Uncertainty-aware tree height change regression

Conference: ECCV 2026
Paper: ECCV Poster
Code: https://sid.erda.dk/sharelink/eP4ENGhKTv
Area: Remote Sensing
Keywords: Canopy Height Change, Geospatial Foundation Models, Uncertainty-Aware Regression, Airborne Laser Scanning, Satellite Time Series

TL;DR

The paper introduces the Canopy Height Change (CHC) benchmark dataset covering 10,598 km² at 3 m resolution with pixel-wise continuous canopy height difference and rigorous uncertainty estimates, demonstrating that uncertainty-weighted regression losses substantially improve downstream performance while foundation models still lag behind supervised baselines.

Background & Motivation

Accurate monitoring of canopy height dynamics over time is critical for quantifying terrestrial carbon sinks and sources, assessing ecosystem responses to climate disturbances, and tracking tree growth as well as deforestation. Recent breakthroughs in high-resolution satellite imagery coupled with Geospatial Foundation Models (GFMs) have enabled wall-to-wall mapping of static canopy heights at continental and global scales. However, the majority of remote sensing change detection studies continue to frame temporal canopy analysis as binary change detection (e.g., forest clearing versus undisturbed land), fundamentally overlooking the continuous nature of forest dynamics where subtle annual growth, progressive thinning, and selective logging represent essential continuous variables.

A principal bottleneck impeding continuous change estimation is the absence of large-scale, high-resolution benchmark datasets that pair multi-temporal satellite imagery with ground-truth height change targets equipped with reliable uncertainty metrics. Although bi-temporal Airborne Laser Scanning (ALS) serves as the gold standard for elevation measurements, repeated flight campaigns invariably differ in pulse density, flight altitude, sensor geometry, and scan angle. The discrete sampling of complex geometric tree canopies induces heteroscedastic measurement noise and systematic crown-apex clipping biases. When models are trained or evaluated on continuous height differences without accounting for reference label uncertainty, subtle biological signals are easily overwhelmed by acquisition artifacts.

To address this challenge, this work bridges airborne LiDAR physical error propagation and multi-temporal satellite vision. Core idea: build the Canopy Height Change (CHC) benchmark dataset covering 10,598 km² across Spain at 3 m resolution with pixel-wise systematic and statistical uncertainty estimates, establishing an uncertainty-aware regression framework for training and evaluating Geospatial Foundation Models.

Method

Overall Architecture

The proposed methodology encompasses the construction pipeline for the Canopy Height Change (CHC) dataset and the standardized evaluation protocol for fine-tuning Geospatial Foundation Models (GFMs) on uncertainty-aware continuous height change regression. The input consists of bi-temporal airborne laser scanning point clouds from the Spanish PNOA campaigns (Cobertura 2a in 2018 and Cobertura 3a in 2023) paired with co-registered PlanetScope multi-spectral optical image time series (2018–2023) at 3 m resolution. The output is a dense pixel-wise regression map predicting continuous canopy height change \(\Delta h\).

The data processing pipeline first rasterizes point clouds to digital surface models (DSMs) using the 95th height percentile on a 3 m grid to circumvent digital terrain model (DTM) interpolation errors. Next, pixel-level systematic uncertainty \(\sigma_{\text{syst}}\) and heteroscedastic statistical uncertainty \(\sigma_{\text{stat}}\) are derived via static asphalt surface calibration and pulse-density thinning simulations. Finally, during model fine-tuning, the reference variance is leveraged through variance-weighted MSE (wMSE) or confidence-thresholded MSE (tMSE) to guide GFM decoders.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Bi-temporal ALS Point Clouds (2018/2023) + PlanetScope Imagery"] --> B["1. DSM Grid Rasterization & Direct Differencing<br/>95th percentile height subtraction on ASPRS vegetation"]
    B --> C["2. Systematic & Heteroscedastic Uncertainty Estimation<br/>Static road residual analysis + pulse thinning MLP simulation"]
    C --> D["3. Uncertainty-Aware Regression Criteria<br/>Inverse-variance wMSE weighting & confidence tMSE thresholding"]
    D --> E["4. Multi-Scale Temporal Adaptation Architecture<br/>Frozen GFM encoder + LTAE temporal fusion + UPerNet dense decoder"]
    E --> F["Continuous Canopy Height Change & Uncertainty-Filtered Metrics"]

Key Designs

1. DSM Grid Rasterization & Direct Differencing: Circumventing DTM Error Propagation

Standard forestry approaches compute height differences by subtracting Canopy Height Models (\(\text{CHM} = \text{DSM} - \text{DTM}\)). However, deriving DTMs across dense canopies relies on scarce ground laser returns, introducing substantial interpolation artifacts. The authors bypass DTM dependence by rasterizing filtered point clouds (ASPRS vegetation classes 3–5) to digital surface models using the 95th height percentile within regular \(3\times 3\,\text{m}\) grid cells. The direct surface difference \(\Delta h(95) = h(95)_{2023} - h(95)_{2018}\) captures canopy dynamics cleanly. Pixels with vegetation lower than 3 m in both epochs are masked out to avoid seasonal crop and low herbaceous changes.

