Transfer learning of U-Net convolutional networks for TanDEM-X forest/non-forest mapping at large scales

Kurzfassung

Accurate mapping of forest/non-forest areas supports monitoring, verification and reporting on forest resources. The potential of TanDEM-X coherence (due to volume decorrelation) to distinguish between forest and non-forest classes has been recognized. Growing literature shows that deep learning classifiers are able to automatically learn relevant features in image data and mostly outperform traditional classifiers at image classification tasks. This study assesses the potential of U-Net convolutional networks for TanDEM-X forest/non-forest classification and their transferability in time and space. Trained U-Net convolutional networks are used to discriminate forest/non-forest classes with TanDEM-X data and their potential to transfer on TanDEM-X data collected at different time steps and from different geographical areas is evaluated. First, we trained the networks on 50% of reference TanDEM-X images and evaluated their learning rate and prediction power using the remaining 50% of reference images. Second, transfer of the networks across time as well as geographical area was done after evaluating environmental similarity between the reference and transfer TanDEM-X images using Mahalanobis distances. Finally, model performance was assessed using, overall accuracy (OA) and Area Under Curve (AUC), considering complete transfer scenarios as well as transfer with augmentation training; in which additional training of the networks with parts of transfer images was done. Preliminary results show that the networks classify forest/non-forest areas in reference images with high accuracy in German temperate forests (OA>0.95, AUC > 0.92). We expect that transfer of the networks across time will have marginal influence on model performance compared to transfer of the networks across geographic areas. In general, lower environmental similarity may result in bigger losses in model performance. However, marginal augmentation training (e.g. up to 20% of transfer images) may boost model performance back to the model performance achieved with reference images. This work demonstrates how deep learning algorithms can support applications in forest resource and disturbance mapping at regional to global scales using TanDEM-X data.