【Objective】This study aims to improve the classification accuracy of dominant tree species in typical natural secondary forests in northeast China, a convolutional neural network (CNN)-based framework for tree species classification using UAV hyperspectral images was proposed.【Method】Four dominant tree species—Fraxinus mandshurica, Juglans mandshurica, Ulmus sp., and Betula platyphylla—from the Maoershan Experimental Forest Farm of Northeast Forestry University were studied. Hyperspectral images of seven different regions were acquired using a novel UAV-mounted hyperspectral imager. A single-tree dataset with varying crown sizes was constructed using ground-measured data, divided into training and test sets at a 7∶3 ratio. A hybrid 3D-2D-CNN model integrating 3D and 2D convolutional layers was developed: 3D convolutional layers extracted spectral-spatial coupled features, while 2D layers captured detailed spatial features, enhancing the model’s holistic learning capability. The model was compared with 2D-CNN, 3D-CNN, and feature-selection-based machine learning models (random forest (RF), support vector machine (SVM), and gradient boosting machine (GBM)). Additionally, the band importance was analyzed using a progressive band removal method, and spectral feature sensitivity was investigated.【Result】The proposed 3D-2D-CNN model achieved a classification accuracy of 87% and an F1 score of 0.86 for the four tree species, outperforming other algorithms with an overall accuracy improvement of 5%-6%. Band importance analysis highlighted the significant contribution of the near-infrared band classification.【Conclusion】The 3D-2D-CNN model, by effectively integrating spectral and spatial information, significantly enhanced the classification performance of natural secondary forest tree species compared to traditional methods. This approach provides technical support for forest resource management and ecosystem monitoring via remote sensing.
【Objective】This research aims to accurately and efficiently detect mildly discolored pine trees infected with pine wilt disease in real time using multispectral unmanned aerial vehicle (UAV) remote sensing images, a Multispectral Images YOLO (MI-YOLO) model is proposed based on YOLOv8n.【Method】First, multispectral images are rapidly aligned using the cross power spectrum in the frequency domain. Second, a tiny multi-branch auxiliary feature pyramid network is introduced as the neck network to enhance feature utilization while maintaining model lightweight. Finally, the original C2f feature fusion module in YOLOv8n is replaced with a lightweight C2f-Faster module to reduce redundant computation.【Result】The proposed MI-YOLO model achieves an average precision of 84.5% at an IoU threshold of 0.5 (AP50), with 2.1 MB parameters and 7.25 GB floating point operations (FLOPs). Compared with YOLOv8n, AP50 is improved by 10.5 percentage points, while the number of parameters and FLOPs are reduced by 30% and 13%, respectively.【Conclusion】The MI-YOLO object detection model exhibits high accuracy and a lightweight structure, enabling real-time detection of mildly discolored pine trees infected with pine wilt disease in multispectral images.
【Objective】This study aims to reveal the dynamic evolution patterns of reed wetland land types in Zhalong Wetland and quantify the spatiotemporal variations of their coverage characteristics, providing scientific basis for regional wetland ecological conservation and sustainable development.【Method】Based on Sentinel-2 multispectral remote sensing imagery from 2017 to September 2023 in Zhalong Wetland, a multi-temporal remote sensing dataset was constructed, containing five land categories: lake, reed bed, construction land, arable land, and saline-alkali land. An integrated classification approach combining the attention-based Vmamba model with NDWI water mask was proposed to extract spatial distributions and area changes of each land type. Meanwhile, the fractional vegetation cover (FVC) was inverted using the dimidiate pixel model, and leaf area index (LAI) and ecosystem quality index (EQI) were calculated.【Result】Classification results demonstrated that the overall accuracy (OA) of our algorithm was 80.85%, the mean intersection over union (MIoU) was 71.59%, and macro-F1 score (MF1) was 79.93%. During the study period, lake and reed bed areas in Zhalong Wetland showed expanding trends, while cultivated land and built-up land areas continuously decreased. Saline-alkali land area fluctuated dynamically. Both FVC and EQI exhibited a first-increasing-then-decreasing trend, which was generally consistent with the content of Chinese Climate Bulletin.【Conclusion】The proposed change-monitoring model integrating attention-based Vmamba and water mask demonstrates high applicability in wetland information extraction, significantly improving classification accuracy and dynamic monitoring precision. The collaborative monitoring results of FVC, LAI, and EQI provides reference for wetland resource management and sustainable utilization.
