Identification of tree species using machine learning and phenological characteristics from a 4-year time series of remote sensing data

Accurate identification of tree species is essential to monitor forest resources for climate change mitigation, biodiversity, and forest certification schemes. However, differentiating among species using remotely sensed data is challenging due to the similarity of spectral features. Here, we show that accuracy of species identification can be improved by incorporating phenological information derived from a time-series of multisource remote sensing data. Using the Google Earth Engine platform, we obtained a 4-year (2019-2022) time series of satellite imagery covering multiple phenological periods in mixed-species forests. This dataset was processed using Savitzky Golay filtering and first-order spectral differential transformation identify five dominant tree species through the Forest-Evergreen and Deciduous Forest-Tree Species Hierarchical Classification System (FEDT) with the Random Forest (RF) machine learning algorithm. The integration of phenological data, spectral indices and differential transformations achieved an overall accuracy of 0.82 and kappa coefficient of 0.75, compared to an overall accuracy of 0.76 and kappa coefficient of 0.68, respectively, when using spectral indices alone. Our findings highlight the value of time-series phenological analysis for enhancing the accuracy of tree species identification, providing a scalable method for improved monitoring of forest resources at regional to global scales.