Automated segmentation of land use from overhead imagery

Antoine Richard (1) , Assia Benbihi (1) , Cédric Pradalier (1) , V. Perez , Rosalinde van Couwenberghe

1 Georgia Tech Lorraine [Metz]

Reliable land cover or habitat maps are an important component of any long-term landscape planning initiatives relying on current and past land use. Particularly in regions where sustainable management of natural resources is a goal, high spatial resolution habitat maps over large areas will give guidance in land-use management. We propose a computational approach to identify habitats based on the automated analysis of overhead imagery. Ultimately, this approach could be used to assist experts, policy and decision makers who promote sustainable agroecology by evaluating habitat services and prioritizing land uses. The overall objective of our project is to classify the evolution of land usage since the advent of aerial imagery. In this paper, our goal is to bring automatic habitat classification to the level achieved by a human expert performing a high spatial resolution classification. This classification consists in identifying habitats such as hedges, lakes, fields, pastures or forests. Therefore, we train a machine vision algorithm to segment an overhead imagery into a dozen of expert-specified land use classes. Relying on the recent developments in machine learning, and in particular deep learning, the best machine vision model appears to be convolutional neural networks (e.g. SegNet, DeepLab). The training was performed using data from a hand-labelled high-resolution (0.5m/pixel) database around the Orne River (Moselle, France-2000km²). Aerial orthophoto are available for two time periods: 2015 and 1955. In addition, we also generated artificial 1955 data from 2015 imagery and used them as learning base for the 1955 imagery as the data available in 2015 provides more quantity and more diversity. The paper highlights the performances of these state-of-the-art machine learning algorithms for land use recognition and segmentation. It shows their potential in the context of studies in environment sciences and environmental decisions. The automatic approach presents an alternative for detailed and accurate land cover maps acquired manually, which are labor intensive and time consuming. The paper also illustrates the potential benefits of generating artificial imagery to pre-train the machine vision model and requires less annotated data. This approach may prove useful for time periods where there is few labeled data.

Mots clés: pixel-wise classification; deep learning; environment monitoring; conservation ecology

Domaines: Automatique / Robotique

Antoine Richard, Assia Benbihi, Cédric Pradalier, V. Perez, Rosalinde van Couwenberghe. Automated segmentation of land use from overhead imagery. International Conference on Precision Agriculture, Jun 2018, Montreal, Canada. ⟨hal-02573189⟩