Green-spaces, especially trees, influence climate in urban areas in a number of ways and can contribute to climate change mitigation as well as adaptation. Information about this resource is therefore an important basis for decision-making in spatial planning and urban management. While a plethora of remote sensing data is available, in many cases it cannot be used efficiently due to the lack of suitable analysis tools. With regard to image data processing, artificial neural networks, especially convolutional neural networks (CNNs), have become a widely established method during the last decade, boosted by the increased availability of training data and computing power. This work investigates their suitability for the derivation of information on trees in urban areas from remote sensing data.
Multiple CNNs are trained for three different input formats (24x24, 50x50, 100x100 pixels) in order to derive a variety of information, namely tree location, genus, height, age and crown diameter. Digital orthophotos (DOPs) as well as digital surface and elevation models (DSM, DEM) are used as input data. Example data is created using the street tree inventory of the city of Leipzig (Saxony, Germany). The trained models are applied to new data using a sliding window.
The results of this work confirm the great potential of CNNs as generic tools for the analysis of image or raster data shown in previous studies. Upon application to test data, the detection of input images containing visually distinguishable tree crowns is performed with an accuracy of up to 99%. For the classification of tree genera, an overall accuracy of up to 72% is reached, whereas confusion matrices show differences in accuracies for single genera. The remaining target variables are predicted with minimal error values (RMSE) of 9 a for the tree age, 1.8 m for the tree height and 1 m for the crown diameter. As the amount of example data is limited, a strong influence of its composition and quality can be observed.