Estimating Canopy Height at Scale

Authors: Jan Pauls, Max Zimmer, Una M. Kelly, Martin Schwartz, Sassan Saatchi, Philippe Ciais, Sebastian Pokutta, Martin Brandt, Fabian Gieseke

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental A comparison between predictions and ground-truth labels yields an MAE / RMSE of 2.43 / 4.73 (meters) overall and 4.45 / 6.72 (meters) for trees taller than five meters, which depicts a substantial improvement compared to existing global-scale maps. We conduct several ablation studies to evaluate the impact of the involved model components.
Researcher Affiliation Academia 1Department of Information Systems, University of M unster, Germany 2Department for AI in Society, Science, and Technology, Zuse Institute Berlin, Germany 3Laboratoire des Sciences du Climat et de l Environnement, LSCE/IPSL, France 4Jet Propulsion Laboratory (JPL), California Institute of Technology, USA 5Department of Geosciences and Natural Resource Management, University of Copenhagen, Denmark 6Department of Computer Science, University of Copenhagen, Denmark.
Pseudocode No The paper describes the approach and model architecture in text and figures but does not provide pseudocode or algorithm blocks.
Open Source Code Yes Our source code, as well as detailed documentation, are publicly available on Git Hub.6
Open Datasets Yes For the Sentinel-1 data, we adopt a methodology similar to that used by Schwartz et al. (2024)... For the Sentinel-2 data... we resort to a cloud reduction algorithm adapted from Braaten (2024)... As ground-truth labels, we resort to the data provided by the GEDI mission. In particular, we make use of the so-called RH100 metric... To mitigate the impact induced by incorrect canopy height values in mountainous areas, we resort to data from the Shuttle Radar Topography Mission (SRTM)...
Dataset Splits Yes Overall, 100 000 patches are extracted, out of which 80% are used for training, 10% for validation, and 10% for testing. The validation set was used for hyperparameter selection.
Hardware Specification Yes The computational process consumes approximately 1 500 GPU hours on a GPU cluster with various GPU devices (e.g., Nvidia RTX3090s and A100s). We also appreciate the hardware donation of an A100 Tensor Core GPU from Nvidia.
Software Dependencies No The paper mentions architectural components and optimizers (U-Net, ResNet-50, AdamW, Huber loss) but does not provide specific software version numbers for dependencies like Python, PyTorch, or TensorFlow.
Experiment Setup Yes We optimize the model weights using the Adam W (Loshchilov & Hutter, 2017) optimizer with a weight decay of 0.001, a batch size of 32, and an initial learning rate of 0.001. We also resort to a linear learning rate warm-up for the first 10% of the iterations and a linear learning rate scheduler for the remaining 90%. Additionally, gradient clipping was applied to prevent gradient explosion. Finally, the shifted Huber loss was used during training, i.e., LS with the Huber loss as pixel-wise loss function L in Equation (4).