【Objective】Pinus koraiensis needles exhibit a significant forest fire risk due to their high oil content, and surface fire spread is the main fire spread mode. Developing a predictive model for surface fire spread rates can provide scientific basis and valuable insights for fire prevention and control in Pinus koraiensis plantations.【Method】The dead coniferous layer of Pinus koraiensis plantation in Liangshui area of Heilongjiang province was used as the material, 360 sets of indoor point burning tests were conducted with water content of 0, 5%, 10%, 15%, 20%, slope of 0°, 5°, 10°, 15° and wind speed of 0, 1, 2, 3, 4 and 5 m/s. Based on the fire spread rate measured by thermocouple method, an improved WOA(IWOA)-BP neural network model was constructed to predict the fire spread rate, and the prediction results were compared with those of three models (WOA-BP neural network, GA-BP neural network and PSO-BP neural network).【Result】Slope, wind speed and fire spread rate were significantly positively correlated (P<0.01), while water content exhibited a negative correlation with fire spread rate (P<0.05). The fire spread rate decreased with an increase in fuel water content, and increased with the increase of wind speed and slope. When the wind speed was 4 m/s, the fire spread growth rate reached the maximum. The improved whale optimization algorithm (IWOA) included Tent chaotic mapping, improved nonlinear convergence factor, adaptive weighting and Levy flight motion. These enhancements increased the algorithm’s randomness and diversity, thereby improving its convergence speed and reducing the likelihood of becoming trapped in local optima, with high prediction accuracy and robustness. The accuracy and stability of the BP neural network model optimized by the IWOA demonstrated significant improvements compared to three other models, exhibiting the highest model fitness to the measured data.【Conclusion】The IWOA-BP neural network model can effectively predict the fire spread rate of the dead coniferous layer of the Pinus koraiensis plantation, and providing scientific guidance for forest fire prevention and control and forest litter fire spread rate prediction model.
【Objective】This study developed a processing method based on terrain scale conversion and image reconstruction to address the challenges of blurred kilometer-scale satellite imagery and the absence of micro-topography information in forest fire monitoring. Aiming to enhance the clarity and visualization effect of the fire environments in satellite images, thereby supporting rapid-fire information extraction and informed decision-making for fire prevention and suppression.【Method】The study focused on two forest fire events: the “5·17” fire in Ganhe Village, Shuanghe Yi Township, Jinning District, Kunming City, and the “5·18” fire in Yutaizi Mountain, Yongqing Village, Xiaobaihu Town, Luliang County, Qujing City, both occurring in 2019. Utilizing FY-3B satellite imagery and 30 m resolution Digital Elevation Model (DEM) data, terrain units at varying scales were constructed using the point spread function (PSF) and sub-pixel decomposition method. The optimal reconstruction scale was determined through information entropy theory. A weighted Hue, Saturation, Value (HSV) transformation was employed to fuse shaded relief model (SRM) data with FY-3B images. The reconstruction effects of the scaling up and scaling down methods were systematically compared.【Result】(1) The PSF-based scaling up method demonstrated superior performance to conventional interpolation techniques, exhibiting the minimum mean elevation error and RMSE and the highest alignment with original DEM contour lines. (2) Terrain information entropy analysis revealed that the optimal DEM reconstruction scale for the study area was ≤120 m. (3) Reconstructed images at multiple scales significantly improved the visualization of the original 1.1 km resolution FY-3B imagery, effectively correcting anti-stereo phenomena and enhancing the representation of terrain details in fire environments. (4) The 15 m scale fused image exhibited the best performance in quantitative metrics, including improved terrain feature contrast and the highest accuracy in forest fire area extraction.【Conclusion】Integrating multi-scale DEM data through terrain scale conversion and image reconstruction methods markedly enhanced the fire environment representation and visualization quality of FY-3B forest fire monitoring imagery. These advancements provide critical technical support for improving satellite-based forest fire monitoring efficacy and optimizing fire management strategies. Future research should explore the statistical characteristics of disaster-affected regions across varying scales to further refine the matching of terrain units for enhanced visualization outcomes.
