Explainable artificial intelligence for understanding cross-scale diversification in agricultural landscapes

Agricultural production is expected to cope with numerous challenges in the future. Therefore, the diversification of agricultural systems has gained attention as a promising solution to meet ecological, economic, and social needs in the long term. Despite the overall positive reports of diversification effects, the joint distribution of temporal and spatial diversity and their influencing factors remains largely unexplored. This thesis aims to assess the interrelations of diversity within the agricultural system at the field, farm, and landscape levels and to investigate the links between temporal and spatial crop diversity in Brandenburg, Germany, from 2011 to 2020 using machine learning (ML) combined with explainable artificial intelligence (XAI) tools. This approach involves identifying influencing factors of agricultural diversity such as farm attributes, landscape complexity, and environmental conditions. Data on agricultural land use were derived from the Integrated Administration and Control System (IACS), covering more than 5,642 farms and over 150,000 fields annually. The results revealed (1) a negative relationship between temporal crop diversity at the field level and landscape level complexity associated with soil quality and (2) a positive link between temporal and spatial crop diversity associated with higher crop diversity at the farm level. The potential of using XAI as a trust-enhancing tool for machine learning in such an interdisciplinary research setting was explored by reviewing the recent use of XAI tools in environmental and Earth system sciences in 575 studies. This revealed (3) that despite the increasing use of XAI tools in the environmental and Earth system sciences over time, the question of whether XAI can contribute to increased trust, e.g., in high-stakes decision contexts, has been addressed in only 1.2% of the studies. These findings contribute to understanding synergies and trade-offs between diversity within agricultural systems across levels and scales and its influencing factors. They also provide insights into the potential of XAI tools toward trustworthy ML applications. These findings can serve to develop targeted diversification strategies that enhance the resilience of agricultural systems and the sustainability of agricultural landscapes and improve the trustworthiness of ML and XAI in interdisciplinary research settings.