2. Systematic & Heteroscedastic Uncertainty Estimation: Road Residuals and Pulse Thinning Simulation

ALS height differences carry both constant systematic sensor misalignment \(\sigma_{\text{syst}}\) and variable statistical sampling error \(\sigma_{\text{stat}}\): - Systematic uncertainty \(\sigma_{\text{syst}}\) is calibrated from vertical elevation differences over static flat asphalt roads identified via land-use databases (SIOSE). The empirical residual distributions yield standard deviations between \(0.17\,\text{m}\) and \(0.23\,\text{m}\), with vertical datum offsets between \(-0.09\,\text{m}\) and \(-0.01\,\text{m}\). - Statistical uncertainty \(\sigma_{\text{stat}}\) arises from discrete point sampling over complex canopy geometry. Cells with high pulse density (\(\rho_i \ge 10\,\text{ppm}^2\), corresponding to \(\ge 90\) pulses per \(3\times 3\,\text{m}\) cell) serve as ground-truth reference surfaces \(h(95)_i^{\rho(10)}\). Numerical thinning generates degraded pulse samples (\(1 < \rho < 10\,\text{ppm}^2\)) to calculate density-induced deviations \(d_i\) alongside geometric covariates \(v_i\). A multi-layer perceptron (MLP) adjustment model \(f(\rho, v)\) is trained on \(7.935\times 10^5\) cells using a Gaussian Negative Log-Likelihood (NLL) objective:

\[\mathcal{L}_{\text{NLL}} = \frac{1}{N}\sum_{i=1}^N \left( \frac{\ln \hat{\sigma}_i^2}{2} + \frac{(d_i - \hat{b}_i)^2}{2\hat{\sigma}_i^2} \right)\]

The final total reference variance for each pixel combines both sources: \(\sigma_i^2 = \sigma_{\text{syst},i}^2 + \sigma_{\text{stat},i}^2\).

3. Uncertainty-Aware Regression Criteria: Mitigating Noisy Labels and Spurious Deviations

Standard Mean Squared Error (MSE) treats all target labels uniformly, causing deep networks to overfit high-noise samples. Two uncertainty-guided training losses are formulated: - Weighted MSE (wMSE) scales squared errors by the inverse target variance, attenuating the gradients of unreliable labels:

\[\mathcal{L}_{\text{wMSE}} = \frac{1}{\sum_{i=1}^N \frac{1}{\sigma_i^2}} \sum_{i=1}^N \frac{1}{\sigma_i^2} (\hat{h}_i - h_i)^2\]

• Thresholded MSE (tMSE) filters out noise by computing the loss exclusively over samples with absolute \(z\)-score \(|z| > 1.65\) (representing deviations that have less than a 10% probability of occurring under a null no-change hypothesis).

4. Multi-Scale Temporal Adaptation Architecture: LTAE Fusion and Residual Refinement

To evaluate heterogeneous GFMs with diverse input channel specifications within the PANGAEA benchmark framework, encoders remain entirely frozen. For models lacking native multi-temporal capability, multi-year PlanetScope feature representations are fused via a Lightweight Temporal Attention Encoder (LTAE) using temporal position encodings. The resulting multi-scale feature pyramids feed a trainable UPerNet decoder, followed by bilinear upsampling and a residual refinement block for 3 m dense prediction.

Loss & Training

Models are trained for 50 epochs on the 592 training tiles of CHC using the AdamW optimizer with learning rate warmup and cosine annealing. A 3 m vegetation mask eroded by one pixel prevents boundary artifacts. Performance on the 269 test tiles is evaluated on statistically significant change pixels (\(|z| > 1.65\)) via RMSE, MAE, nMAE, and \(R^2\).

Key Experimental Results

Main Results

Evaluation of fully supervised baselines, the task-specific ECHOSAT model, and 9 state-of-the-art Geospatial Foundation Models across three loss formulations on the CHC test set (Table 2 from the paper):

Encoder	Best Loss Config	Overall RMSE (m) ↓	Overall MAE (m) ↓	Overall nMAE ↓	Overall \(R^2\) ↑
ECHOSAT (Domain-Specific)	-	5.25	3.70	0.67	0.55
Supervised UNet (End-to-End)	tMSE (t)	5.65	3.89	0.70	0.48
Prithvi-EO 2.0	wMSE (w)	5.90	4.32	0.78	0.43
DINOv3	tMSE (t)	5.95	4.20	0.76	0.42
DOFA	tMSE (t)	5.99	4.14	0.75	0.41
SatlasNet	wMSE (w)	6.43	4.63	0.83	0.33
CROMA	tMSE (t)	6.61	4.34	0.78	0.29
Supervised ResNet-50	tMSE (t)	6.98	4.55	0.82	0.21
RemoteCLIP	tMSE (t)	7.12	5.10	0.92	0.17
TerraMind	sMSE (s)	7.17	5.14	0.93	0.16
SpectralGPT+	sMSE (s)	7.23	5.27	0.95	0.15
Scale-MAE	tMSE (t)	7.83	5.59	1.01	0.00

