🤖 machine learning

ECHOSAT: Estimating Canopy Height Over Space And Time

The paper introduces ECHOSAT, a global, 10-meter resolution, multi-year tree canopy height map generated using a specialized vision transformer and self-supervised growth loss to accurately capture temporal forest dynamics for improved carbon accounting and disturbance assessment.

Original authors: Jan Pauls, Karsten Schrödter, Sven Ligensa, Martin Schwartz, Berkant Turan, Max Zimmer, Sassan Saatchi, Sebastian Pokutta, Philippe Ciais, Fabian Gieseke

Published 2026-03-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jan Pauls, Karsten Schrödter, Sven Ligensa, Martin Schwartz, Berkant Turan, Max Zimmer, Sassan Saatchi, Sebastian Pokutta, Philippe Ciais, Fabian Gieseke

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand the life story of a forest by looking at a single, frozen photograph. You can see the trees, but you can't tell if they are growing, if a storm knocked some down last year, or if a fire swept through. For a long time, that's exactly what scientists had: static maps of the world's forests that showed height at one specific moment, but nothing about how those forests change over time.

Enter ECHOSAT. Think of ECHOSAT not as a photograph, but as a high-definition, 7-year-long movie of the entire world's forests, where every single frame is incredibly detailed (10 meters per pixel).

Here is the simple breakdown of how they made this movie and why it matters.

1. The Problem: The "Snapshot" Trap

Existing maps are like a yearbook photo. They tell you how tall the trees were in 2020, but they don't tell you if those trees grew an inch or lost a branch in 2021. To calculate how much carbon forests are absorbing (which is crucial for fighting climate change), we need to know the change, not just the current state.

2. The Ingredients: A Multi-Sensor Smoothie

To build this movie, the team didn't just use one camera. They blended data from three different types of "eyes" in space:

Optical Eyes (Sentinel-2): Like a standard camera taking photos in visible light. Great for seeing colors and shapes, but useless when it's cloudy.
Radar Eyes (Sentinel-1 & ALOS): Like a bat using sonar. It sends out radio waves that bounce off trees. It works day or night, rain or shine.
The "Ruler" (GEDI): This is the most important ingredient. It's a laser scanner on the International Space Station that shoots a laser beam down to the ground. It measures the exact height of the trees it hits, acting as the "truth" or the teacher for the AI.

3. The Brain: A Time-Traveling Vision Transformer

The team built a special AI brain called a Temporal-Swin-Unet.

The Analogy: Imagine a student trying to learn how trees grow. If you show them a picture of a sapling and then a picture of a giant oak 50 years later, they might guess the growth. But if you show them a video of the tree growing year by year, they understand the process.
The Innovation: Most AI models look at one year at a time. ECHOSAT looks at the whole 7-year sequence at once. It learns the "rules" of nature: trees usually grow slowly and steadily, but if a fire hits, they drop in height instantly.

4. The Secret Sauce: The "Growth Loss"

This is the paper's biggest trick. In machine learning, you usually teach the AI by showing it the right answer (the GEDI laser measurements). But GEDI only measures a tiny dot of the forest (like a few pixels out of a million). The AI has to guess the rest.

If you just let the AI guess, it might get confused and say a tree grew 10 meters one year and shrank 10 meters the next, just because of noise. That's physically impossible.

So, the team invented a "Growth Loss" rule.

The Analogy: Think of a strict gym coach. If the AI predicts a tree grew 10 meters in a year, the coach slaps its hand and says, "No! Trees can only grow about 3 meters a year max. Try again."
The Twist: If the AI predicts a tree suddenly shrank because of a fire (a "disturbance"), the coach says, "Okay, that's allowed! Nature is chaotic."
This forces the AI to learn realistic physics. It learns that trees generally go up, but sometimes they crash down.

5. The Result: A Global Forest Movie

The result is a map that covers the whole globe, updated every year from 2018 to 2024.

Accuracy: It is more accurate than previous maps, even when looking at just a single year.
Detail: It can spot small changes, like a single tree dying or a small patch of forest being cut down, which older, blurry maps missed.
Insight: It shows that in places like the French Landes forest (where trees are farmed for wood), the height goes up and down rapidly due to logging. In the Amazon, it stays mostly stable, showing a healthy, slow-growing ecosystem.

Why Should You Care?

Forests are the Earth's lungs. To know if they are helping us fight climate change, we need to know if they are getting bigger (absorbing more carbon) or smaller (releasing carbon).

Before ECHOSAT, we were guessing based on old photos. Now, we have a living, breathing map that tracks the heartbeat of the planet's forests. It helps governments and scientists make better decisions about protecting nature and managing our carbon footprint.

In short: ECHOSAT turned a static photo album of the world's trees into a dynamic, physics-aware movie, allowing us to finally see how our forests grow, struggle, and recover over time.

1. Problem Statement

Forest monitoring is critical for climate change mitigation, as forests absorb approximately 3.5 Pg of carbon annually. However, existing global tree height maps suffer from two major limitations:

Static Nature: Current maps provide only single-time snapshots (e.g., a specific year) and cannot capture temporal dynamics like tree growth, degradation, or recovery.
Lack of Temporal Consistency: Previous attempts to create time-series maps often rely on training independent models for each year or applying extensive post-processing to smooth inconsistencies. This leads to unrealistic pixel-level fluctuations and fails to capture small-scale disturbances (e.g., selective logging, individual tree mortality) or natural growth curves.

