Andri Simeon

Covering over a third of all terrestrial land area, forests are crucial environments; as ecosystems, for farming, and for human leisure. However, they are challenging to access for environmental monitoring, for agricultural uses, and for search and rescue applications. To enter, aerial robots need to fly through dense vegetation, where foliage can be pushed aside, but occluded branches pose critical obstacles. Therefore, we propose pixel-wise depth regression of occluded branches using three different U-Net inspired architectures. Given RGB-D input of trees with partially occluded branches, the models estimate depth values of only the wooden parts of the tree. A large photorealistic simulation dataset comprising around 44K images of nine different tree species is generated, on which the models are trained. Extensive evaluation and analysis of the models on this dataset is shown. To improve network generalization to real-world data, different data augmentation and transformation techniques are performed. The approaches are then also successfully demonstrated on real-world data of broadleaf trees from Swiss temperate forests and a tropical Masoala Rainforest. This work showcases the previously unexplored task of frame-by-frame pixel-based occluded branch depth reconstruction to facilitate robot traversal of forest environments. All models, code, and data are available online.

The destructive power of avalanches demands fast detection and early warning systems to mitigate avalanche risk in mountain areas. Since avalanches are extended moving sources of seismic energy, array systems based on seismic sensors can be used to monitor avalanches over wide areas, regardless of weather and visibility. Operationally, one of the main challenges of these detection systems is the automatic discrimination of avalanche signals from other sources. Various machine learning techniques have, therefore, been used to automatically differentiate avalanche events within seismic recordings. So far, these models mainly relied on expert knowledge and a time-consuming selection process to extract signal attributes from the time and frequency domains. In this study, we therefore investigate the detection performance of models trained with expert-derived features and fully automatic extracted features using autoencoders. Autoencoders are neural networks that efficiently compress (encode) input data, such as seismic signals, down to its  essential features to reconstruct (decode) the original input. Hence, meaningful features can be automatically learnt during this process. To develop the models, we use a catalogue of labelled events recorded by a seismic array system deployed above Davos (Switzerland) since the winter season 2020-2021. We assess the performance of different classifiers developed with data from individual sensors by aggregating the predictions from all array components through ensemble predictions. Our results show that models trained with autoencoder-extracted features performed as well as those relying on expert features and spectrogram images. When tested on a continuous seismic dataset, the autoencoder model based on the frequency content of the signals achieved the best balance between avalanche detection accuracy and false alarm rate. Looking forward, we aim to integrate the best approach into an operational, near-real-time avalanche detection system.

Snow avalanches are the deadliest natural hazards in Switzerland and cause significant economic losses annually. Therefore, avalanche detection and prediction are crucial in ensuring safety and mobility in the Swiss Alps. Operationally, avalanche forecasts are issued daily during the winter season to inform the public and local authorities about the avalanche hazard. Whereas traditional avalanche forecasting has been a human-expert decision-making process entirely, machine-learning models have become increasingly used in recent years. In this context, the Swiss avalanche warning service has been at the forefront of using these innovative data-driven tools to supplement current forecasting practices. These data-driven approaches have shown the potential to provide objective decision tools with higher spatiotemporal resolution than human-based forecasts. Nevertheless, the predictions generated by these machine learning models are opaque, usually referred to as 'black boxes', making it challenging for avalanche forecasters to interpret how variables are combined to make predictions. To address this, we employed an explanation method to interpret predictions from a machine learning model developed to predict danger levels for dry snow conditions in Switzerland. Moreover, precise avalanche data collected from an automated seismic detection system at the Vallée de la Sionne test site (Switzerland) are used to validate these danger level forecasts during the winter season 2020-2021. A novelty in future model implementations would be integrating these timely avalanche detections into forecasting models.