South Korea Improves Forest Greenhouse Gas Reporting with Remote Sensing and Deep Learning

Accurate forest monitoring is critical for managing carbon sinks and ensuring trustworthy national greenhouse gas (GHG) inventories. By combining remote sensing with deep learning, countries can replace broad statistical assumptions with observation-based data that better reflect real forest dynamics. This case study highlights how ENVI® and the ENVI Deep Learning module enabled South Korea to map forest types with greater precision, providing reliable data that directly improves national GHG reporting.

The Challenge

Historically, mixed forests in Korea were reported as 50% conifer and 50% broadleaf, even when one type clearly dominated. While this simplification made reporting easier, it introduced inaccuracies that masked real ecological changes and distorted stock-difference accounting. The Korean University team needed map-ready, observation-based activity data that could be updated annually and aligned with Intergovernmental Panel on Climate Change (IPCC) standards.

Fig. 1. Chiaksan National Park, Gangwon-do, South Korea.

The Solution

Joon Kim and the team at Korea University developed a Phenological Classification Framework (PCF) in ENVI, using Sentinel-2 imagery. They created five seasonal composites per year to capture leaf-on and leaf-off conditions. Then, using ENVI and the ENVI Deep Learning Module, they trained a U-Net model to classify forests as coniferous or broadleaf across the entire country, producing national maps for 2019, 2020, and 2021.

Fig. 2. Study area.

ENVI provided an integrated environment for preparing large satellite image stacks, creating phenology-based composites, and managing training data, all within a single interface. The ENVI Deep Learning module was especially critical, enabling the team to train and apply a U-Net model without writing custom code. This allowed them to perform national-scale classification efficiently on standard workstations and easily repeat the workflow each year.

The team also appreciated ENVI’s flexibility to integrate indices, topographic data, or custom IDL® scripts, offering an adaptable foundation for future research and model refinements.

How They Did It

Using ENVI, the team prepared five seasonal composites per year from Sentinel-2 imagery, capturing phenological variations across seasons. These composites were stacked into 20-layer datasets that served as the foundation for classification.

Fig. 3. Data flow and output process.

With the ENVI Deep Learning module, they trained a U-Net model using forest-type maps from the Korea Forest Service as reference data, achieving a balanced classification between coniferous and broadleaf forests. Once trained, the model was applied nationwide to produce annual forest-type maps for 2019, 2020, and 2021.

The resulting datasets were then incorporated into carbon stock calculations using the IPCC stock-difference method, allowing the team to replace broad statistical estimates with accurate, observation-based data that significantly improved national GHG inventory precision.

The Results

The ENVI-based workflow produced forest-type maps with 83.13 percent overall accuracy and a kappa value of 0.6755 when validated against official reference data. A separate visual inspection using winter imagery showed about 90 percent accuracy.

Most notably, the analysis revealed that South Korea’s mixed forests were not evenly split. In fact, only about 15 percent were conifer-dominant while roughly 85 percent were broadleaf-dominant. Incorporating these observation-based results into IPCC reporting produced carbon-stock trajectories that diverged from prior national estimates, demonstrating the value of ENVI for transparent, data-driven, and policy-relevant GHG reporting.

Fig. 4. Phenological forest-type classification for the entire forest in (a) 2019, (b) 2020, and (c) 2021.

Fig. 5. Phenological spectral reflectance of NIR band based on forest type.

“As the study used satellite imagery specific to South Korea, the ENVI Deep Learning module was employed for training efficiency and convenience. Unlike Python-based training methods, ENVI Deep Learning automatically divides satellite imagery and trains the model, allowing large datasets to be processed even on standard personal computers.”

— Project team, Korea University

Conclusion

By combining Sentinel-2 time series imagery with ENVI and its deep learning module, the Korea University research team developed a scalable, repeatable, and observation-based workflow for national forest classification. The approach replaced long-standing statistical assumptions with accurate, map-based results, revealing that mixed forests are overwhelmingly broadleaf-dominant.

These findings improve the accuracy of carbon stock estimates and demonstrate how ENVI supports transparent, IPCC-aligned data pipelines that strengthen both scientific understanding and policy decision making.

Fig. 6. Annual carbon stock change by forest type.

If you would like to see how your organization can benefit from ENVI and ENVI Deep Learning technology, just email us.

You can read the full case study “Advancing forest GHG inventory accuracy with a phenological classification framework: Toward an observation-based approach 3 in South Korea,” published in Science Direct.