📄 earth_science

Spatiotemporal Estimation of Daily Peatland Gross Primary Production with Landsat Remote Sensing Data in Germany, Estonia and Finland

This study demonstrates that a Landsat-based random forest model, trained with eddy covariance data, can effectively estimate daily gross primary production across diverse peatland management regimes in Germany, Estonia, and Finland, thereby enabling consistent large-scale monitoring of peatland productivity under varying climate and land-use conditions.

Original authors: Nur Hussain, Adrià Descals, Alisa Krasnova, Kaido Soosaar, Josep Penuelas, Ülo Mander, Parvez Rana

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nur Hussain, Adrià Descals, Alisa Krasnova, Kaido Soosaar, Josep Penuelas, Ülo Mander, Parvez Rana

Original paper licensed under CC BY 4.0 (https://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

The Big Picture: Weighing the Earth's "Carbon Sponges"

Imagine peatlands (wet, boggy areas full of decaying plants) as giant, ancient carbon sponges. For centuries, these sponges have been soaking up carbon dioxide from the air and storing it underground. This is crucial for keeping our climate stable. However, humans have been draining these sponges to make way for forests or farms, and sometimes we try to "rewet" them to fix the damage.

The big question the researchers asked was: How much carbon are these sponges actually absorbing right now?

To answer this, they needed to measure something called Gross Primary Production (GPP). Think of GPP as the "daily lunch" the plants in the bog eat. It's the total amount of carbon dioxide the plants pull out of the air to grow. If the plants are eating a lot, the sponge is working hard. If they are eating little, the sponge is struggling.

The Problem: Too Big to Measure by Hand

Scientists used to measure this "daily lunch" using tall towers equipped with sensors (called Eddy Covariance towers) that sit right in the middle of the bog. These towers are like super-accurate but very narrow microscopes. They can see exactly what's happening in a small circle around them, but they can't see what's happening a mile away.

Since peatlands are huge and messy (some are wet, some are dry, some have trees, some have just grass), using just a few towers is like trying to understand the weather of an entire country by standing in one backyard.

The Solution: The Satellite "Drone"

The researchers decided to use Landsat satellites as their eyes in the sky. Think of Landsat as a high-definition drone flying overhead that takes pictures of the entire landscape every few days.

The Old Way: They tried using older, lower-resolution satellites (like MODIS), but those were like looking at a landscape through a foggy window. The details were too blurry to see the specific differences between a drained forest and a restored bog.
The New Way: They used Landsat, which is like looking through a crystal-clear window. It has a resolution of 30 meters (about the size of a small house), allowing them to see the specific "texture" of the vegetation.

The Experiment: Testing the "Recipe"

The team looked at four specific "kitchens" (study sites) in Finland, Estonia, and Germany. Each kitchen had a different "chef's style" (management regime):

Natural/Restored: Bogs that are wet and growing back (like a garden left to grow wild).
Drained/Managed: Bogs where water was removed to grow trees (like a manicured lawn).
Clear-Cut: Areas where all trees were chopped down (like a freshly mowed field).
Partial-Cut: Areas where only some trees were removed (like a forest with gaps).

The Recipe:
They took the "lunch" data from the ground towers (the truth) and combined it with the satellite photos. They fed this information into a Random Forest model.

Analogy: Imagine a super-smart detective (the AI model). The detective is shown thousands of photos of the bogs and told, "On days when the plants ate this much lunch, the satellite photo looked this specific way."
The detective learns the pattern: "Okay, when the near-infrared light bounces back strongly and the ground feels cool, the plants are eating a big lunch."

What They Found

The "detective" (the model) turned out to be quite good at its job.

It Works Across the Board: The model successfully predicted how much carbon the plants were eating in all four countries, regardless of whether the bog was natural, drained, or restored.
The "Secret Ingredients": The model figured out that two specific clues from the satellite were the most important:
- Near-Infrared Light (NIR): This is like a health check. Healthy, green plants reflect a lot of this invisible light. It tells the model how "leafy" and dense the vegetation is.
- Land Surface Temperature (LST): This is like a thermometer. It tells the model how hot or cold the ground is, which affects how fast the plants can "eat."
Different Management, Different Patterns:
- Clear-Cut Sites: These showed very clear, strong peaks in "eating" (productivity). Because the trees were gone, the sun hit the ground directly, and the new grass/shrubs grew fast and uniformly. The satellite could easily see this "uniform feast."
- Partial-Cut Sites: These were messier. Because some trees were left standing, the sunlight was patchy, and the ground was a mix of shade and sun. This made the "eating" pattern more complex and harder for the satellite to predict perfectly, but the model still did a decent job.

The Verdict

The paper concludes that satellites can act as a reliable "remote chef's assistant." By using Landsat images and a smart computer model, we can now estimate how much carbon peatlands are absorbing across large, messy landscapes without needing a tower in every single spot.

Accuracy: The model's predictions matched the ground measurements about 57% of the time (a solid score for such a complex system), with some sites matching even better (up to 69%).
Limitation: The model isn't perfect. It sometimes struggled in very specific, complex spots (like the Agali-II site in Estonia or the German sites), likely because the ground conditions were too tricky for the satellite to see clearly through the "fog" of vegetation and water.

