结合坡度增强与多尺度注意力的滑坡检测算法

Landslide detection combining slope enhancement with multi-scale attention mechanism

遥感影像滑坡区域快速精确识别在灾害应急响应、防灾减灾等方面发挥重要作用。传统滑坡区域遥感影像目视解译高度依赖人力、经验等,导致处理效率较低,难以满足快速响应需求。为此,文中提出一种结合坡度增强机制与多尺度注意力机制(SMARDN)的多源遥感影像滑坡自动检测算法,对多源数据进行特征提取与分割,高效生成滑坡检测二值图。在该模型框架下,提出多尺度注意力模块进行地物目标的尺度差异化特征提取与特征映射权重校准;充分考虑坡度影像与滑坡图的强相关性,提出坡度增强机制,强化坡度信息挖掘与特征重用;采用交叉熵与Dice联合损失函数对网络模型训练进行约束,提升模型分割精度。文中采用LandSlide4Sense多源数据集对提出的滑坡检测模型进行验证。定性实验表明,SMARDN网络在中分辨率多源数据上的滑坡检测能力优于主流滑坡检测方法,表明文中方法能更精准地刻画滑坡区域。定量实验表明,相较于经典Unet网络,文中方法在IoU和F 1分数两项定量指标上分别提升4.31%和3.33%,具有更高的滑坡检测精度。

Rapid and accurate identification of landslide areas in remote sensing images plays an important role in disaster emergency response,disaster prevention and mitigation.The visual interpretation of remote sensing images in traditional landslide areas is highly dependent on manpower and experience,and the processing efficiency is low,making it difficult to meet the needs of rapid response.To this end,this paper proposes a multi-source remote sensing image landslide automatic detection algorithm SMARDN that combines the slope enhancement mechanism and the multi-scale attention mechanism.Segmentation to efficiently generate binary images for landslide detection.Under the framework of this model,the multi-scale attention module is used to extract the scale-differentiated feature of the surface object and the weight calibration of the feature map;fully considering the strong correlation between the slope image and the landslide map,a slope enhancement mechanism is proposed to strengthen the slope information mining and Feature reuse;the joint loss function of cross entropy and Dice is used to constrain the network model training and improve the model segmentation accuracy.In this paper,the LandSlide4Sense multi-source dataset is used to verify the proposed landslide detection model.Qualitative experiments show that the landslide detection ability of SMARDN network on medium-resolution multi-source data is better than mainstream landslide detection methods,indicating that the method in this paper can describe landslide areas more accurately.Quantitative experiments show that compared with the classic Unet network,the method in this paper improves the two quantitative indicators of IoU and F1 scores by 4.31%and 3.33%,respectively,and has higher landslide detection accuracy.