Status: 05.05.26
Experimental exploration of the global canopy height model by Lang et al.
This repository tracks the exploration of the Model presented in A high-resolution canopy height model of the Earth. Goal of this work is to understand how the model works and predicts, if any ecological relationships between spectral data and canopy height can be identified and how the model reacts to changes in input data. For a full explanation of the installation, setup and deployment see the original instructions: Installation. Additional Info for the installation can be found in the Section Setup Notes.
Thesis documented under esa99.github.io/pages/gchm
deploy.R is the main script containing the full workflow. All variables are set here before deployment. It should be deployed inside a virtual environment.
environment.yml This file contains all requirements for the virtual environment and is used for its setup. I used miniforge3 and conda.
/results Here are the result tables exportet by deploy.R. The final ones are labled "main", older and other ones are moved to /results/archive
/R Containes all scripts for the analysis of the result tables. Creation of plots, maps and tools for data preperation and handeling of backup files. /R/data inherits the plot locations.
/gchm Here are all scripts by Lang et al. including the models and their deployment. The band manipulation function was included here by adding it to: /gchm/utils/transforms.py
/documentation Short documentation of the process and maps of the tiles.
The following plot shows the results of the "America" Tiles (see section Calculation Groups), comparing the differrence to the original prediction when manipulating single bands of the input by up to 25%. The x-axis shows the degree of manipulation applied to each spectral band, expressed as a percentage (e.g., +10% indicates Band × 1.10). The y-axis represents the resulting average change in the predicted variable (in meters). Multiple spectral bands are visualized, color-coded similar to their conventional band colors (but in a colour blind freindly way), allowing comparison across different tile locations.
Overview of the average relative difference by band.
Distribution of canopy heights between the tiles.
Differences to original prediction after manipulation.
To systematically analyze the role of spectral data in deep learning models for estimating canopy height, an automated workflow was developed. This workflow encompasses data preparation, variable creation, file management, model deployment, and the export of results. Following the key elements of the process and the sequence of operations are illustrated. The spectral manipulation function is integrated into the model deployment scripts. The original code of the model remained unchanged.
The deploy.R script contains the full workflow and is deployed from bash after setting the correct conda environment and directory. At the beginning Tile-Name, Bands, Increment and direction are Set and then the script is deployed. The original code (deploy.py etc.) was modified to fit the script as well as extended by a manipulation-function performing the band manipulation inside the deployment. The original (pretrained) models and all other code are used unchanged. Functions from the package "dandelion" (https://github.com/ESA99/dandelion) were used and written specifically for the usecase of this modified deployment.
The selection process is being coordinated in consultation with the University of Munich. Corresponding Worldcover as tiles are needed, and were cropped accordinlgy.
Selection: 3 Tiles presented in the paper + 1 Demo Tile from the Paper + 3 Tiles in Europe for management comparison + 4 tiles spread globally to ensure latitudinal variance and global coverage (+ Biome Diversity). Mongolia: Old selection 50TPT, new proposal by Lukas 49UCP.
| Continent | Latitude | Name | Country | Biome | Centeroid_Elevation | Source |
|---|---|---|---|---|---|---|
| Africa | 5.8 | 33NTG | Cameroon | Tropical Moist Broadleaf | 754 | PAPER |
| Asia | 47.3 | 49UCP | Mongolia | Temp. Grasslands + Temp. Conifer | 902 | Munich |
| Asia | 2.2 | 49NHC | Malaysia | Tropical Moist Broadleaf | 401 | PAPER |
| Europe | 47.4 | 32TMT | Switzerland | Temp. Coniferous + Broadleaf | 590 | PAPER |
| Europe | 48.2 | 32UQU | Germany | Temp. Broadleaf + Coniferous | 422 | Munich |
| Europe | 52.7 | 34UFD | Poland | Temp. Broadleaf | 154 | Munich |
| Europe | 63.5 | 35VML | Finland | Boreal Forest | 201 | Münster |
| North America | 46.5 | 10TES | USA West | Temp. Broadleaf Forest | 392 | PAPER |
| North America | 37.5 | 17SNB | USA East | Temp. Broadleaf Forest | 682 | BOTH |
| Oceania | -36.6 | 55HEV | Australia | Temp. Broadleaf + Montane Grass/Shrub | 562 | Münster |
| South America | -1.4 | 20MMD | Brazil | Trop. Moist Broadleaf | 56 | Münster |
Tiles will be "grouped" for the analysis in three deployment groups, to split calculation time in blocks and allow for more efficient workflow. 1. Americas (10TES, 17SNB, 20MMD) 2. Europe (32TMT, 32UQU, 34UFD, 35VML) 3. AfAsOc (33NTG, 49NHC, 49UCP, 55HEV)
The following timing values correspond to the model deployment on the ILÖK-RS Supercomputer. Each loop represents a single combination of tile, band, and increment.
| Loops | Total [h] | Average [min] | Tile Group |
|---|---|---|---|
| 27 | 04:49 | 10:43 | |
| 99 | 20:15 | 12:16 | |
| 123 | 24:11 | 11:48 | Americas |
| 246 | 75:45 | 18:29 | EuAfMalasia |
The installation was performed generally following option B of the Installation Instructions.
miniforge3 replaces the outdated mambaforge. The command conda has to be used instead of mamba at some steps (Step 5 was a problem for example). There may be some dependencies not installed, check for those if an error occurs -> Ex.: libtiff 5 was required but there where four different versions, resulting in an error when deploying the demo script.
This code sets up to treat a folder (for example gchm) as a python package after looking for setup.py
pip install -e .Real time GPU supervision:
watch -n 1 nvidia-smiIf conda uses a local R -> use global r installation...
alias Rscript=/usr/bin/RscriptThis work is based on the paper and repository by 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}
}