Estimating optical vegetation indices and biophysical variables for temperate forests with Sentinel-1 SAR data using machine learning techniques: A case study for Czechia

Overview

• Current vegetation indices (VIs) for monitoring forests have limitations due to atmospheric effects.
• Synthetic aperture radar (SAR) data can provide consistent forest monitoring through clouds.
• This study explores using SAR data and machine learning to estimate optical VIs for forests.

Plain English Explanation

Vegetation indices are measurements that use satellite data to track the health and characteristics of vegetation, like forests. The most common vegetation indices are things like the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI).

However, these vegetation indices can be limited by cloud cover and other atmospheric effects that interfere with the satellite data. In contrast, radar satellites can see through clouds and take images day or night, providing more consistent data.

The researchers in this study wanted to see if they could use the radar satellite data, together with machine learning, to accurately estimate the common vegetation indices for forests. This could provide a way to monitor forests that is less affected by clouds and weather.

The researchers focused on healthy and disturbed temperate forests in Czechia, using data from the European Sentinel-1 radar satellite and other datasets like terrain and weather information. They were able to estimate four key vegetation indices - Leaf Area Index (LAI), Fraction of Absorbed Photosynthetically Active Radiation (FAPAR), EVI, and NDVI - with high accuracy compared to ground truth data.

This SAR-based approach to estimating vegetation indices could provide a powerful way to continuously monitor forest health and changes, even in cloudy conditions. The researchers were able to detect abrupt forest changes on a weekly timescale, which is difficult to achieve with traditional optical satellite data.

Technical Explanation

The researchers created a paired multi-modal time series dataset in Google Earth Engine, including temporally and spatially aligned data from the Sentinel-1 radar satellite, the Sentinel-2 multispectral satellite, a digital elevation model, weather data, and land cover information. This allowed them to train and test machine learning models to estimate four key vegetation indices - LAI, FAPAR, EVI, and NDVI - using only the Sentinel-1 SAR data and auxiliary features.

They compared the performance of traditional machine learning algorithms like Random Forest Regression (RFR) and XGBoost to an AutoML approach (auto-sklearn). The traditional ML models slightly outperformed the AutoML, achieving R^2 values between 70-86% and low mean absolute errors (0.055-0.29) in estimating the target vegetation indices.

The inclusion of terrain and weather features further improved the results. The researchers were able to achieve up to 240 measurements per year at a 20m spatial resolution using the SAR-based vegetation indices, with high accuracy compared to the Sentinel-2 ground truth data. A key advantage is the ability to detect abrupt forest changes on a sub-weekly timescale using the SAR data.

Critical Analysis

The researchers acknowledge that their SAR-based approach to estimating optical vegetation indices is less direct than using the multispectral data itself. However, they successfully demonstrate that the SAR signal contains enough relevant information, when combined with machine learning, to accurately predict these important forest metrics.

One limitation is that the study is focused on a specific region (Czechia) and time period (2021). Further research would be needed to assess the generalizability of the approach to other forest types and geographic areas. Additionally, the researchers did not explore the potential causes or explanations for why the SAR data was able to effectively estimate the optical vegetation indices.

While the results are promising, there may be cases where the direct use of multispectral data is still preferable, especially if the goal is to understand the biophysical mechanisms driving changes in vegetation. The SAR-based approach is more of an indirect proxy that leverages the power of machine learning.

Overall, this research raises interesting scientific questions about the information content of SAR data and the potential to leverage it for forest monitoring in a way that is more resilient to atmospheric conditions. Further work could explore the limits of this approach and investigate the underlying relationships between SAR signatures and optical vegetation indices.

Conclusion

This study demonstrates that synthetic aperture radar (SAR) data, combined with machine learning, can be used to accurately estimate key optical vegetation indices for monitoring forest ecosystems. This provides a promising alternative to relying solely on multispectral satellite data, which can be limited by atmospheric effects like clouds.

The researchers were able to achieve high accuracies in estimating leaf area index, fraction of absorbed photosynthetically active radiation, the Enhanced Vegetation Index, and the Normalized Difference Vegetation Index using Sentinel-1 SAR data and auxiliary features. This approach enables more consistent, high-frequency forest monitoring that is less affected by weather conditions.

While the SAR-based approach is less direct than using multispectral data, it demonstrates the potential to leverage the unique capabilities of radar satellites for environmental applications. Further research could explore the broader applicability of this technique and investigate the underlying relationships between SAR signatures and optical vegetation indices. Overall, this work highlights an innovative way to overcome limitations in optical satellite data using advanced machine learning techniques.