基于无监督表达学习的森林地貌特征建模及林火易发性评估

An unsupervised representation learning approach for modelling forest landform characteristics and fire susceptibility assessment

近几十年来,极具破坏性的森林火灾在全球范围内造成了巨大的损失,且频率仍在逐年提高。基于历史统计数据对林火发生风险进行预测是一个较为可行的防控火灾的方法。传统的统计学习方法多使用人工指定的方式对数据进行降维及特征提取。而随着遥感技术的不断发展,高精度网格化的多维森林地貌信息的获取难度不断降低。使用人工提取特征的方式很难利用这类数据,从而限制了这类方法在真实复杂环境下的性能。为此介绍一种通过深度无监督表达学习对森林地理信息进行建模的全新方法,并借助一组区域火灾风险预测实验对比无监督学习与其对应的有监督模型的性能。结果表明该方法对森林地貌特征建模的有效性。

Destructive wildfires have caused extraordinary losses in both economic and natural property worldwide with an even increasing frequency in recent decades.One practical approach in forest fire susceptibility prediction is using statistical learning methods to learn from historic data.Conventional methods use handcraft feature to reduce data dimension.With the continuous development of remote sensing technology,the difficulty of obtaining high-precision gridded multidimensional forest landform information is constantly decreasing.It is difficult to make full use of such data through handcraft features,which limits the performance of conventional methods when applied in the real world.This paper introduces a novel approach to model forest geographic information through deep representation learning,which is,leveraging deep convolutional neural network and state-of-the-art representation learning methods to extract feature embedding for a given area of interest.Fire susceptibility assessment experiments are used to evaluate the proposed method and compare the unsupervised learning and its supervised counterpart to show its effectiveness.