Supervised machine learning of fullcube hyperspectral data

Kurzfassung

This thesis is part of multiple studies aimed at using remote sensing technologies for monitoring the changes in the Bavarian Forest National Park (48°580 N 13°230 E), which has an overall area of 24.369 hectares. In order to detect changes of biodiversity, additional methods contributing to regular forest inventories were used. One key component of that goal is the tree species classification. The scientific team of the National Park has been working together with the German Aerospace Center conducting multiple airborne campaigns collecting hyperspectral data. The fullcube hyperspectral data (0.4 - 2.498 µm) with a resolution of 3.2 m, acquired in July 2013 is the basis data set for the pixel based supervised machine learning of tree species. Along with two field campaigns and the forest inventory data, 4775 pixels of ground truth data were derived. The spectral data were processed using the established CATENA processing chain developed at DLR, and for further enhancing the predictive capabilities smoothed and brightness normalized. BRDF effects were minimized using the novel approach of BREFCOR included in the ATCOR4 software for correcting atmospheric disturbances of airborne remote sensing data. The classification was carried out using common open source software.1 Additional LiDAR data and a set of vegetation indices were also used as input data. Seven classifiers of extremely randomized trees were trained using different feature combinations. Three levels of predictions were made based on I. species, II. species groups, and III. coniferous / broadleaf trees. The classification accuracy was evaluated using Kappa scores, F1-measurements and confusion matrices. Over fitting was detected as a problem, when using LiDAR based DTM data, because of the small size of available training data and the specific behaviour of random forests. The large number of training pixels, which would be needed for representing the multitude of differences in species distribution over the height above zero was not achieved. The greedy behaviour of the used forest of randomized trees lead to a biased learning behaviour. Apart from comparing machine learning metrics retrieved from the ground truth data, the overall tree species composition of the both parts of the Bavarian Forest National Park was calculated and the northern part was evaluated by comparing predicted results to the latest forest inventory. The fullcube hyperspectral spectrum combined with selected vegetation indices showed an overall better suitability for classifying the selected tree species reaching a kappa score of 0.589 for the test data set. The highest F1-scores were recorded for the species Pinus mugo with 0.88,followed by the species Fagus sylvatica (0.80), Picea abies (0.65) and Fraxinus excelsior (0.64). Difficulties in the classification were observed within the conifers and broadleaved species, rather than between these two groups. The coniferous minority class species Pseudotsuga menziesii (0.14) showed low F1-scores based on high misclassification as Abies alba and Picea abies. While the broadleaved species Acer pseudoplatanus (0.29) showed high misclassification as Fagus sylvatica.