TL;DR
This paper presents a novel approach to map and quantify carbon stocks in dynamic agroforestry systems, specifically poplar plantations, by integrating satellite data (Sentinel-1, Sentinel-2) with spaceborne LiDAR (GEDI) using a deep learning U-Net model. This method achieves a high-resolution canopy height model, enabling more accurate estimations of tree height, biomass, and carbon stocks, crucial for climate change mitigation efforts.
What Problem Does This Paper Address?
Agroforestry systems are vital for climate change mitigation and enhancing landscape resilience, but their dynamic nature, especially short-rotation plantations like poplars, poses a challenge for traditional monitoring methods like National Forest Inventories (NFIs). These methods are often too infrequent and labor-intensive for timely updates. Accurately monitoring tree height and associated carbon stocks in these rapidly changing environments is essential for effective management and policy-making. Existing remote sensing approaches have limitations in terms of resolution, cost, or spatial coverage, creating a need for integrated, high-resolution, and freely available data solutions.
Methods
The core of this study is the development of a 10-meter resolution Canopy Height Model (CHM) using a U-Net deep learning architecture. This model takes Sentinel-1 (radar) and Sentinel-2 (optical) satellite imagery as input and uses GEDI waveform data as a reference to predict tree height. The U-Net architecture, known for its effectiveness in image segmentation, was adapted for regression to capture fine-scale canopy details and broader contextual information. After generating the CHM, the researchers used poplar-specific yield tables derived from terrestrial laser scanning to estimate key forestry variables: Diameter at Breast Height (DBH), Growing Stock Volume (GSV), Aboveground Biomass (AGB), and Carbon Stock (CS).
Datasets
- Sentinel-1: Dual-band cross-polarization (VV and VH) GRD scenes were used, preprocessed for radiometric calibration and terrain orthorectification.
- Sentinel-2: Atmospherically corrected surface reflectance images (L2A) were utilized.
- GEDI: Level 2A products, specifically the RH95 (relative height) metric, were used as the reference for training the U-Net model. A large number of GEDI pulses (1,371,490) covering the survey area were downloaded.
- Field Truth Data:
- Italian National Forest Inventory (NFI) plots: Data from 35 circular 530-m² plots were used for validation.
- Traditional forestry field surveys: 60 square plots (50x50m) with measured DBH, tree height, and GSV were used for independent validation.
- Poplar Plantation Maps: Annual maps for 2021 and 2022, derived from semi-automatic classification and refined through photointerpretation, were used to delineate plantation areas.
Key Insights and Ideas
- Synergistic Power of Multi-Sensor Data: The study demonstrates that combining radar (Sentinel-1), optical (Sentinel-2), and LiDAR (GEDI) data through deep learning significantly enhances the accuracy of forest structure mapping compared to using individual sensors or simpler models.
- U-Net for High-Resolution CHM: The U-Net architecture proved highly effective in generating a detailed 10m CHM, outperforming other global CHM products in accuracy for this specific agroforestry context.
- Dynamic Agroforestry Monitoring: The methodology provides a framework for frequent, annual updates of forest variables, which is crucial for managing dynamic systems like short-rotation poplar plantations.
- Bridging the Gap in Validation: By using both NFI data and independent field surveys, the study offers robust validation of its derived forest metrics.
Key Results
- Canopy Height Model (CHM) Accuracy: The U-Net CHM achieved a Mean Absolute Error (MAE) of 2.6 m and an RMSE of 22.6% when compared to NFI data.
- Forestry Variable Accuracy:
- Compared to NFI data: RMSE values of 30.7% for DBH, 46.2% for GSV, and 63.2% for AGB.
- Compared to independent field surveys: RMSE values of 19% for DBH and 37.7% for GSV.
- Carbon Stock Estimates (2021):
- Average GSV: 66.28 m³ ha⁻¹
- Average AGB: 23.83 Mg ha⁻¹
- Average Carbon Stock (CS): 11.92 MgC ha⁻¹
- Harvested Stock Estimates (2021-2022):
- Total harvested GSV: 370,000 m³ (equivalent to 66,000 MgC)
- Average harvested area: 1.5 ha
- Average yield: 130 m³ ha⁻¹
Limitations and Future Work
- Forest Masking: The study acknowledges the lack of a precise, up-to-date forest mask for the entire area, which could have influenced the CHM's accuracy at the boundaries of plantations.
- Temporal Discrepancies: While efforts were made to harmonize data, slight temporal differences between NFI data (2015-2019) and the CHM (2021) could introduce some bias, particularly due to the rapid growth of poplars.
- Generalizability: The study focused on poplar plantations in the Padan Plain. Further research is needed to assess the transferability of this approach to other types of agroforestry systems or forest structures.
- Future Work: The authors suggest that the continuous availability of Sentinel data allows for annual updates of the CHM and subsequent forest variable estimations, enabling effective yearly monitoring of poplar plantation stocks. The integration of such Earth observation data into national forest monitoring systems is highlighted as a key future direction. The potential of upcoming satellite missions like EDGE for even more advanced forest monitoring is also mentioned.
Summary updated