel priors for subsequent query-feature interaction. Experiments on the self-constructed Electric Pylon Remote Sensing Dataset (EPRD) show that the proposed method improves AP from 72.7% to 74.1% and APs from 47.2% to 50.9% compared with the baseline Deformable-DETR. Evaluations on the public Electric Pylon Dataset (EPD) further demonstrate its generalization capability. These results indicate that context-aware query modulation is effective for Transformer-based electric pylon detection in complex remote sensing scenarios.

To address misclassification and reduced accuracy in semantic change detection caused by class imbalance and variable object sizes, this paper improves BGSNet and proposes a new change detection network, SRDFNet (Semantic Refinement and Differential Features). Based on BGSNet’s framework, it introduces three complementary modules: (1) a hierarchical graph module (HGM) that converts multi-scale feature maps into compact semantic graph nodes, using graph attention for intra-layer and cross-level semantic interaction to enhance topological relationship perception; the HGM mitigates the effects of class imbalance by compacting multi-scale features into semantic nodes; (2) a difference enhancement (DE) module that extracts multi-receptive-field difference information from bi-temporal concatenated features via multi-scale parallel convolution branches; (3) a semantic refine (SR) module that performs lightweight residual refinement on bi-temporal semantic features to improve the segmentation accuracy. The DE and SR modules mitigate the degradation in semantic segmentation accuracy caused by variable object sizes. It is trained and tested with BGSNet and three other models on the SECOND and HRSCD datasets. For the SECOND dataset, in terms of five quantitative indicators, namely OA, mIoU, SeK, F1 and recall, SRDFNet achieves 87.64%, 70.31%, 20.36%, 60.25% and 65.27%, respectively. Compared with BGSNet, it gains performance increases of 1.34%, 0.73%, 1.44%, 0.81% and 2.72%, respectively. For the HRSCD dataset, SRDFNet achieves 98.13% (OA), 52.67% (mIoU), 73.77% (SeK), 88.86% (F1) and 88.18% (recall), ranking first among the four methods. Compared with BGSNet, it gains performance increases of 3.96%, 3.93%, 9.69%, 2.33% and 4.00%, respectively.

Hyperspectral unmixing is an important technique for mineral mapping because natural geological scenes commonly contain mixed pixels composed of multiple spectrally overlapping materials. In mineral environments, these mixtures are often intimate rather than purely areal, and nonlinear scattering effects may weaken the validity of linear mixing assumptions. Although autoencoder-based hyperspectral unmixing methods can jointly estimate endmembers and abundances in an unsupervised manner, they often suffer from insufficient physical constraints, unstable endmember learning, and limited geological interpretability. To address these issues, this study proposes a physically constrained and library-guided convolutional autoencoder for mineral hyperspectral unmixing. The method retains an interpretable linear reconstruction backbone while introducing a Hapke-consistency regularization term to incorporate physically motivated nonlinear scattering behavior during endmember optimization. In addition, a library-aware endmember anchor module is designed to improve initialization quality, reduce endmember drift, and guide optimization toward spectrally meaningful solutions. The proposed method was evaluated on both simulated hyperspectral datasets and real airborne SASI data. On the simulated datasets, the method achieved improved endmember spectral fidelity and lower abundance estimation error than several representative autoencoder-based baselines, with the advantage being more evident under nonlinear mixing conditions. Ablation experiments further showed that the Hapke-consistency term mainly improved physical plausibility, whereas the anchor module enhanced optimization stability and spectral consistency. On the real airborne dataset, the proposed method produced endmember spectra that were more consistent with field and laboratory mineral references and generated spatially more coherent abundance maps. These results indicate that incorporating physically motivated constraints and mineral-library priors into deep autoencoder frameworks can improve the robustness and interpretability of mineral hyperspectral unmixing. The proposed framework provides a practical direction for hyperspectral mineral mapping in mixed and spectrally complex geological environments.

