Comparison of Different Machine-Learning Algorithms for Tree Species Classification Based on Sentinel Data

In this study we use freely available Sentinel-2 data, orthophotos (RGB and CIR) and inventory data (11 or 12m radius circles with a percentage composition of tress within) by the BaySF (Bayerische Staatsforsten) to evaluate the potential of different machine-learning approaches to classify tree species. Our study areas, the “Ebersberger Forest” and the “Freisinger Forest” are both located in Bavaria and show a homogenous intermixture of coniferous and deciduous tree patches, dominated by the spruce tree type. Sentinel-2 satellite imagery offers new opportunities for vegetation mapping, climate change analysis or agricultural applications by monitoring the earth surface in a high temporal resolution (ideally in a five-day interval in Europe) and appropriate spectral resolution. Using the Sen2Cor processor, atmospheric correction was applied to the level 1C data, gaining true surface reflectance values to create a bottom of atmosphere (BOA) output. Reducing atmospheric effects is especially relevant when analyzing multitemporal images. As one of the goals of this study was to deliver a semiautomatic workflow for the classification of beech and oak trees, different classification algorithms (object- and pixel-based) were evaluated and the best performing setup was implemented into a model. A hierarchical approach was used to evaluate different band combinations and algorithms (Support Verctor Machines (SVM) and Random Trees (RT)) for the separation of deciduous vs. coniferous trees, followed by a more detailed evaluation for tree species within the respective classes. The Ebersberger Forest was the main project region of interest, the Freisinger Forest was used as a reference validation region. Training and accuracy assessment of the algorithms was based on inventory data. The validation process was conducted using an independent dataset. A confusion matrix, with a focus on User´s and Producer´s Accuracies per class, as well as Overall Accuracies (OA), were calculated for all analyses. In total, we tested 16 different classification setups for coniferous vs. deciduous trees, achieving the best performance of 97.2% OA for an object-based multitemporal SVM approach using scenes from May, August and September by combining the infrared bands. For the separation of beech and oak trees we evaluated 55 different setups, the best result was reached by an multitemporal SVM classification based on objects with an accuracy of 90.9% (OA) using the May scene for segmentation and the August scene with its principal components as classification image. The transferability of the model was tested for the the Freisinger Forest and showed similar results. These results point out, that S-2 has only marginally worse results than comparable commercial high-resolution satellite sensors, the less spatial resolution is compensated by the high temporal resolution, which supports well the classification of different tree types and is therefore well suited for forest analysis.