When debris flows occur in medium to large gullies,they can cause significant damage.Accurately assessing their disaster risk is key to disaster prevention and reduction.Due to the difficulty that Convolutional Neural Networks(CNNs)face in capturing long-range dependencies,which hampers the extraction of global features for medium to large gully-type debris flow,this paper proposes a Convolutional Transformer Net(CTNet)that combines CNNs and Vision Transformer(ViT).First,positional encoding is added at the start to en-hance the model's spatial awareness,and the network is widened along with the use of grouped convolutions to improve feature extraction capabilities.Next,the partitioned image is linearly projected into a sequence and fed into the ViT Block,where the multi-head self-attention mechanism captures dependencies across different subspaces.Finally,CTNet classifies the gullies into high-risk and low-risk categories based on the risk score vector.In experiments on the Nujiang Gully dataset,CTNet significantly improved the recognition per-formance for medium to large gullies,achieving an accuracy of 88.97％and a precision of 90.85％,which notably outperforms other comparison models.Therefore,CTNet effectively overcomes the limitations of CNNs in global modeling and provides a new paradigm for high-precision disaster-risk gully identification.

Gully-type debris flows are typical sudden-onset geological hazards widely distributed in mountainous regions. Their formation is controlled by complex geomorphological conditions and heterogeneous environmental factors, making the accurate identification of debris-flow formative environments a persistent challenge. To address the limitations of traditional approaches in representing irregular gully boundaries, modeling cross-regional spatial dependencies, and integrating multi-source data, this study proposes a novel deep learning framework named DSTENet (Deformable Semantic-Topological Enhanced Network). First, deformable RoI pooling and a triple-branch convolutional pathway were incorporated to improve the model's adaptability to irregular gully geomorphological boundaries. Second, a hybrid ConvNeXt-Transformer architecture was adopted to jointly capture local texture features and global contextual information. Finally, a feature fusion module combining local refinement convolution and global average pooling was designed to enhance feature discrimination and multi-scale information integration. Using typical debris-flow-forming gullies in Yunnan Province, China, as the study area, a multi-source dataset comprising topographic and multispectral features was constructed for classification experiments. The results show that DSTENet achieves an overall accuracy of 87.54%, outperforming standard CNN baselines, and demonstrates stable, robust performance in identifying gully-type debris-flow formative environments across complex geomorphic settings.

To improve the accuracy and intelligence of geological disaster risk assessment, this study proposes a novel method based on a Multi-strategy Improved Grey Wolf Optimization algorithm combined with the Elman Neural Network (MIGWO-Elman). The model utilizes spatial factors such as NDVI extracted from remote sensing imagery, and slope and aspect derived from DEM, integrated with multi-source data including geological and rainfall information. A regional feature database is constructed using GIS techniques. In the model design, strategies including chaotic reverse learning, nonlinear convergence control, Levy flight perturbation, and self-history best are introduced to enhance the global search capability and convergence stability of the algorithm. Wenchuan County is selected as the study area for experimental validation. Results show that the MIGWO-Elman model outperforms other comparative models in terms of accuracy (91.76%), recall (93.57%), F1 score (0.919), Kappa coefficient (0.835), and AUC value (0.948). Moreover, 93.38% of known disaster points fall within high and very high-risk zones, demonstrating a strong agreement with actual disaster distribution. This study confirms that the proposed model, supported by remote sensing and GIS, can provide effective technical support for geological disaster risk zoning and emergency management on digital earth platforms.

Rapid and accurate identification of gully-type debris flows is vital for safeguarding lives and property in mountainous regions. To address the issues of inaccurate and insufficient gully feature extraction, we propose a dual-branch feature extraction module based on multi-scale and self-attention mechanisms. Integrated into the feature layer of Convolutional Neural Network (CNN) to solve the multi-scale feature extraction problem of gullies. First, the feature extraction component consists of a dual-branch structure with a global feature extraction part based on self-attention mechanisms and a local feature extraction part based on multi-scale methods, designed to extract gully features at different scales and establish connections among them. Next, during the multi-scale feature map output stage, feature maps from different scales are fused by adaptively updating their weights. Finally, a multi-layer perceptron classifies gully images. Results indicate that the proposed module effectively extracts feature at different scales, accurately representing the intrinsic characteristics of gullies. When integrated with traditional CNNs, the module significantly improves the recognition performance of gully-type debris flows. For instance, with ResNet18, the model achieves an accuracy of 89.7% and recall of 85.7%, representing improvements of 17.3% and 21.4%, respectively, over the baseline.

