Knowledge from Imagery
Extracting Species and Habitat Information from the UAV Drone Campaign 2017
RGB drone images provide high-resolution raster graphics in 3 color channels (R-red, G-green, B-blue) with an information depth of 8 bits (gray values ranging from 0 to 255). This information can be used to teach an algorithm to automatically detect different surface properties. For example, it is possible to differentiate green vegetation from soil and deadwood, but also to differentiate vegetation according to different species based on color information. Each pixel is assigned the class of the respective type and displayed on a map. The algorithms that are used to automatically classify species from the color space are self-learning systems- one speaks of machine learning.
In addition to the color information, drone imagery provide digital surface models (DSM / DOM) in a resolution of 20 cm. By intersecting these with a LIDAR elevation model (provided by LGB-Brandenburg) one can calculate a so-called normalized digital surface model (nDSM / nDOM). In this, the actual terrain height is subtracted, so that finally only the height of the object, hence in the case of test field A, the vegetation height, is left. On test field A we found out that Calluna vulgaris reaches heights of 5 to 70cm depending on age.
Three methods for the automatic and quantitative determination of the recorded drone image were developed. As an example, the results for a section of the entire test field (detail red rectangle) are shown. In the first image an automatic tree mask was applied, which automatically detects the edges of larger trees, so that a separation of open land and forest habitats can be made (1st image - green boundary). In the second, we succeeded in separating Calluna vulgaris plants from their environment using a machine learning algorithm (Random Forest). Every single plant is there recognized and identified in the image pixel (2nd image - green). Different age-, blooming- and vitality stages are allocated to the single plants and their spatial distribution was mapped. In the last step, each of the individual Calluna heath plants is assigned the height from the nDOM (3rd image - same continuous scale as above from 0 to 70cm). Together, all 3 algortihms allow to identify each single Calluna plant from a drone image and characterize it by height or color attributes.