Cliff change detection using siamese kpconv deep network on 3d point clouds

Abstract

Mainly depending on their lithology, coastal cliffs are prone to changes due to erosion. This erosion could increase due to climate change leading to potential threats for coastal users, assets, or infrastructure. Thus, it is important to be able to understand and characterize cliff face changes at fine scale. Usually, monitoring is conducted thanks to distance computation and manual analysis of each cliff face over 3D point clouds to be able to study 3D dynamics of cliffs. This is time consuming and inclined to each one judgment in particular when dealing with 3D point clouds data. Indeed, 3D point clouds characteristics (sparsity, impossibility of working on a classical top view representation, volume of data, …) make their processing harder than 2D images. Last decades, an increase of performance of machine learning methods for earth observation purposes has been performed. To the best of our knowledge, deep learning has never been used for 3D change detection and categorization in coastal cliffs. Lately, Siamese KPConv brings successful results for change detection and categorization into 3D point clouds in urban area. Although the case study is different by its more random characteristics and its complex geometry, we demonstrate here that this method also allows to extract and categorize changes on coastal cliff face. Results over the study area of Petit Ailly cliffs in Varengeville-sur-Mer (France) are very promising qualitatively as well as quantitatively: erosion is retrieved with an intersection over union score of 83.86 %.