Earthquakes deposit loose materials in gullies, making seismic mountain gorges prone to landslides and debris flows. Monitoring and predicting ground deformation in these areas is essential. This study introduces a CS-Elman prediction model based on SBAS-InSAR monitoring. SBAS-InSAR technology analyzes 36 Sentinel-1A im- ages from Wenchuan County, Sichuan Province, China (June 2021 to October 2023), focusing on deformation areas and precipitation data. The CS algorithm optimizes the Elman network's parameters, using SBAS-InSAR data as training samples. Validation shows that: (1) Wenchuan County experiences varied deformation, with Banzi Gully in Miansi Town showing the highest uplift at 183.74 mm/a due to heavy rain. (2) As sample size increases, prediction error decreases and accuracy improves. Predictions suggest ongo- ing uplift at about 4.50 mm per month above Cutou Gully over the next four months, highlighting the need for continued monitoring.

This study addresses the challenges of conducting extensive debris flow surveys in the rugged terrain of Yunnan, China. It introduces the "Residual Multi-Source Data Fusion Network (RMDFNet)," a deep learning model developed through multi-source data fusion. RMDFNet utilizes various geographic data, including Digital Elevation Models (DEM), remote sensing imagery, soil, rock type, and vegetation data, to assess the susceptibility of debris flow gullies for early warnings. Its enhanced residual structure enables effective extraction of lower-level features. Model training incorporates an improved loss function based on prior probability, thereby enhancing recognition across diverse gully types. Gully susceptibility levels are evaluated by comparing gullies under assessment to debris flow occurrences.Experimental results show RMDFNet's outstanding performance, achieving an 84.62% accuracy rate. Compared to conventional deep learning networks, there is a significant improvement in accuracy, nearly 4%, and in comparison to other machine learning methods, the accuracy is enhanced by almost 10%. In evaluating susceptibility levels in debris flow gullies across the entire Nujiang Prefecture, the model identified 82 out of 85 gullies as extremely high risk, constituting 96.47% of gullies categorized as extremely high risk. This underscores the feasibility of utilizing image-based learning to assess the susceptibility levels of debris flow-prone gullies. The study introduces a novel approach to preventing and managing debris flow disasters, emphasizing the potential applications of data fusion and deep learning in the field of geological susceptibility early warning systems. It provides robust support for disaster prevention in mountainous regions.

In response to the problem of excessive human intervention and poor accuracy of traditional methods for assessing the susceptibility of large-scale debris flow by selecting disaster-causing factors combined with statistical models, this paper proposes a DMANet (Dense-Multiscale-Attention Net) model for assessing the susceptibility of gully-type debris flows based on an improved Dense Net. Firstly, to improve assessment efficiency without reducing accuracy, the network structure of Dense Net is optimized. Secondly, the model uses an improved inception module to extract features of different scales to obtain local and overall features of the gully and integrates shallow features to reduce feature loss. Finally, to reduce the noise impact in gully images, the CBAM attention mechanism is introduced into the network, allowing the network to focus more on target samples. Based on the similarity between the gully to be assessed and the gully where debris flow has occurred, the susceptibility is predicted, and the results of susceptibility prediction are divided into five levels: extremely low susceptibility zone, low susceptibility zone, medium susceptibility zone, high susceptibility zone, and extremely high susceptibility zone using the natural breakpoint method. Experimental results show that the extremely high susceptibility zone of debris flow is mainly concentrated in the areas with abundant water systems on both sides of the Nu Jiang River, accounting for 56.53% of the entire study area, and the proportion of debris flow is 85.36%. The ACC value and AUC value of the DMANet model reach 77.43% and 0.7903 respectively, indicating that the model is a high-performance method for assessing the susceptibility of debris flow. This method can objectively and efficiently assess the susceptibility of debris flow, providing some ideas for the prevention and control of debris flow in the future.

