Work

Evaluating Tree Species Mapping: Probability Sampling Validation of Pure and Mixed Species Classes Using Convolutional Neural Networks and Sentinel-2 Time Series

Year: 2024

Type: article

Abstract: Accurate large-scale classification of tree species is crucial for monitoring, protecting, and managing Earth’s invaluable forest ecosystems. Numerous previous studies have recognized the suitability of satellite imagery, particularly Sentinel-2 imagery, for this task. However, research that specifically addresses the challenges posed by the heterogeneity of forests, including factors such as mixed species stands and stands with sparse canopy cover, as well as research conducted on a large-scale, remains remarkably scarce.In this study, we utilized a dense phenology Sentinel-2 time series, which offered consistent data across multiple granules, to map tree species across the entire forested area in Austria. Alongside eight pure tree species classes (spruce, larch, black pine, white pine, mountain pine, beech, oak, and green alder), as well as a separate class for other deciduous species, we expanded the classification scheme to include mixed tree species and sparsely populated classes (classes with sparse canopy cover), resulting in a total of 25 classes. The incorporation of mixed and sparse classes aimed to enrich the training data, which was curated through visual orthophoto interpretation, to better represent actual forest conditions. Furthermore, synthetic data creation was introduced to enhance the training data.Autocorrelation has significant implications for the validation of thematic maps. To investigate the impact of spatial dependency on validation data, two methods were employed at diverse split and buffer distances: spatial split validation and a validation method based on a buffered ground reference probability samples provided by the National Forest Inventory (NFI). In accordance with the results of this investigation the tree species maps were assessed with a spatial split distance of 4 000 meters and a buffer for the NFI validation data of 1 500 meters. While a random training data holdout set yielded 99 % accuracy, the spatial split validation resulted in 74 % accuracy, emphasizing the importance of accounting for spatial autocorrelation when validating with holdout sets derived from polygon-based training data. The validation based on NFI data resulted in 55 % overall accuracy, 91 % post hoc pure class accuracy, 65 % post hoc mixed class accuracy, and 79 % accuracy when confusions in taxonomic proximity were disregarded (e.g., spruce-larch confused with spruce-beech). The significant decline in accuracy observed between spatial split and NFI validation underscores the disparity between polygon-based training data and ground reference forest data, particularly in areas with diverse forests. This discrepancy is further accentuated by the pure class accuracy of 91 %, revealing the substantial impact of mixed stands on the accuracy of tree species maps.

Source: Remote Sensing

DOI: 10.3390/rs16162887