Research map / Learning theory and algorithms / Prediction and decision losses
Contrasting Local and Global Modeling with Machine Learning and Satellite Data: A Case Study Estimating Tree Canopy Height in African Savannas
Esther Rolf, Lucia Gordon, Milind Tambe, Andrew Davies
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train local FCN on local labels instead of using the best global map | Local FCN vs Meta/ETH/GLAD maps · Karingani TCH | RMSE: 4.51 -> 1.64 (ETH); 3.09 -> 1.65 (GLAD); 3.51 -> 1.65 (Meta); 2.43 -> 1.65 | Improved |
| initialize XceptionS2 from global weights instead of random init (ablation) | XceptionS2 · Karingani TCH | RMSE: 1.98 vs 2.05 (random), 1.94 (random, no latlon) | No clear effect |
| initialize U-Net with SSL4EO self-supervised weights (ablation) | U-Net ResNet-18 · Karingani TCH | RMSE: 2.27 -> 2.05 | Improved |
| fine-tune entire U-Net instead of only the decoder (ablation) | U-Net SSL init · Karingani TCH | RMSE: 2.05 -> 2.08 | No clear effect |
| remove location embedding from XceptionS2 (ablation) | XceptionS2 random init · Karingani TCH | RMSE: 2.05 -> 1.94 | Improved |
| use small FCN instead of globally initialized models | Local FCN vs best globally initialized model · Karingani TCH | RMSE: 0.34m lower average RMSE, 0.28m lower MAE | Improved |
| add near-infrared band to RGB input (ablation) | Local FCN · Karingani TCH | RMSE: about 10% lower average RMSE | Improved |
| add remaining 8 Sentinel-2 bands (ablation) | Local FCN · Karingani TCH | RMSE: small positive effect; 3-channel vs 12-channel gap 0.22m | Improved |
| add 3-channel location embeddings (ablation) | Local FCN · Karingani TCH | RMSE: did not change performance greatly, if anything decreased | No clear effect |
| increase number of training sites from 3 to 12 (ablation) | Local FCN · Karingani TCH | RMSE: 0.25m average gap between 3 and 12 sites | Improved |
| use FCN instead of XceptionS2 (ablation) | Local models · Karingani TCH | RMSE: 0.30m gap in average RMSE | Improved |
| use local FCN predictions instead of ETH/GLAD maps for biomass calibration | AGB estimation · Aboveground biomass calibration | R2: 0.11 (ETH) -> 0.52; 0.22 (GLAD) -> 0.52 | Improved |
Closest papers by meaning
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.86)
- Estimating Canopy Height at Scale (ICML 2024 · similarity 0.86)
- Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation (ICML 2025 · similarity 0.86)
- Open-Canopy: Towards Very High Resolution Forest Monitoring (CVPR 2025 · similarity 0.86)
- AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 · similarity 0.86)
- HarvestNet: A Dataset for Detecting Smallholder Farming Activity Using Harvest Piles and Remote Sensing (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
