Forest age estimation in northern Arkhangelsk region based on machine learning pipeline on Sentinel-2 and auxiliary data

Alina Smolina, Svetlana Illarionova, Dmitrii Shadrin, Alexander Kedrov, Evgeny BurnaevScientific 2023 · DOI

Abstract

Tree age is one of the key characteristics of a forest, along with tree species and height. It affects management decisions of forest owners and allows researchers to analyze environmental characteristics in support of sustainable development. Although forest age is of primary significance, it can be unknown for remote areas and large territories. Currently, remote sensing (RS) data supports rapid information gathering for wide territories. To automate RS data processing and estimate forest characteristics, machine learning (ML) approaches are applied. Although there are different data sources that can be used as features in ML models, there is no unified strategy on how to prepare a dataset and define a training task to estimate forest age. Therefore, in this work, we aim to conduct a comprehensive study on forest age estimation using remote sensing observations of the Sentinel-2 satellite and two ML-based approaches for forestry inventory data, namely stand-based and pixel-based. We chose the CatBoost algorithm to assess these two approaches. To establish the robustness of the pipeline, an in-depth analysis is conducted, embracing diverse scenarios incorporating dominant species information, tree height, Digital Elevation Model (DEM), and vegetation indices. We performed experiments on forests in the northern Arkhangelsk region and obtained the best Mean Absolute Error (MAE) result of 7 years in the case of the stand-based approach and 6 years in the case of the pixel-based approach. These results are achieved for all available input data such as spectral satellites bands, vegetation indices, and auxiliary forest characteristics (dominant species and height). However, when only spectral bands are used, the MAE metric is the same both for per-pixel and per-stand approaches and equals 11 years. It was also shown that, despite high correlation between forest age and height, only height can not be used for accurate age estimation: the MAE increases to 18 and 26 years for per-pixel and per-stand approaches, respectively. The conducted study might be useful for further investigation of forest ecosystems through remote sensing observations.

Introduction

Forest age is a critical metric for sustainable management, carbon stock estimation, and ecological health monitoring. However, gathering this data through traditional field studies is labor-intensive, expensive, and nearly impossible for remote or vast territories. While remote sensing (RS) offers a solution for large-scale monitoring, there is no "gold standard" for how to translate satellite data into accurate age estimates. Most existing methods rely on intermediate variables (like tree height) to infer age, but these relationships are often non-linear and vary wildly by species and environment. This paper establishes a robust machine learning pipeline to determine which data features and training strategies yield the most reliable results.

Methodology and Datasets

The study utilized the CatBoost algorithm, a high-performance gradient boosting framework, to test two primary spatial strategies:

  1. Stand-based approach: Aggregating pixel statistics (min, max, mean, std) within known forest inventory boundaries (polygons) to create training samples.
  2. Pixel-based approach: Treating every 10-meter resolution pixel as an independent data point.
Forest stand boundaries overlaid on satellite imagery, illustrating the spatial units used for the stand-based approach.

The dataset integrated multispectral imagery from Sentinel-2 (Level-2A) with ground-truth forest inventory data from the Arkhangelsk region, covering 12,617 forest stands with an average size of 10 ha. The feature set included 10 spectral bands, vegetation indices (NDVI, NDWI, SAVI, ARVI, and EVI), Digital Elevation Models (DEM) for topography, and auxiliary data for tree height and dominant species.

Key Insights: The "Height Fallacy"

A significant finding of this research is the debunking of the "Height Fallacy." While foresters traditionally use height as a proxy for age, this study found that height alone is a poor predictor. Using only height resulted in a high MAE of 18–26 years. The relationship between height and age becomes increasingly non-linear as trees mature, particularly after the 100-year mark.

Scatter plot showing the relationship between forest age and height for mixed species, highlighting the non-linear correlation (R2=0.454).

Furthermore, the B08 (Near-Infrared) band was identified as the most critical spectral feature for age prediction. The study also demonstrated that training individual models for specific tree species (e.g., a "Spruce-only" model) significantly outperformed a single "one-size-fits-all" model, suggesting that the spectral signature of aging is species-dependent.

Results and Performance

The pixel-based approach achieved the highest accuracy with an MAE of 6 years when incorporating all available data (spectral + height + species + DEM). When restricted to only Sentinel-2 spectral bands, the MAE increased to 11 years for both spatial approaches.

In terms of generalization, models trained on specific dates performed well across the growing season, though accuracy slightly decreased for late-summer images (late August) due to changing foliage characteristics. The stand-to-stand approach yielded an average $R^2$ of 0.91, while the pixel-to-pixel approach averaged 0.79.

Conclusion

This study demonstrates that machine learning pipelines using Sentinel-2 data can accurately estimate forest age, especially when augmented with species and height data. While the stand-based approach is highly accurate, it requires pre-existing boundary maps. The pixel-based approach offers a versatile alternative for unmapped territories. Future work should focus on estimating height and species directly from remote sensing data to reduce dependency on manual inventory data and explore deep learning architectures to capture more complex spatial patterns.

Summary updated

Abstract

Tree age is one of the key characteristics of a forest, along with tree species and height. It affects management decisions of forest owners and allows researchers to analyze environmental characteristics in support of sustainable development. Although forest age is of primary significance, it can be unknown for remote areas and large territories. Currently, remote sensing (RS) data supports rapid information gathering for wide territories. To automate RS data processing and estimate forest characteristics, machine learning (ML) approaches are applied. Although there are different data sources that can be used as features in ML models, there is no unified strategy on how to prepare a dataset and define a training task to estimate forest age. Therefore, in this work, we aim to conduct a comprehensive study on forest age estimation using remote sensing observations of the Sentinel-2 satellite and two ML-based approaches for forestry inventory data, namely stand-based and pixel-based. We chose the CatBoost algorithm to assess these two approaches. To establish the robustness of the pipeline, an in-depth analysis is conducted, embracing diverse scenarios incorporating dominant species information, tree height, Digital Elevation Model (DEM), and vegetation indices. We performed experiments on forests in the northern Arkhangelsk region and obtained the best Mean Absolute Error (MAE) result of 7 years in the case of the stand-based approach and 6 years in the case of the pixel-based approach. These results are achieved for all available input data such as spectral satellites bands, vegetation indices, and auxiliary forest characteristics (dominant species and height). However, when only spectral bands are used, the MAE metric is the same both for per-pixel and per-stand approaches and equals 11 years. It was also shown that, despite high correlation between forest age and height, only height can not be used for accurate age estimation: the MAE increases to 18 and 26 years for per-pixel and per-stand approaches, respectively. The conducted study might be useful for further investigation of forest ecosystems through remote sensing observations.
BibTeX
@inproceedings{smolina2023forest,
  author = {Alina Smolina and Svetlana Illarionova and Dmitrii Shadrin and Alexander Kedrov and Evgeny Burnaev},
  title = {{Forest age estimation in northern Arkhangelsk region based on machine learning pipeline on Sentinel-2 and auxiliary data}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-49207-w},
}
Figure