Canopy height mapping from optical remote sensing data: reassessing deep learning methods

Introduction

This page documents the progress of my research on the global canopy height model by Nico Lang (2024). For the full data and scripts please visit my GitHub and for other projects and more information about me visit my Main Site.

Research question

Can canopy height be derived from spectral data?

The main objective of this research is to uncover the prediction mechanisms of the CNN. Central question of the work is What can be learned about canopy height from optical remote sensing.

Hypothesis

1. The model has not learned a connection of spectral values to canopy height.
   
2. Other variables like location and biome play a more important role.

Methods

This research builds on the model architecture created by Lang et al. (2023). Their promising global canopy height prediction model serves as a great base for investigating the mechanisms behind optical data canopy height predictions. By observing changes in the resulting predic- tions, after manipulating the input data, we would expect to detect (linear) patterns, if spectral properties would provide a strong explainatory value.

Consistent with the original paper, Sentinel-2 data and ESA-WordCover data from 2020 was used. The code was mainly left unchanged and one function added to allow controlled manipulation of single input bands. A workflow was created in a single script to allow easy application,

Research locations

In total 11 tiles were used for the analysis of the model. They were selected to be distributed globally, cover multiple different biomes and based on the presented tiles by the original paper.

Results

The preliminary results already show no clear trend after spectral manipulation (Fig. 1). Differences between locations have shown to be more significant than spectral manipulation, suggesting that the actual influence of spectral properties in the creation of the prediction is very low or none. Relationships between the location of the image tile and the predicted canopy height could already be observed. They will be properly analysed and shown at a later stage of this project.

Fig. 1: Average absolute relative difference to the original prediction by degree of manipulation, for all bands and locations.

Slides

Conference slides (PDF)

Sources

Lang, N., Jetz, W., Schindler, K., & Wegner, J. D. (2023).
A high-resolution canopy height model of the Earth.
Nature Ecology & Evolution, 1-12.

@article{lang2023high,
  title={A high-resolution canopy height model of the Earth},
  author={Lang, Nico and Jetz, Walter and Schindler, Konrad and Wegner, Jan Dirk},
  journal={Nature Ecology \& Evolution},
  pages={1--12},
  year={2023},
  publisher={Nature Publishing Group UK London}
}