In large-scale debris flow susceptibility assessments, there is often excessive manual intervention, low efficiency, and inadequate model accuracy. To address these issues, this paper integrates multiple data sources and proposes a Multi-channel and Multi-scale Residual Network (MMRNet) for automatic extraction of gully features. Firstly, MMRNet employs a multi-scale feature fusion module to capture both local and global features of gullies, enhancing the model’s feature representation capabilities. It then uses an improved residual structure to fuse shallow features, compress features, and improve assessment efficiency. Additionally, channel rearrangement techniques are used to enhance feature flow. Finally, susceptibility prediction is made based on the similarity between the gully under evaluation and gullies where debris flows have occurred. The natural breakpoint method is used to classify susceptibility results into five levels. Experimental results show that the very high susceptibility zones for debris flows are mainly concentrated in areas with abundant river systems along the Nujiang River, covering 61.68% of the entire study area, with a debris flow proportion of 98.78%. The MMRNet model achieves an accuracy (ACC) of 81.6% and an area under the curve (AUC) of 0.8320, indicating that this model is a high-performance method for debris flow susceptibility assessment.

In response to issues such as incomplete segmentation and the presence of breakpoints encountered in extracting debris-flow fans using semantic segmentation models,this paper proposes a local feature and spatial attention mechanism to achieve precise segmentation of debris-flow fans.Firstly,leveraging the spatial inhibition mechanism from neuroscience theory as a foundation,an energy function for the local feature and spatial at-tention mechanism is formulated.Subsequently,by employing optimization theory,a closed-form solution for the energy function is derived,which ensures the lightweight nature of the proposed attention mechanism algorithm.Finally,the performance of this algorithm is compared with other mainstream attention mechanism algorithms embedded in semantic segmentation models through comparative experiments.Experimental results demonstrate that the proposed method outperforms both the original models and mainstream attention mechanisms across various classic models,effectively enhancing the performance of net-work models in debris-flow fan segmentation tasks.

In response to the challenges posed by rugged terrain in Yunnan, hindering large-scale mudslide screening efforts, this article introduces a dual-channel Convolutional Neural Network (CNN) constructed using elevation data from historical mudslide-prone valleys (Digital Elevation Model, DEM) and remote sensing imagery. The network is designed to facilitate the comprehensive assessment of potential mudslide hazards in gullies, serving as a crucial tool for early mudslide disaster warning. The model initially employs an enhanced residual structure to extract fundamental features from both types of data. Subsequently, it leverages the SE module and deep separable structure to emphasize the importance of relevant features and expedite model convergence. Finally, the model classifies the gullies under evaluation based on their similarity to gullies where mudslides have previously occurred. Experimental results demonstrate the model’s robust performance in assessing mudflow-prone gullies, achieving an impressive precision rate of up to 81.10% and a recall rate of 82.76%. When applied to evaluate the potential hazard of mudslide gullies across the entirety of Nujiang Prefecture, the model predicts that 87.80% of the mudslide locations are at an extremely high risk. These findings underscore the viability of utilizing image-based gully feature analysis for assessing the hazard levels of mudslide-prone gullies.

Debris flow susceptibility evaluation plays a crucial role in the prevention and control of debris flow disasters. Therefore, this article proposes a convolutional neural network model named multi-level feature extraction network (MFENet). First, a dual-channel CNN architecture incorporating the Embedding Channel Attention mechanism is used to extract shallow features from both digital elevation model images and multispectral images. Subsequently, channel shuffle and feature concatenation are applied to the features from the two channels to obtain fused feature sets. Following this, a deep feature extraction is performed on the fused feature sets using a residual module improved by maximum pooling. Finally, the susceptibility index of gullies to debris flows is calculated based on the similarity scores.

