TL;DR
This paper introduces the Canopy Height Change (CHC) dataset, a novel benchmark for tracking continuous forest canopy height changes and pixel-level uncertainties using satellite imagery and airborne LiDAR. By framing tree growth and loss as a continuous regression task rather than binary detection, the authors expose a major gap in current Geospatial Foundation Models (GFMs) and demonstrate how incorporating label uncertainty directly into training significantly improves model performance.
What Problem Does This Paper Address?
Monitoring how forest canopy height changes over time is vital for quantifying carbon sinks, tracking deforestation, and understanding ecosystem dynamics. While remote sensing and deep learning have made wall-to-wall canopy height mapping possible, tracking changes over time remains problematic.
Most existing approaches simplify the problem into a binary predicate (e.g., "loss" vs. "no loss"). This completely overlooks the continuous nature of vegetation dynamics, such as slow incremental growth or partial canopy degradation. Furthermore, the reference data used to train these models—typically derived from Airborne Laser Scanning (ALS)—is inherently noisy. Variations in flight configurations, scan geometries, and pulse densities between different campaigns introduce significant errors. Without quantifying this label uncertainty, models end up overfitting to random noise rather than learning true vegetation changes.
Methods
The authors tackle these challenges by developing an uncertainty-aware regression framework:
- Continuous Change Mapping: They generated Digital Surface Models (DSMs) by calculating the 95th height percentile on a 3-meter grid from two national ALS campaigns (2018 and 2023) in Spain. Subtracting these DSMs yielded continuous height changes ($\Delta h^{(95)}$).
- Uncertainty Modeling: They split uncertainty into two components:
- Systematic uncertainty ($\sigma_{syst}$): Estimated by analyzing residuals on flat, unchanged road surfaces.
- Statistical uncertainty ($\sigma_{stat}$): Modeled by simulating lower pulse densities on high-density reference cells. A Multi-Layer Perceptron (MLP) was trained using a Gaussian Negative Log-Likelihood (NLL) loss to predict pixel-level variance based on pulse density and local geometric features.
- Uncertainty-Aware Training: To evaluate how models handle this noise, they integrated the uncertainty layer into the loss functions during fine-tuning using two methods:
- Weighted MSE (wMSE): Scales errors by the inverse target variance, down-weighting highly uncertain pixels.
- Thresholded MSE (tMSE): Computes the loss only for pixels where the change is statistically significant ($z\text{-score} > 1.65$).
To benchmark performance, they used the PANGAEA framework to fine-tune several state-of-the-art GFMs. Frozen GFM encoders were paired with a Lightweight Temporal Attention Encoder (LTAE) for temporal fusion and a UPerNet decoder for dense pixel prediction.
Datasets
The paper introduces the Canopy Height Change (CHC) dataset:
- Coverage: $10,598\text{ km}^2$ across northern and western Spain (Cantabria and Extremadura).
- Resolution: 3-meter spatial resolution, organized into 1,123 tiles of $1024 \times 1024$ pixels.
- Data Layers: Continuous height change, maximum canopy height for both years, and a spatially resolved change variance (uncertainty) layer.
- Imagery: Co-located time series of PlanetScope optical imagery (RGB + Near-Infrared) spanning 2018 to 2023.
Unlike existing datasets like Open-Canopy (which only provides binary change labels for large losses), CHC provides a continuous metric for both growth and loss alongside quantified uncertainties.
Key Insights and Ideas
- Regression is a Blind Spot for GFMs: While large-scale pretrained GFMs excel at classification tasks, they struggle significantly with fine-grained, pixel-level continuous regression.
- Uncertainty Mitigation Works: Incorporating target uncertainty during training prevents models from wasting capacity on noisy labels. Both thresholding and weighting the loss function consistently improved regression accuracy across different architectures.
- Domain Specificity Beats Scale: A task-specific model trained explicitly on canopy height dynamics (ECHOSAT) outperformed all general-purpose GFMs, despite operating at a coarser spatial resolution ($10\text{ m}$ vs $3\text{ m}$).
Key Results
- Baseline Comparison: A fully supervised, standard UNet baseline trained with thresholded MSE (tMSE) achieved the best overall performance among the 3m models with a Root Mean Squared Error (RMSE) of 5.65 m.
- GFM Performance: None of the frozen GFM encoders managed to beat the supervised UNet baseline. The best-performing GFMs were Prithvi (RMSE: 5.90 m) and DOFA (RMSE: 5.99 m).
- The Independent Champion: The $10\text{-meter}$ resolution ECHOSAT model achieved the lowest overall error with an RMSE of 5.25 m.
- Impact of Uncertainty Weighting: Introducing uncertainty-aware loss functions yielded clear performance boosts. For instance, training Prithvi with wMSE decreased its test RMSE by 8%, while tMSE improved DOFA and the baseline UNet by 4% and 9% respectively.
- Directional Detection: While regressing the exact magnitude was difficult, models were reasonably good at predicting the direction of change. The CROMA GFM achieved an F1-score of 0.85 for detecting height increases.
Limitations and Future Work
The primary limitation highlighted is that current GFMs struggle to resolve the fine-grained, subpixel structural cues required to accurately regress continuous height changes from optical satellite time series. The models can often tell if a change happened (direction), but struggle with how much (magnitude).
The authors suggest that future research should focus on developing GFM architectures specifically optimized for continuous regression tasks and exploring better ways to fuse multi-sensor data (like combining optical PlanetScope data with radar or spaceborne LiDAR) to better capture subtle vegetation growth signals.
Summary updated