A novel approach for landform classification based on salience detection integrating expert knowledge and deep learning

几年来，地球测量学在数据获取方面的进步使在非常大的区域内获得非常高分辨率的数据成为可能。特别是，我们现在拥有无与伦比的详细程度的数字地形模型，用于分析地形和环境。然而，这需要健壮和自动的处理方法。在其他问题中，地貌的检测和分析仍然是一个难题。地貌是特定背景下地貌作用的结果。它的定义本质上是模糊的，因此目前的方法是针对某些类型的数据和某些地形。近年来，由于图像处理中的深度学习方法取得了一些进展，但这些方法很难推广。因此，我们提出了一种基于地形显著要素的新方法：遵循地貌学家的认知方法，地貌不是以形态测量标准来表征，而是以显著程度来表征。这些突出点包括在描述地形的点(峰、坑)和线(海沟、脊)的网络(图)中。本研究的目的是提出一种新的基于深度学习技术的地貌检测方法。我们认为，这种方法更稳健，因为它对地形模型分辨率的依赖较小，而且它可以更容易地整合专家的描述。学习方法的一个限制是拥有已经标记了显著程度的数据。这样的数据很少，而且它们的构建也很繁琐。因此，本计划提出了三个目标：1-提供一种方法，从专家知识和数据库中自动计算所需的显著属性，并将它们集成到深度神经网络中。2-设计一个以半监督方式从图表中自动分类显著的应用程序。只有部分数据被标记，系统必须从它派生分类规则。3-设计一个应用程序，以一种无监督的方式从图形中自动分类显著。数据不加标签，检测完全取决于专家提供的定义的质量。将使用关于不同形式(冰川、冰川峡谷、滑坡地带)的几个案例研究来验证这些方法。该项目将培养8名高素质人才。它将提供一种新的地貌分析方法，供地貌学家和环境专家使用，并使他们能够分析大型数据集，以便更好地了解环境现象。

For several years, advances in data acquisition in geomatics have made it possible to acquire very high-resolution data over very large areas. In particular, we now have digital terrain models at unequaled levels of detail for the analysis of relief and the environment. However, this requires robust and automatic processing methods. Among other issues, the detection and analysis of landforms remains a difficult problem. A landform is the result of geomorphological processes in a specific context. Its definition is inherently vague and current methods are therefore specific to certain types of data and certain landforms. Lately, progress has been made thanks to deep learning methods coming from image processing, but they are difficult to generalize. We therefore propose a new approach based on salient elements of the terrain: a landform is not characterised by morphometric criteria but by saliences, following the cognitive approach performed by geomorphologists. These saliences are included in a network (graph) of points (peaks, pits) and lines (thalwegs, ridges) which describes the terrain. The objective of this research program is to propose new methods for landform detection based on deep learning techniques on graphs. We consider that this approach is more robust since it is less dependent on the terrain model resolution and that it makes it possible to more easily integrate the descriptions made by the experts. A limit to the learning methods is to have data where saliences have already been labelled. Such data are rarely available and their construction is tedious. Three objectives are therefore proposed for this program: 1- Provide a methodology to calculate the required salience properties automatically from expert knowledge and the database and integrates them into a deep neural network. 2- Design an application for automatic classification of saliences from a graph in a semi-supervised way. Only part of the data is labelled and the system must derive classification rules from it. 3- Design an application for automatic classification of saliences from a graph in an unsupervised way. Data are not labelled and the detection depends entirely on the quality of the definition provided by experts. Several case studies on various forms (moraines, glacial valleys, landslide zones) will be used to validate the methods. This project will train eight highly qualified persons. It will provide a new approach to landform analysis that can be used by geomorphologists and environmental experts and will allow them to analyse large datasets for a better understanding of environmental phenomena.