Debris flow is a natural geological disaster that frequently occurs in mountainous areas, posing a serious threat to the lives and property of local residents. However, conducting large-scale on-site investigations of debris flows is challenging due to the complex terrain of these areas. To address this issue, a CNN model based on dual-channel feature fusion is proposed to assess the susceptibility of debris flows. First, a dual-channel architecture is constructed, where a 2D CNN extracts the spatial features of the DEM image, and a 3D CNN extracts the spatial-spectral features of the multispectral image. Second, the residual structure is improved for feature extraction, and a CBAM block is added to enhance the network's ability to extract key features of valley images. Then, a fusion module is designed to fully integrate the features of the two channels. Finally, the susceptibility index is calculated based on the similarity score, and the susceptibility assessment results are divided into five levels and verified. The proposed model achieves an accuracy of 78.62% in the valley classification task and shows promising results in assessing the susceptibility of debris flows. Specifically, the proportion of high and moderately susceptible areas is 75.52%, the debris flow ratio is 96.38%, and the frequency ratio precision is 93.45%. These results demonstrate the feasibility of the proposed susceptibility assessment method and highlight its potential as a reference for debris flow prevention and disaster reduction efforts.

Susceptibility mapping plays a crucial role in debris flow prevention and control. One of the hardest hit areas by debris flows and a typical mountainous region, Nujiang Prefecture in the southwest of China, is selected as the study area. This paper conducts susceptibility mapping by directly applying a CNN model to DEM, remote sensing, lithology, soil type, lithology and precipitation data. First, each type of data is used as an independent input to extract shallow features. Then, feature fusion is conducted by channel shuffle and dense connection. Finally, precipitation is included to give the susceptibility. A novel loss function based on focal loss is proposed to improve the model performance. Compared with 7 conventional CNN models, our model reaches the highest accuracy of 86%. The kappa coefficient is at least 0.1 higher than other models. The kappa coefficient is also the highest compared with back propagation neural network, support vector machine and random forest. The obtained susceptibility map confirms well with the historical debris flow records and is more accurate than the previous study in this region. Furthermore, mid-feature visualization is used to demonstrate the feature extraction ability of our model. The extracted features, including slope and NDVI are highly consistent with the manually calculated ones. Our model also captures the water-flowing process. The new method has the potential to be applied to debris flow susceptibility assessment all over the world.

图像超分辨率重构是指将低分辨率图像生成对应的高分辨率图像,在许多领域有着重要作用.文章在SRCNN方法的基础上,提出了改进模型.首先,在SRCNN基础上使用小卷积代替大卷积.其次,加入残差结构.最后,在前两层网络后加入ReLU激活函数.结果表明,scale为3、4、6、8的PSNR分别提升了0.140 3 dB、0.084 5 dB、0.147 2 dB、0.113 5 dB,模型性能较改进前有所提升.

森林生物量调查监测是正确认识和管理森林生态系统的基础性工作,分析已有的调查资料是提高抽样效率的有效途径.采用四川省森林资源连续清查第六次至第九次数据,即 2002 年、2007 年、2012 年和 2017 年 4 个年度固定样地的调查数据,进行区域化特征分析,基于聚类分布模式进行空间分层抽样,采用不等概率抽样估计总体特征值.结果表明:区域化特征聚类分层能有效降低层内的方差,作为空间分层抽样的先验信息;在95％可靠性下,空间分层抽样活立木生物量估计精度的均值为93.41％,显著减少了外业样地调查工作量,能有效地提高抽样效率.

沟谷的泥石流危险度评价是泥石流防治工作中基础且重要的一环,针对该问题,以怒江州为例,提出了一个结合立体卷积与残差结构,能同时对数字高程模型(Digital Elevation Model,DEM)数据与多光谱数据进行特征学习的神经网络模型.以整沟为研究对象,将模型在历史泥石流灾害沟谷的数据上训练后,根据相似度对沟谷的泥石流危险度进行评分,并绘制了怒江州的泥石流危险度评价图.在所有 214 条沟谷中,高风险沟谷共 114条,中风险沟谷共 40 条,低风险沟谷共 60 条.实验结果表明,该模型能在沟谷泥石流分类任务上达到最高 80%的正确率、88%的召回率以及 0.81 的Kappa系数.此外,在使用更少训练数据的实验以及对比各个不同模型的实验中,所设计的模型均表现优异.模型给出的危险度与历史灾害记录和实地考察结果基本相符.

