The primary objective of this research project was to investigate the application of semantic image segmentation using neural networks to study biological soil crusts (biocrusts). These crusts, forming a thin layer on the soil surface in hot and cold deserts, play a pivotal role in ecological processes such as carbon and nitrogen cycling, biodiversity preservation, erosion protection, and soil dust emission reduction. However, biocrusts are highly sensitive to climate variations and changes in land use intensity.

A significant challenge encountered in this study was the divergence in statistical properties between the training and target data, often captured by different camera devices. This discrepancy includes variations in scale, resolution, brightness, and other factors, potentially impacting the model's performance. To address this issue, we proposed a novel domain adaptation method using a joint energy-based approach.

The project involved the evaluation of a semantic segmentation approach for remote sensing and biomonitoring using biocrust imagery sourced from Utah (United States of America) and two sub-datasets from the National Park Gesäuse (Austria). Impressively, the results demonstrated a robust performance, achieving an overall classification accuracy of 85.9% for the USA data and 88.6% and 91.4%, respectively, for the two sub-datasets from the National Park Gesäuse.

Furthermore, we focused on testing the efficacy of our newly introduced joint energy-based domain adaptation approach, specifically on the sub-datasets from the National Park Gesäuse, which were captured using different camera devices. The implementation of this approach resulted in a noteworthy improvement in segmentation accuracy on the unlabeled sub-dataset, increasing from 70.4% to 75.3%.

The outcomes of this study underscore the effectiveness of our joint energy-based modeling as a viable domain adaptation method for semantic segmentation. This approach exhibits promise in addressing various challenges associated with deep learning and image-based biomonitoring, showcasing its versatility and potential applicability in diverse contexts.