Regionalized classification of stand tree species in mountainous forests by fusing advanced classifiers and ecological niche model

ABSTRACT

Though many new remote sensing technologies have been introduced to analyze forests, regional-scale species-level mapping products are still rare, especially in large mountainous areas. Tree species abundance, low spectral separability among species and huge computing demand are hindrances for obtaining an accurate stand tree species map. This study addressed these problems by synergizing regionalization, multiple feature fusion, and model fusion and proposed a new machine learning workflow. The whole area, i.e. Yunnan province in China (approximately 390,000 km²), was firstly divided into 8 distinct floristic regions according to the distributions and phylogenetic relationships of native tree species. Thereafter, with Google Earth Engine (GEE) platform, multiple data sets, including Sentinel-2 imagery, SRTM DEM, and WorldClim bioclimatic, were collected to construct a high-dimensional feature pool for each region. Thirdly, the maximum entropy model （MaxEnt）, generally used for predicting ecological niche, and three classifiers, i.e. Random Forest (RF), Support Vector Machines (SVM), and Extreme Gradient Boosting (XGBoost), were used to pre-classify environmental and remote sensing data, respectively. After that, two types of decision fusion strategies, parallel and serial ensembles, were applied to fuse pre-classification probability maps for each sub-region. Finally, the spatial distribution of 19 forest stand species over the whole Yunnan Province was obtained by mosaicking the best classification results from 8 sub-regions. Our method achieves an overall accuracy of 72.18% on the entire validation dataset. The decision fusion models significantly improve the classification accuracy, with the eight partitioned best fusion models improving the accuracy by 7.33%–25.39% on average compared to base classifiers. This study demonstrates that the spatial partitioning strategy and the decision fusion integrating a proper machine learning algorithm and ecological niche model can significantly improve the classification accuracy of forest stand species in montane forests.

KEYWORDS:

• Tree species classification
• Vegetation regionalization
• Maximum entropy model
• Machine learning
• Decision fusion

Acknowledgments

We thank anonymous reviewers and editors for their valubale comments and suggestions, which greatly improved the quality of our maniucrpt.

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15481603.2023.2211881.

Additional information

Funding

The work was supported by the National Natural Science Foundation of China [32160369]; Major Scientific and Technological Projects of Yunnan Province [202202AD080010]; National Natural Science Foundation of China [31860182]; National Natural Science Foundation of China [32060320]; National Natural Science Foundation of China [41961053].