Retrieving non-optically active water quality variables, such as total hardness (TH) and permanganate index (CODMn), from hyperspectral data remains challenging because these parameters are not directly linked to spectral reflectance. To improve their estimation from UAV hyperspectral imagery, a GA-MHSA-BPNN framework was developed by combining a genetic algorithm (GA), multi-head self-attention (MHSA), and a backpropagation neural network (BPNN). In this framework, MHSA was introduced to strengthen the representation of informative spectral features, while GA was applied to optimize the initial network parameters and thus enhance convergence stability. The proposed framework was evaluated against BPNN, GA-BPNN, MHSA-BPNN, and 1D-CNN models. Among the tested approaches, GA-MHSA-BPNN produced the most favorable results for both TH and CODMn, with R² values of 0.878 and 0.843, respectively. Additional experiments using different proportions of training samples showed that the model remained relatively stable when the training data were reduced to 70% and 50% of the original dataset. These results indicate that integrating GA and MHSA into a UAV hyperspectral retrieval framework can improve the estimation of non-optically active water quality variables and provide useful methodological support for efficient and refined monitoring of drinking water source areas.

Traditional diagnostic spectroscopy provides a physically interpretable basis for mineral identification. However, how modern classifiers balance spectral and spatial information remains insufficiently understood. This study investigates this issue using CASI airborne hyperspectral data from the Liuyuan area, China. A geologically constrained ground-truth dataset was constructed based on expert knowledge and a semi-automatic Spectral Hourglass workflow. We evaluated representative shallow machine learning methods and deep learning models, including a three-dimensional convolutional neural network (3D-CNN), Vision Transformer (ViT), and SpectralFormer. The Support Vector Machine (SVM) achieved the highest overall accuracy but showed a strong bias toward dominant background classes and failed to reliably detect rare minerals such as jarosite. Deep learning models improved class balance by incorporating broader spectral features. However, excessive spatial aggregation reduced their sensitivity to small and fragmented alteration zones. SpectralFormer models hyperspectral data as ordered spectral sequences and showed more stable performance for spectrally similar and rare minerals. Multi-scale experiments reveal a spectral-dominant discrimination mechanism. Increasing the spectral receptive field improves classification up to an optimal level. In contrast, overly large spatial patches introduce background interference and obscure diagnostic absorption features. These findings highlight the fundamental role of spectral continuity in airborne hyperspectral alteration mineral mapping and clarify the trade-offs involved in integrating spatial context.

Electric pylons are crucial components of power infrastructure, requiring accurate detection and identification for effective monitoring of transmission lines. This paper proposes an innovative model, the EP-YOLOv8 network, which incorporates new modules: the DSLSK-SPPF and EMS-Head. The DSLSK-SPPF module is designed to capture the surrounding features of electric pylons more effectively, enhancing the model’s adaptability to the complex shapes of these structures. The EMS-Head module enhances the model’s ability to capture fine details of electric pylons while maintaining a lightweight design. The EP-YOLOv8 network optimizes traditional YOLOv8n parameters, demonstrating a significant improvement in electric pylon detection accuracy with an average mAP@0.5 value of 95.5%. The effective detection of electric pylons by the EP-YOLOv8 demonstrates its ability to overcome the inefficiencies inherent in existing optical satellite image-based models, particularly those related to the unique characteristics of electric pylons. This improvement will significantly aid in monitoring the operational status and layout of power infrastructure, providing crucial insights for infrastructure management and maintenance.

