Optimal Input Features for Tree Species Classification in Central Europe Based on Multi-Temporal Sentinel-2 Data

Markus Immitzer, Martin Neuwirth, Sebastian Böck, Harald Brenner, Francesco Vuolo, Clement AtzbergerRemote 2019 · DOI

TL;DR

This study demonstrates that multi-temporal Sentinel-2 satellite data, especially when combined with vegetation indices, significantly improves tree species classification accuracy in Central European forests. By analyzing all possible combinations of 18 cloud-free scenes, the researchers found that a strategic selection of just 6-7 images, particularly from spring to late summer, can achieve classification performance comparable to using all available data.

What Problem Does This Paper Address?

Accurate knowledge of tree species composition is crucial for effective forest management, especially in areas like Central Europe with diverse forest ecosystems. Traditional methods often rely on field surveys or outdated forest management plans, which can be time-consuming, costly, and lack comprehensive spatial coverage. While satellite imagery offers a promising alternative, achieving high accuracy in distinguishing individual tree species, particularly in complex forest landscapes, remains a challenge. This paper addresses the need for a robust, scalable, and cost-effective method for detailed tree species mapping using freely available satellite data, specifically exploring the optimal use of Sentinel-2's multi-temporal capabilities.

Methods

The researchers employed a two-step classification approach. First, they classified the study area into six broad land-cover classes (broadleaf forest, coniferous forest, grassland, cropland, built-up, and water) using a Random Forest (RF) classifier on all 18 Sentinel-2 scenes. Second, they performed separate RF classifications for coniferous and broadleaved forest strata to identify individual tree species. A key aspect of their methodology was exploring the benefit of multi-temporal data by testing 262,143 possible combinations of the 18 Sentinel-2 scenes. To optimize feature selection, they used a recursive feature reduction process based on 'Mean decrease in Accuracy' (MDA). They also investigated the added value of incorporating 28 vegetation indices alongside the raw spectral bands. For change detection, they compared NDVI values from August scenes across three years to identify areas of forest management activity, which were then masked out from the final tree species map.

Datasets

  • Sentinel-2 Data: 18 cloud-free Sentinel-2 acquisitions from June 2015 to the end of 2017 over the Wienerwald biosphere reserve in Austria.
  • Reference Data:
    • For land cover classification: Visually interpreted data from a regular grid using four-band orthoimages (20 cm resolution).
  • For tree species classification: Derived from stand maps and forest management databases, with reference points cross-checked using color-infrared orthoimages. This resulted in an average of 85 reference samples per tree species.

Key Insights and Ideas

  • The Power of Temporal Information: The study strongly emphasizes that the temporal dimension of Sentinel-2 data is critical for accurate tree species classification. Different species exhibit distinct phenological cycles, which are captured by multi-temporal imagery and are essential for differentiation.
  • Strategic Scene Selection is Key: While using all available scenes can yield high accuracy, the research reveals that a carefully selected subset of 6-7 images can achieve nearly identical results. This is a significant finding for operational applications, reducing computational load and data storage requirements.
  • Spring to Late Summer is Crucial: Images acquired between April and August proved most important for achieving high classification accuracy, likely capturing key phenological events like leaf flush and peak greenness.
  • Stratification Matters: Separating the classification into broadleaf and coniferous strata, and then classifying within those strata, generally led to higher overall accuracies compared to a single classification of all species.
  • Vegetation Indices Boost Performance: The inclusion of vegetation indices, particularly those incorporating SWIR, NIR, and Red-Edge bands, consistently improved classification accuracy by approximately 5 percentage points.

Key Results

  • Overall Accuracy: The multi-temporal approach achieved an out-of-bag (OOB) overall accuracy of 85.7% for broadleaved species and 95.3% for coniferous species, a significant improvement over mono-temporal classifications (72.9% and 83.8%, respectively).
  • Optimal Scene Combination: A combination of six to seven scenes achieved model quality as high as using all 18 scenes.
  • Best Performing Species: European Beech and European Larch achieved the highest user's accuracies at 96.3% and 95.9%, respectively.
  • Important Spectral Bands: For broadleaved species, SWIR bands (B11, B12) were most important, while for coniferous species, the Red band (B4) was most crucial. The NIR band (B8) was important for distinguishing between the two main groups.
  • Land Cover Classification: The initial land cover classification achieved a high overall accuracy of 96%.

Limitations and Future Work

  • Class Imbalance: The study acknowledges that class imbalances (uneven numbers of reference samples per species) negatively affect class-specific results, with species having fewer samples achieving lower accuracies.
  • Species with Low Accuracy: Maple, cherry, and European hornbeam were classified with lower accuracies (46–75%), suggesting that their spectral-temporal features are less distinct or harder to capture with Sentinel-2 resolution. Further research is needed to understand why not all species show distinct spectral-temporal features.
  • Resolution Limitations: The 10m resolution of Sentinel-2 data does not allow for individual tree crown segmentation, meaning the classification of mixed classes remains a challenge.
  • Cloud Cover: While 18 cloud-free scenes were used, cloud cover can still be a significant issue, especially in certain seasons and regions, limiting the availability of useful data.
  • Future Work: The authors suggest increasing the number of tree species involved in classifications, developing methods to handle class imbalances and mixed classes, and improving temporal revisit frequency of optical sensors, potentially through virtual constellations. They also call for more research into the bio-physical factors driving species-specific classification accuracies.

Summary updated

Abstract

Detailed knowledge about tree species composition is of great importance for forest management. The two identical European Space Agency (ESA) Sentinel-2 (S2) satellites provide data with unprecedented spectral, spatial and temporal resolution. Here, we investigated the potential benefits of using high temporal resolution data for classification of five coniferous and seven broadleaved tree species in a diverse Central European Forest. To run the classification, 18 cloud-free S2 acquisitions were analyzed in a two-step approach. The available scenes were first used to stratify the study area into six broad land-cover classes. Subsequently, additional classification models were created separately for the coniferous and the broadleaved forest strata. To permit a deeper analytical insight in the benefits of multi-temporal datasets for species identification, classification models were developed taking into account all 262,143 possible permutations of the 18 S2 scenes. Each model was fine-tuned using a stepwise recursive feature reduction. The additional use of vegetation indices improved the model performances by around 5 percentage points. Individual mono-temporal tree species accuracies range from 48.1% (January 2017) to 78.6% (June 2017). Compared to the best mono-temporal results, the multi-temporal analysis approach improves the out-of-bag overall accuracy from 72.9% to 85.7% for the broadleaved and from 83.8% to 95.3% for the coniferous tree species, respectively. Remarkably, a combination of six–seven scenes achieves a model quality equally high as the model based on all data; images from April until August proved most important. The classes European Beech and European Larch attain the highest user’s accuracies of 96.3% and 95.9%, respectively. The most important spectral variables to distinguish between tree species are located in the Red (coniferous) and short wave infrared (SWIR) bands (broadleaved), respectively. Overall, the study highlights the high potential of multi-temporal S2 data for species-level classifications in Central European forests.
BibTeX
@inproceedings{immitzer2019optimal,
  author = {Markus Immitzer and Martin Neuwirth and Sebastian Böck and Harald Brenner and Francesco Vuolo and Clement Atzberger},
  title = {{Optimal Input Features for Tree Species Classification in Central Europe Based on Multi-Temporal Sentinel-2 Data}},
  booktitle = {Remote},
  year = {2019},
  doi = {10.3390/rs11222599},
}
Figure