Ablation Study: Direction of Change

Model capability in detecting the sign of height change (increase versus decrease) evaluated on the test set (Table 3 from the paper):

Encoder	Loss Config (s/t/w)	Increase Prec.	Increase Recall	Increase F1 ↑	Decrease Prec.	Decrease Recall	Decrease F1 ↑
Supervised UNet	t	0.85	0.80	0.83	0.60	0.69	0.64
ECHOSAT	-	0.83	0.85	0.84	0.63	0.59	0.61
Prithvi-EO 2.0	w	0.84	0.79	0.81	0.57	0.65	0.61
DINOv3	t	0.83	0.76	0.80	0.55	0.65	0.59
DOFA	t	0.83	0.77	0.79	0.54	0.63	0.58
CROMA	t	0.77	0.96	0.85	0.79	0.33	0.47
Supervised ResNet-50	t	0.76	0.94	0.84	0.70	0.34	0.45
Scale-MAE	t	0.69	0.00	0.00	0.31	1.00	0.47

Key Findings

• GFMs struggle to beat supervised and specialized baselines: None of the frozen foundation models outperformed an end-to-end trained UNet (RMSE 5.65 m) or the specialized ECHOSAT model (RMSE 5.25 m). This highlights that self-supervised pretraining objectives often fail to retain the fine-grained geometric and radiometric cues required for dense physical regression.
• Uncertainty weighting delivers consistent gains: Incorporating pixel-level reference uncertainty markedly improves regression accuracy. Switching from standard MSE to tMSE reduced RMSE by 9% for the supervised UNet and 4% for DOFA, while wMSE reduced RMSE by 8% for Prithvi-EO 2.0.
• Divergence between direction classification and magnitude regression: While models achieve competitive F1 scores in identifying whether height increased (CROMA achieves 0.85 F1), precisely quantifying continuous magnitude in meters remains challenging across all architectures.

Highlights & Insights

• First continuous canopy height change benchmark with pixel uncertainty: Unlike existing binary datasets, CHC models real-world vegetation dynamics at 3 m resolution across 10,598 km² alongside rigorous physical error quantification.
• Decoupled LiDAR error modeling: Decomposes elevation errors into static road-derived systematic calibration and pulse-thinning simulation-derived heteroscedastic sampling variance via Gaussian NLL.
• Critical reality check for Geospatial Foundation Models: Demonstrates that impressive performance on classification benchmarks does not readily transfer to sub-pixel dense physical regression tasks.

Limitations & Future Work

• Optical phenological and shadow confounders: Inter-annual variation in sun angle, soil moisture, and phenological timing across PlanetScope acquisitions can mimic or obscure subtle canopy height differences.
• Signal-to-noise limits for incremental growth: Over 5 years, natural tree growth (often \(<1\,\text{m}\)) can approach the uncertainty envelope of low-density ALS acquisitions, constraining the dynamic range of \(R^2\).
• Future directions: Full end-to-end fine-tuning of GFM backbones; multi-modal integration with Synthetic Aperture Radar (SAR, e.g., Sentinel-1) to capture 3D structural backscatter; and physical growth constraint modeling.

Related Work & Insights

• vs OpenCanopy: OpenCanopy provides 1.5 m resolution data but simplifies temporal dynamics to a binary clearing task (\(>15\,\text{m}\) loss over \(>200\,\text{m}^2\)), whereas CHC provides bidirectional continuous metric height changes.
• vs ECHOSAT: ECHOSAT performs global 10 m multi-temporal regression using GEDI spaceborne LiDAR and optical/radar data, whereas CHC focuses on high-resolution (3 m) wall-to-wall ALS change with per-pixel simulated uncertainties.
• vs Binary Change Detection (BCD): BCD only classifies categorical transitions, while CHC poses a heteroscedastic regression problem that models physical growth and loss magnitudes.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First public benchmark framing continuous canopy height change with explicit pixel-wise uncertainty.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous benchmarking across 9 GFMs, multiple baselines, and uncertainty-aware loss functions.
• Writing Quality: ⭐⭐⭐⭐⭐ Clear methodology, rigorous physical error formulation, and well-structured empirical analyses.
• Value: ⭐⭐⭐⭐⭐ Foundational benchmark for high-resolution forest carbon accounting and remote sensing foundation model assessment.

Related Papers

• [ECCV 2026] Estimating Individual Tree Height and Species from UAV Imagery
• [ECCV 2026] Beyond Attention: Convolutional Global Context for Remote Sensing Change Detection
• [ECCV 2026] Beyond 2D Matching: A Unified Single-Stage Framework for Geometry-Aware Cross-View Object Geo-Localization
• [ICLR 2026] TAMMs: Change Understanding and Forecasting in Satellite Image Time Series with Temporal-Aware Multimodal Models
• [ECCV 2026] CRD-Net: Frequency-Adaptive Feature Injection and Change Decoupling for Building Damage Assessment