The alterations in vegetation cover in the Three North Protection Forest Project area influence its ecological and environmental management. It helps to study soil erosion, environmental change, and ecosystem protection to provide basic data support. Based on the Google Earth Engine cloud platform, this paper selects MODIS 3Q1 data from 2000–2020 and uses the image element dichotomous model to estimate the fractional vegetation cover (FVC) of the Three North Protection Forest Project area, evaluates the stability and temporal and spatial variation of FVC, investigates the coupling relationship between the FVC and temperature and rainfall through evaluation indexes such as the slope of inter-annual variation and partial correlation coefficient, and then analyzes the influence of land use changes on the FVC. The results show that the fractional vegetation cover of the Three North Protection Forest Project area as a whole has increased significantly over the past 20 years, the medium and high vegetation coverage areas have reached 36.4%, the high vegetation FVC has increased from 20.66% in 2000 to 21.59% in 2020, and the vegetation activity is increasing. The vegetation cover is significantly affected by the topographic effect, with the vegetation cover improving relatively well at slopes between 5–13° and elevations ranging from 2000–2500 m. The overall positive correlation between vegetation and temperature and vegetation and precipitation in the study area was 54.08% and 70.04%, respectively, and precipitation was the key factor influencing vegetation growth in the research region. Human activities have a stronger effect on vegetation construction than destruction, and this study contributes to the evaluation of the benefits of the Three North Protection Forest Project and the understanding of human influence on environmental change.

Monitoring the lake waterbody area in the Qinghai–Tibet Plateau (QTP) is significant in dealing with global climate change. The latest released Landsat-9 data, which has higher radiation resolution and can be complemented with other Landsat data to improve imaging temporal resolution, have great potential for applications in lake area extraction. However, no study is published on identifying waterbodies and lakes in large-scale plateau scenes based on Landsat-9 data. Therefore, we relied on the Google Earth Engine (GEE) platform and selected ten waterbody extraction algorithms to evaluate the quantitative evaluation of waterbody and lake area extraction results on the QTP and explore the usability of Landsat-9 images in the relationship between the extraction accuracy and the algorithm. The results show that the random forest (RF) algorithm performs best in all models. The overall accuracy of waterbody extraction is 95.84%, and the average lake waterbody area extraction error is 1.505%. Among the traditional threshold segmentation waterbody extraction algorithms, the overall accuracy of the NDWI waterbody extraction method is 89.89%, and the average error of lake waterbody area extraction is 3.501%, which is the highest performance model in this kind of algorithm. The linear regression coefficients of NDVI and reflectance of Landsat-8 and Landsat-9 data are close to 1, and R² is more significant than 0.91. At the same time, the overall accuracy difference of water extraction between the two data is not better than 1.1%. This study proves that Landsat-9 and Landsat-8 data have great consistency, which can be used for collaborative analysis to identify plateau waterbodies more efficiently. With the development of cloud computing technologies, such as Gee, more complex models, such as RF, can be selected to improve the extraction accuracy of the waterbody and lake area in large-scale research.

The monitoring of lake waterbody area in the Qinghai Tibet Plateau (QTP) is of great significance to deal with global climate change. As the latest generation of Landsat series satellites, Landsat-9 data not only have higher radiometric resolution, but also cooperate with other Landsat satellites to greatly improve the temporal resolution. It has great application potential in lake waterbody area monitoring. In order to explore the performance of different algorithms for extracting waterbody and lake waterbody area in Landsat-9 data under large-scale QTP regions, this study relies on Google Earth Engine (GEE) platform and selects 10 waterbody extraction algorithms as the basis to realize the quantitative evaluation of QTP lake waterbody area extraction results. The results show that the Random Forest (RF) algorithm performs best in all models. The overall accuracy of waterbody extraction is 95.84%, and the average error of lake waterbody area extraction is 1.505%. Among the traditional threshold segmentation waterbody extraction algorithms, the overall accuracy of the NDWI waterbody extraction method is 89.89%, and the average error of lake waterbody area extraction is 3.501%, which is the highest performance model in this kind of algorithms. This study proves that Landsat-9 data can effectively classify QTP waterbodies. With the development of cloud computing technologies such as Gee, more complex models such as RF can be selected to improve the extraction accuracy of water body and Lake area in large-scale research.