The core challenge is to generate a globally consistent, high-resolution (10m) tree height time-series that inherently learns realistic temporal dynamics (gradual growth and abrupt disturbances) without relying on post-processing.

2. Methodology

The authors propose ECHOSAT, a deep learning framework that combines multi-sensor satellite data with a specialized architecture and a novel self-supervised loss function.

A. Data Integration

The model utilizes a multi-temporal dataset spanning 2018–2024 (7 years, 84 monthly timesteps):

Optical: Sentinel-2 (12 spectral bands, 10m resolution).
Radar: Sentinel-1 (C-band, VH polarization) and ALOS PALSAR-2 (L-band, HH/HV).
Auxiliary: TanDEM-X DEM and forest classification maps.
Ground Truth: GEDI (Global Ecosystem Dynamics Investigation) LiDAR data serves as sparse labels. Due to the sparse nature of GEDI (25m footprints), the model must generalize from limited points to a global 10m grid.

B. Model Architecture: Temporal-Swin-Unet

The core architecture is a modification of the Swin Transformer, adapted for spatio-temporal regression:

Base: Combines Video Swin Transformer (for temporal processing) and Swin-Unet (for dense prediction).
Input: A tensor of shape $18 \times 84 \times 96 \times 96$ (Channels $\times$ Time $\times$ Height $\times$ Width).
Patch Size: Unlike standard Vision Transformers that use $4\times4$ or $16\times16$ patches, ECHOSAT uses a $1\times1$ pixel patch size. This is crucial for preserving fine-grained spatial details (e.g., forest boundaries) at 10m resolution.
Temporal Downsampling (TD): Encoder layers reduce the temporal dimension via linear projection and concatenation of adjacent years.
Temporal Skip Connection (TSC): Decoder layers enrich features using corresponding encoder features from the same year via a Transformer attention mechanism, ensuring temporal coherence.
Dual Heads:
1. Reference Head: Pre-trained using Huber loss to learn a baseline representation.
2. Prediction Head: Fine-tuned using the novel Growth Loss to output the final time-series.

C. Novel Contribution: Growth Loss

To enforce physically realistic growth patterns without post-processing, the authors introduce a self-supervised Growth Loss:

Disturbance Detection: The model identifies "disturbance years" where tree height drops by $>50\%$ and $>4$ m, or falls below 10m within two years.
Constrained Linear Regression: For each pixel, the model fits a piecewise linear regression to the reference output:
- Pre-disturbance: A linear growth curve.
- Post-disturbance: A linear regrowth curve.
- Constraints: Slopes are bounded within realistic biological limits (e.g., $0$ to $3$ m/year).
Pseudo-Labeling: These fitted curves serve as "pseudo-labels" for the prediction head.
Loss Function: The model minimizes the distance between its predictions and these physically constrained pseudo-labels. This forces the network to learn monotonic growth and abrupt declines naturally.

3. Key Contributions

First Global Spatio-Temporal Map: ECHOSAT is the first high-resolution (10m) global tree height map covering seven years (2018–2024), enabling the monitoring of forest dynamics at an unprecedented scale.
Inherent Temporal Learning: Unlike previous methods requiring post-processing, ECHOSAT uses a specialized Growth Loss to inherently learn realistic growth curves and disturbance events during training.
Architecture Innovation: The Temporal-Swin-Unet with $1\times1$ patching and specific temporal skip connections allows for pixel-level regression that maintains both spatial detail and temporal consistency.
Superior Performance: The model outperforms existing state-of-the-art single-year maps in accuracy and reduces prediction variance.

4. Results and Evaluation

The model was evaluated against GEDI LiDAR labels (filtered for quality) and compared against four existing global height maps (Potapov et al., Lang et al., Pauls et al., Tolan et al.).

Accuracy (2020 Benchmark):
- MAE: 5.85 m (vs. 6.85–11.89 m for baselines).
- RMSE: 10.87 m (vs. 11.76–16.13 m for baselines).
- $R^2$ : 0.59 (vs. 0.45–0.52 for baselines).
- ECHOSAT consistently outperformed all baselines across all metrics (MAE, MSE, RMSE, MAPE, $R^2$ ).
Temporal Dynamics:
- Growth Patterns: The model correctly captured that shorter/younger trees grow faster than taller ones.
- Disturbances: It successfully identified abrupt height declines (e.g., logging in Le Landes, France) and subsequent regrowth, while maintaining stability in undisturbed regions (e.g., Amazon rainforest).
- Consistency: The model produced smooth, realistic time-series without the "flickering" or unrealistic fluctuations seen in independent single-year models.
Limitations:
- Performance degrades in very dense clouds (tropical rainforests) where GEDI signals fail to penetrate.
- Small annual growth increments (<1m) are difficult to distinguish from noise.
- Validation of year-to-year changes is limited by the lack of independent, high-frequency ground truth (GEDI is sparse).

5. Significance

ECHOSAT represents a significant leap forward in global carbon accounting and forest management:

Carbon Monitoring: By providing accurate, time-resolved height data, it allows for better estimation of Above-Ground Biomass (AGB) changes, crucial for tracking carbon sinks and sources.
Disturbance Assessment: It can detect small-scale degradation and selective logging that were previously invisible to coarse-resolution or static maps.
Policy Impact: The data supports evidence-based climate policy and forest management by offering a dynamic view of forest health rather than a static snapshot.

The maps and code are made available to the public to advance global efforts in climate change mitigation.