Day 4

Detailed paper information

Back to list

Paper title Towards tree species mapping with deep learning using lidar and hyperspectral satellite data
  1. Delphine NOBILEAU Capgemini Speaker
  2. Imane SOUFFER Capgemini
  3. Edouard Martins EUMETSAT
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Forests hold an essential role in the planet balance on several aspect such as water supplying, biomass production and in the climate regulation. However, the alarming changing rate of Forest diversity threats its sustainability and makes tree species mapping and monitoring one of the major worldwide challenges. Despite all the deployed efforts for tree species detection, Forest inventories databases still relay on field surveys that give inconsistent data with a highly restrictive cost which is unsuitable for large scale monitoring. Earth observation satellite sensors such as LiDAR (Light Detection and Ranging) altimeters and Hyperspectral sensors would take the lead in improving the forest tree’s occupation detection by coupling surface spectral resolved data and 3D canopy information. Although some previous research carried out tree species classification using these two technologies, those studies were mainly based on high resolution Unmanned Aerial imagery (UAV imagery) instead of remote sensing satellite data.
This paper explores GEDI (Global Ecosystem Dynamics Investigation), PRISMA (Hyperspectral Precursor of the Application Mission) and MSI (MultiSpectral Instrument) Sentinel-2 potentials in tree species identification. The work baseline also reduces data processing limitations through the use of hyperspectral dimensionality reduction techniques and data augmentation approaches. Furthermore, the paper reviews machine learning algorithms and deep learning models for tree mapping. Along with those studies, we propose a supervised deep learning framework based on the Hyper3DNet CNN model to locate the major tree species within an image pixel.
Different experiments are led to first provide a performance comparison between the proposed framework and other machine learning models, and secondly report a performance comparison between different satellite imagery products. The established work plan is applied on four different region datasets (England, Spain, France and Scotland) for accuracy assessment
Results showed that hyperspectral data are critical for tree species detection, scoring a 95 % average classification accuracy. Thus, the hyperspectral profile is a robust discriminative source of information for tree species classification. Moreover, we concluded that Lidar and multispectral data unfit the automated established training approach, and that deep leaning performs better than random forest and svm classifiers which reach only a 70% average classification accuracy. Even if the study endorses the robustness of hyperspectral satellite data in tree species mapping and proves that CNN models are inadequate for lidar data, further tests with multilayer perceptrons on the laser altimeter data could be considered for a global tree species automatic discriminative solution.