Classification of Tree Species in Poland Using CNNs Tabular-to-Pseudo Image Approach Based on Sentinel-2 Annual Seasonality Data

Abstract

Tree species classification provides invaluable information across various sectors, from forest management to conservation. This task is most commonly performed using remote sensing; however, this method is prone to classification errors, which modern computational approaches aim to minimize. Convolutional neural networks (CNNs) used to model tabular data have recently gained popularity as a highly efficient classification tool. In the present study, a variation of this method is used to classify satellite multispectral data from the Sentinel-2 mission to distinguish between 18 common Polish tree species. The novel model is trained and tested on data from species-homogeneous forest stands. The data form a multi-seasonal time series and cover five years of observations. The model achieved an overall accuracy of 80% and Cohen Kappa of 0.80 of the raw output and increased to 93% with post-processing procedures. Considering the large number of species classified, this is a promising and encouraging result. The presented results indicate the importance of early vegetation season reflectance data in model training. The spectral bands representing the infrared, red-edge and green wavelengths had the greatest impact on the model.