Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation
Authors: Jan Pauls, Max Zimmer, Berkant Turan, Sassan Saatchi, Philippe Ciais, Sebastian Pokutta, Fabian Gieseke
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our model accurately predicts canopy height over multiple years given Sentinel-1 composite and Sentinel 2 time series satellite data. Using GEDI Li DAR data as the ground truth for training the model, we present the first 10 m resolution temporal canopy height map of the European continent for the period 2019 2022. ... All results in this section are based on 1,500 randomly selected validation points... Table 2. Comparison of various model configurations for 2020. ... Table 3. MAE comparison over multiple years for all model configurations. ... Section 4.3. Method Comparison. ... B. Ablation Studies. |
| Researcher Affiliation | Academia | 1Department of Information Systems, University of M unster, Germany 2Department for AI in Society, Science, and Technology, Zuse Institute Berlin, Germany 3Jet Propulsion Laboratory (JPL), California Institute of Technology, USA 4Laboratoire des Sciences du Climat et de l Environnement, LSCE/IPSL, France 5Department of Computer Science, University of Copenhagen, Denmark. |
| Pseudocode | No | The paper describes its methodology using prose and refers to a 3D U-Net architecture. It does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The entire pipeline and model weights are publicly released on Git Hub3 and the resulting tree canopy height maps are accessible through Google s Earth Engine4 (Gorelick et al., 2017), ensuring reproducibility and facilitating research on large-scale forest monitoring, forest structure analysis and above-ground biomass estimation. 3https://github.com/AI4Forest/ Europe-Temporal-Canopy-Height |
| Open Datasets | Yes | Openly available satellite imagery for tree canopy height prediction is primarily sourced from three key missions: Landsat, Sentinel-1, and Sentinel-2. ... Sentinel-2 multispectral imagery is a key part of our approach. We use the Level-2A surface reflectance product (BOA) available via the Copernicus AWS... GEDI Li DAR data provides sparse height measurements that are essential for model supervision. We use the Level-2A product... |
| Dataset Splits | Yes | Our dataset comprises 800,000 randomly selected patches, each measuring 2.56 km 2.56 km (approx. 0.15% pixels have labels per patch), totaling 8 TB in size. To minimize computational load and data transfer, we use a 10% subset for training and hyperparameter tuning for the different baselines, as outlined in Section 4.1. The final model is then trained on the entire dataset. ... All results in this section are based on 1,500 randomly selected validation points, see Figure 2 for the distribution. |
| Hardware Specification | Yes | We also appreciate the hardware donation of an A100 Tensor Core GPU from Nvidia and thank Google for their compute resources provided (Google Earth Engine). |
| Software Dependencies | No | The paper mentions using Google Earth Engine and the Copernicus AWS for data, and refers to Adam as an optimizer, but does not provide specific version numbers for any software libraries, programming languages, or other dependencies. |
| Experiment Setup | Yes | Optimization Setup. We use Adam (Kingma, 2014) with an initial learning rate of 0.001, weight decay of 0.01, and gradient clipping at 1.0. We follow best practices (Li et al., 2020; Zimmer et al., 2023) and use a linear learning rate scheduler with a 10% warmup, training for 400,000 iterations with a batch size of 16 (corresponding to 8 epochs). ... Our final loss function is the Huber loss, as it effectively balances penalizing small errors while being less sensitive to outliers compared to L2 loss. |