Evaluating Tree Species Mapping: Probability Sampling Validation of Pure and Mixed Species Classes Using Convolutional Neural Networks and Sentinel-2 Time Series
The accurate large-scale classification of tree species is crucial for the monitoring, protection, and management of the Earth’s invaluable forest ecosystems. Numerous previous studies have recognized the suitability of satellite imagery, particularly Sentinel-2 imagery, for this task. 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. Aiming for the classification scheme to more accurately represent actual forest conditions, we included mixed tree species and sparsely populated classes (classes with sparse canopy cover) alongside pure tree species classes. To enhance the training data for the mixed and sparse classes, synthetic data creation was employed. 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 numerous 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). 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, and 79% accuracy when confusions in phenological proximity were disregarded (e.g., spruce–larch confused with spruce). The significant differences in accuracy observed between spatial split and NFI validation underscore the challenge for polygon-based training data to capture ground reference forest complexity, particularly in areas with diverse forests. This hardship is further accentuated by the pure class accuracy of 91%, revealing the substantial impact of mixed stands on the accuracy of tree species maps. Keywords: large scale; forest diversity; satellite; ground reference forest data; spatial autocorrelation; synthetic training data
