Page 154 - 《武汉大学学报(信息科学版)》2025年第6期
P. 154
1176 武 汉 大 学 学 报 (信 息 科 学 版) 2025 年 6 月
tion factor system. The MFL-LN method employed the deep feature enhancement module (DFEM) to
construct the network architecture, reducing the number of parameters and computational load to achieve
model lightweighting. Additionally, it dynamically adjusted the attention to different channel features, en‑
hancing the ability to perceive and utilize important features. Subsequently, the multi-scale feature fusion
module (MSFF) was employed to integrate the features extracted through down-sampling in the encoding
phase, enabling information interaction on the fused deep features and adaptively adjusting the feature
weights. It can effectively capture both global and local spatial characteristics of landslides and compensate
for the feature extraction loss caused by down-sampling, thereby improving the predictive capability for
landslide susceptibility. Results: An ablation study was conducted on the MFL-LN method, and it was
compared with the five traditional models. The accuracy of the susceptibility evaluation models was ana‑
lyzed from three perspectives: landslide susceptibility zoning maps, landslide hazard point density zoning
statistics, and evaluation model accuracy. The hazard point density for very low and very high susceptibility
areas in the MFL-LN model was 0.009 and 0.557, respectively, ensuring that the hazard point density is
lower in very low susceptibility areas and highest in very high susceptibility areas. Additionally, the area un‑
der the receiver operating characteristic curve was the highest at 0.837. The proposed model was analyzed
for lightweight optimization based on parameter count and floating-point operations. The parameter count
of the MFL-LN model was 0.081, and the floating-point operation count was 41.102. It effectively re‑
duced the model training costs and improved the accuracy of landslide susceptibility zoning. Conclusions:
By selecting Ankang City in Shaanxi Province as the study area and analyzing the historical landslide hazard
data, the MFL-LN model was constructed utilizing the DFEM and MSFF with an encoder-decoder struc‑
ture to evaluate landslide susceptibility. It improved computational efficiency while maintaining the perfor‑
mance of model, making the landslide prediction model more efficient and practical.
Key words: landslide; susceptibility; lightweighting; attention; multi-scale feature learning
滑坡是中国最常见的地质灾害之一,受多种 利用统计分析、传统机器学习、深度学习等方法,
环境因素影响,具有突发性、高危害性和频发性 从大量数据中分析挖掘滑坡特征和规律,构建预
等特点。频繁的人类工程活动严重影响了自然 测模型实现滑坡易发性评价。其中,统计分析方
[6]
气候和地质环境,加剧滑坡发生的频率和破坏程 法主要包括证据权法 、信息量法 、确定性系数
[7]
[9]
[8]
度。国家统计局发布的《中国统计年鉴 2023》地 法 、频率比 等;传统机器学习方法主要有逻辑
质灾害统计数据显示,全国共发生地质灾害 5 659 回归 [10-12] 、随机森林 [13-16] 、决策树 [17-18] 、支持向量机
起 ,其 中 滑 坡 3 919 起 ,占 地 质 灾 害 总 数 的 (support vector machine, SVM) [19-21] 等;深度学习
63.95%。开展滑坡易发性分析研究对于滑坡灾 方 法 中 最 典 型 的 是 卷 积 神 经 网 络(convolutional
害 预 警 、资 源 规 划 和 灾 后 重 建 具 有 重 要 的 现 实 neural network, CNN),其具有局部连接、权值共
意义 。 享、池化操作等特点,在滑坡易发性评价中应用
[1]
根据驱动模式不同,滑坡易发性评价方法可 十分广泛 [22-28] 。数据驱动模型往往未充分考虑多
分为经验驱动模型和数据驱动模型两类。经验 源滑坡评价因子的冗余性,没有聚焦重要特征对
驱动模型通过对地质、地貌、水文等因素的分析 滑坡的贡献,未顾及复杂网络模型带来的资源需
和评价,基于专家知识和滑坡演化过程与物理机 求和成本较高的不利因素。
制构建规则系统来获取滑坡发生概率。经验驱 针对上述问题,本文提出了多尺度特征学习
动模型主要有层次分析法 、专家打分法 、模糊 的 轻 量 化 滑 坡 易 发 性 评 价 方 法(lightweight net‑
[2]
[3]
逻辑法 [4-5] 等,这类方法原理较简单且易于实现和 work based on multi-scale feature learning,MFL-
理解,对数据和计算资源要求不高,适用于数据 LN),设 计 深 度 特 征 增 强 模 块(deep feature en‑
不完备情况下滑坡易发性预测。经验驱动模型 hancement module, DFEM)和多尺度特征融合模
存在主观性和不确定性,无法挖掘数据之间的复 块(multi-scale feature fusion module, MSFF),并
杂关系,往往导致模型预测精度不高,评价结果 使用编码-解码结构组建多尺度特征学习的轻量
缺乏足够的可靠性和稳定性。数据驱动模型则 化网络。MFL-LN 通过深入挖掘和融合不同尺