In short: The researchers proved that we can use high-resolution satellite photos to "see" how well different types of managed peatlands are working as carbon sponges, filling a major gap in our ability to monitor these vital ecosystems.

Technical Summary: Spatiotemporal Estimation of Daily Peatland Gross Primary Production with Landsat Remote Sensing Data

Problem Statement
Peatlands are critical terrestrial carbon (C) reservoirs, storing approximately one-third of the world's soil carbon despite covering only 3% of the global land surface. However, their role as carbon sinks is increasingly compromised by climate warming and anthropogenic disturbances, particularly drainage and forest management practices in boreal and hemiboreal regions. While eddy covariance (EC) flux towers provide high-frequency, continuous measurements of net carbon exchange and derived Gross Primary Production (GPP), their spatial footprint is limited (typically 100–500 m), making it difficult to capture the heterogeneity of peatland landscapes influenced by micro-topography, vegetation composition, and hydrological gradients. Conversely, existing satellite-based GPP products, such as MODIS (1 km resolution), often fail to accurately represent fine-scale variability in complex peatland ecosystems due to pixel heterogeneity and coarse resolution. There is a significant observational gap in quantifying the spatiotemporal dynamics of GPP across contrasting peatland management regimes (natural, drained, restored, clear-cut, and partial-cut) at landscape scales.

Methodology
To address these limitations, the authors developed a machine learning framework integrating Landsat remote sensing data with in-situ EC flux observations.

Study Sites: The study utilized four case study sites across three countries representing diverse management regimes: Lettosuo (Finland, drained forestry with clear-cut and partial-cut treatments), Agali-II (Estonia, drained forestry), and Anklam and Zarnekow (Germany, representing agricultural conversion and rewetted/restored fen conditions).
Data Sources:
- In-situ Data: Daily total GPP (GPPFLUX) was derived from gap-filled and partitioned EC flux tower data, aggregated to annual cumulative values for analysis periods ranging from 2009 to 2024 depending on the site.
- Remote Sensing Data: Landsat-5, -7, -8, and -9 Level-2 surface reflectance products (Blue, Green, Red, NIR, SWIR1, SWIR2) and Land Surface Temperature (LST) were processed via Google Earth Engine (GEE).
Modeling Approach: A Random Forest (RF) regression model was employed to predict daily GPP.
- Predictors: Spectral bands (Blue, Green, Red, NIR, SWIR1, SWIR2) and LST.
- Training: The model was trained on 70% of the coincident Landsat-EC data and validated on the remaining 30%. Cross-site validation (leave-one-site-out) was performed to test model transferability.
- Variable Importance: Gini importance metrics were used to assess the contribution of each spectral and thermal band to GPP prediction.

Key Results

Model Performance: The Landsat-based RF model demonstrated good agreement with EC-derived GPP observations across all sites. The overall model achieved an $R^2 \approx 0.57$ $R^{2} \approx 0.57$ and an RMSE of approximately 3.0 g C m⁻² d⁻¹.
- Performance varied by site, with the highest accuracy at the Finnish clear-cut site ( $R^2 = 0.72$ ) and the lowest at the Estonian Agali-II site ( $R^2 = 0.29$ ).
- Cross-site validation confirmed the model's robustness and transferability, with $R^2$ values ranging from 0.38 to 0.69 for independent sites.
Variable Importance: Near-Infrared (NIR) reflectance and Land Surface Temperature (LST) were consistently the most important predictors across all sites, highlighting the roles of canopy structure and thermal regulation in GPP. The relative importance of visible (Red) and Short-Wave Infrared (SWIR) bands varied by site and management regime, reflecting differences in soil exposure, peat moisture, and vegetation composition.
Management Regime Dynamics:
- Clear-cut vs. Partial-cut: Clear-cut areas exhibited stronger seasonal peaks and more homogeneous spatial patterns, likely due to simplified canopy structure and increased light penetration to regenerating understory. Partial-cut areas showed reduced seasonal amplitude and greater spatial heterogeneity, attributed to retained canopy cover buffering microclimatic extremes.
- Restoration: The model successfully captured productivity patterns in rewetted sites (DE-Zrk), distinguishing them from drained or natural states.
Spatial Mapping: The study successfully generated spatially explicit GPP maps at 30 m resolution, visually distinguishing productive vegetated zones from less active areas, particularly evident in the Lettosuo site maps.

Significance and Claims
The paper claims that this study provides a scalable, consistent framework for monitoring peatland productivity across diverse management conditions using freely available Landsat data. By integrating multispectral and thermal observations with EC flux data, the approach overcomes the spatial resolution limitations of MODIS and the footprint limitations of flux towers. The findings support the use of satellite-driven approaches for large-scale monitoring of peatland carbon sequestration, offering a tool to assess the impacts of forest management (clear-cutting, partial-cutting) and restoration (rewetting) on carbon uptake. The study emphasizes that while NIR and LST are robust, transferable predictors, the inclusion of visible and SWIR bands is necessary to capture local ecological variability and management-induced spectral changes. This work fills a critical observational gap in managed peatlands, enhancing the capacity to evaluate carbon flux responses to environmental change at landscape-to-regional scales.