In response to the issues of inconsistent factor selection and limited training samples in debris flow factor-based evaluation methods, this study proposed a prototypical network-based approach for assessing the susceptibility of valley debris flow disasters. The method involves organizing the training data through meta-learning and calculating the prototype center for each valley type, serving as a representative of that category. Subsequently, the distance between the features of unknown samples and the prototype center of each class is computed to determine the probability of their classification. Based on the category probabilities, the debris flow susceptibility index of the valley is calculated to obtain the evaluation grade for debris flow susceptibility. The model was applied to evaluate the valleys in Nujiang Prefecture, and its results were compared with historical disaster data, yielding a classification accuracy rate of 67.39%. The evaluation levels provided by the model align well with the severity of debris flow disasters in historical events. Compared to traditional methods such as field surveys and factor evaluation, the method proposed in this paper allows for the rapid identification and evaluation of debris flow disaster areas using remote sensing imagery, presenting new insights for research on early warning and prediction of debris flow disasters.

山区多发沟谷型泥石流,而由于山区地形崎岖,导致无法开展大面积的泥石流危险性评价工作.本文使用遥感数据、DEM(Digital Elevation Model)数据以及岩性、土壤、植被数据,构建了一个基于多源数据,能快速进行大面积排查工作的卷积神经网络模型RSDNet(Residual-Shuffle-Dense residual Net).该模型首先使用最大池化改进的残差结构对各类不同数据进行浅层特征提取,然后使用通道重排以加强各类数据底层特征间的关联性,接着使用密集残差结构对底层特征作进一步的特征提取,学习各类特征间的相互作用对潜在泥石流危险性的影响,最后根据待评价沟谷与已发生过泥石流沟谷的相似度给出沟谷的潜在危险性.在训练过程中,使用了交叉熵和基于焦点损失改进的联合损失函数,使模型能更好地区分各类沟谷的形态特征和致灾特征.RSDNet在沟谷分类任务上可达到89.7％的精确率.在对怒江州全境沟谷进行潜在危险性评价的任务中,132条历史泥石流沟谷有122条被模型判断为高危险或极危险.结果表明模型性能良好,为沟谷泥石流的危险性评价提供了新思路.

针对不同场景下火灾图像的识别问题,提出一种利用残差网络改进VGG16 的模型.首先,将VGG16 原有的 3 层全连接层改为 1 层,并增加dropout层以防止过拟合.其次,在残差块中的卷积层之后添加BatchNorm2d函数,对数据进行归一化处理.结果表明,改进的VGG16 网络准确率、召回率和AUC值等指标性能均优于VGG16和Resnet34网络,能够对火灾图像进行快速、准确的识别.

云南作为泥石流受灾最严重的省份之一,每年均会遭受重大损失.为了应对这种突发性灾害,本文基于DCHNNet(dual-channel hybrid neural network)提出了一个基于双通道的改进残差结构的卷积神经网络——双通道残差网络(two-way residual network,TWRNet).该网络能够广泛应用于泥石流沟谷图像的潜在危险性排查,实现泥石流灾害的预警.TWRNet首先采用切片的方式对数字高程(digital elevation model,DEM)数据和遥感数据分开处理,并使用改进的残差结构进行特征提取;然后将特征进行融合,并使用通道注意力机制SE(squeeze-and-excitation networks)模块进行通道增强;最后给出泥石流沟谷的分类结果.在训练过程中,本文使用了交叉熵和焦点损失构成的联合损失函数.实验结果表明,TWRNet在泥石流沟谷识别方面达到了最高89.28％的识别率和87.50％的召回率,模型性能良好.使用图像学习沟谷特征的方法来进行泥石流孕灾沟谷的识别是可行的.