Hedgerow map of Bavaria, Germany, based on orthophotos and convolutional neural networks

Hedgerows play a significant role in biodiversity preservation, carbon sequestration, soil stability and the ecological integrity of rural landscapes. Understanding their current condition and future development is therefore crucial for a range of stakeholders such as municipalities, state agencies or environmental organizations. The wall-to-wall mapping and characterization of hedgerows in-situ is, however, very labour-, time- and cost-intensive. This impedes a regular monitoring at adequate intervals. In the Federal State of Bavaria, Germany, the hedgerow biotope mapping is repeated every 20-30 years for each district. State-wide consistent and up-to-date data are hence not available. In this study we present an approach for mapping all hedgerows in Bavaria using orthophotos and deep learning. We used hedgerow polygons of the federal in-situ biotope mapping from 5 focus districts as well as additional manually digitized polygons as training and test data and orthophotos as input in a DeepLabV3 Convolutional Neural Network (CNN). The CNN has a Resnet50 Backbone and was optimized using the Dice loss as cost function. The orthophotos were acquired in 2019 – 2021. They have a spatial resolution of 20 cm and were fed to the CNN at tiles of 125 x 125 m. The generated hedgerow probability tiles were post-processed through merging and averaging the overlapping tile boarders, shape simplification and filtering. The resulting hedgerow vector data set achieved medium overall accuracies (precision = 0.43, recall = 0.53, F1-score = 0.48). The model generally overestimated the number of hedgerows, and hedgerows were often confused with riparian as well as urban vegetation. Looking at each hedgerow polygon individually, the mapping accuracy varied considerably, with a median F1-score of 0.51 for all detected objects. In addition, we found differences in accuracies among districts in different landscapes. For example, the Hassberge district, a landscape rich of hedgerows, reached a F1-score of 0.61. A comprehensive comparison with the Copernicus High Resolution Layer (HRL) Small Woody Features (SWF) revealed significant differences between the datasets. About 43 % of the hedgerows in our data set were not present in the SWF layer. Especially narrow, elongated vegetated structures are not captured in the SWF product. This highlights the potential to use our state-wide hedgerow map of Bavaria in combination with the SWF dataset, but also by itself, for a range of administrative, statistical and nature conservation applications. • First EO based state-wide map of hedgerows for Bavaria using deep learning • Approach can map hedgerows effectively and allows short repetition cycles • Accuracies differ among districts within Bavaria • Challenges to distinguish between hedgerows and linear woody vegetation • Significant difference to SWF reveals potential for new usages and joined analyses