Can historical black and white images be used to map changes of tree-line?

Tree-line shifts are evident signs of various aspects of global climatic changes. Remote sensing techniques and aerial imageries are typically used to assess the changes of tree-line in mountainous areas over the past years mostly based on field survey and repeat photography work. The extraction of tree-lines was either done by visual image interpretation or low spatial resolution satellite images. (Bolton et al., 2018)

In Switzerland, the earliest available historical black and white (B&W) images are from the 1920s, and in the other European countries, there is such abundant data as well. Nevertheless, these data sources are currently insufficiently used and there is only limited use of historical aerial images in the analysis of past vegetation. Therefore their automatic and accurate processing still remains challenging. In our previous studies (Wang et al., 2022) covering parts of the Swiss Alps we obtained promising classification results using a deep learning approach. However, difficulties were related to weak and time consuming labeling efforts. In addition, unclear interclass differences between dense forest and group of trees had a negative effect on model accuracies.

In this study, we proposed a BWForest-Unet based on semantic segmentation to access tree cover in Swiss mountain areas. Images from 2019 and labeled tree images using a countrywide canopy height model (CHM) were used in the model. The main advantage of this net is that features from different spatial regions of the image are combined and thus enabling the localization of more precisely regions of interest. The designed BWForest-Unet tries to learn the spatial interdependencies of features by adding an attention model in the decoder processing. Furthermore, suitable data augmentations, e.g., thickness, local elastic, pinch, scratch and grid distortion were applied as an effective method of supplementing the training samples, which intend to efficiently simulate 1980s images by using current 2019 images. The test area consists of 170 1km*1km sample plots distributed over the whole of Switzerland in 1980s.

The study reveals that 1) suitability of semantic segmentation based on BWForest-Unet in combination with B&W aerial images are superior to previous work and therefore promisingto map mountain tree-line change over 35 years in upper tree-line ecotones of the Alps 2) the usese of existing CHMs substantially reduced the labelling workload. 3) The combination of suitable data augmentations simulates the 1980s image to a certain extent.

 

Bolton, D.K., Coops, N.C., Hermosilla, T., Wulder, M.A., White, J.C., 2018. Evidence of vegetation greening at alpine treeline ecotones: three decades of Landsat spectral trends informed by lidar-derived vertical structure. Environmental Research Letters 13, 084022.

Wang, Z., Ginzler, C., Eben, B., Rehush, N., Waser, L.T., 2022. Assessing Changes in Mountain Treeline Ecotones over 30 Years Using CNNs and Historical Aerial Images. Remote Sensing 14, 2135.