Tree Recognition in Bing Maps’ Aerial and Satellite Imagery with a Convolutional Neural Network

Woody vegetation is a common indicator for the health of biomes because they are an important factor for bio diversity. Furthermore, trees and shrubs are also very crucial for ecosystems due to their ability to provide carbon storage, protection from erosion as well as food for animals and humans. In this thesis a workflow was developed as proof of concept that Convolutional Neural Network (CNN)s are a suitable approach for the segmentation of medium resolution remote sensing Red Green Blue (RGB) data. The aim of the segmentation was to recognize woody vegetation along a geographic vegetation gradient. The workflow includes data acquisition, manual annotation and training of a modified U-Shaped Encoder-Decoder CNN (UNet). As sample area for test and training data, I chose a north-south line of approximately 10° longitude in West Africa, stretching from Mali to Ghana. The trained models showed promising results with a decrease in robustness towards the southern more densely vegetated areas. To quantify the spatial transferability of the CNN model with regard to three test areas, I performed spatial Cross Validation (spatial CV). Due to the large vegetation gradient the results of the spatial CV showed losses of approximately 15 % to 30 % in precision and recall when the model is spatially transferred