Abteilung Systemanalyse, Integrated Assessment und Modellierung

Purpose: The aims of this work are (1) to explore deep learning (DL) architectures, spectroscopic input types, and learning designs toward optimal quantification in MR spectroscopy of simulated pathological spectra; and (2) to demonstrate accuracy and precision of DL predictions in view of inherent bias toward the training distribution.
Methods: Simulated 1D spectra and 2D spectrograms that mimic an extensive range of pathological in vivo conditions are used to train and test 24 different DL architectures. Active learning through altered training and testing data distributions is probed to optimize quantification performance. Ensembles of networks are explored to improve DL robustness and reduce the variance of estimates. A set of scores compares performances of DL predictions and traditional model fitting (MF).
Results: Ensembles of heterogeneous networks that combine 1D frequency-domain and 2D time-frequency domain spectrograms as input perform best. Dataset augmentation with active learning can improve performance, but gains are limited. MF is more accurate, although DL appears to be more precise at low SNR. However, this overall improved precision originates from a strong bias for cases with high uncertainty toward the dataset the network has been trained with, tending toward its average value.
Conclusion: MF mostly performs better compared to the faster DL approach. Potential intrinsic biases on training sets are dangerous in a clinical context that requires the algorithm to be unbiased to outliers (i.e., pathological data). Active learning and ensemble of networks are good strategies to improve prediction performances. However, data quality (sufficient SNR) has proven as a bottleneck for adequate unbiased performance - like in the case of MF.

Renewable energy is the key to reducing greenhouse gas emissions, and is one of the most concerning issues worldwide. China has the largest hydropower potential in the world. Yet, how China’s hydropower potential will change under 1.5 °C and 2.0 °C global warming and beyond remains unknown. Here, we find that China’s hydropower will increase greatly because of global warming. Gross hydropower potential (GHP) will increase by about one-half compared to the baseline period (1986-2015) under 1.5 °C and 2.0 °C warming, and about two-thirds under 4.5 °C warming. The spatial and temporal changes in GHP will vary largely. GHP will increase relatively more in summer than in winter, and more in Southwest China than in other regions. Compared to GHP, increases in per-capita GHP will be relatively less under 1.5 °C (5%) and 2.0 °C (7%) warming, but of a similar magnitude under 4.5 °C warming (71%). This study provides important information on China’s hydropower potential changes under global warming.

Study region: Middle and lower Hanjiang River (MLHR) basin, China.
Study focus: Changes in streamflow are often due to intertwined factors like climatic variability, land use change, and hydraulic constructions, whose relative impact is however poorly understood. In this study, we disentangled these effects by comparing real data and modelled scenarios of catchment behaviour. Modelled scenarios employ the soil and water assessment tool (SWAT), whereas the scenario comparison uses the Indicators of Hydrologic Alteration (IHA).
New hydrological insights for the region: Our analyses show that (1) watershed inlet was the major factor altered the streamflow regime at the watershed outlet, which pointed us to consider only the relative catchment contribution in further analyses. (2) Climate variability was the main driver of the net changes in natural hydrological regime, which downplayed the effect of land use change on streamflow variability. (3) The streamflow regulation associated to the progressive increase in reservoirs and their operation significantly altered the flow pattern, causing a general decrease in average streamflow, an attenuation of extreme events indicators, and an alteration of the pulse pattern with more frequent but shorter pulses. (4) The average water withdrawals at MLHR were estimated, which were 8.03 × 10⁹ m³/year. Overall, this research provides a path to assess the hydrologic impact of cascading reservoirs at the basin level under climate variability and land use change elsewhere.

Monitoring biodiversity is paramount to manage and protect natural resources. Collecting images of organisms over large temporal or spatial scales is a promising practice to monitor the biodiversity of natural ecosystems, providing large amounts of data with minimal interference with the environment. Deep learning models are currently used to automate classification of organisms into taxonomic units. However, imprecision in these classifiers introduces a measurement noise that is difficult to control and can significantly hinder the analysis and interpretation of data. We overcome this limitation through ensembles of Data-efficient image Transformers (DeiTs), which not only are easy to train and implement, but also significantly outperform the previous state of the art (SOTA). We validate our results on ten ecological imaging datasets of diverse origin, ranging from plankton to birds. On all the datasets, we achieve a new SOTA, with a reduction of the error with respect to the previous SOTA ranging from 29.35% to 100.00%, and often achieving performances very close to perfect classification. Ensembles of DeiTs perform better not because of superior single-model performances but rather due to smaller overlaps in the predictions by independent models and lower top-1 probabilities. This increases the benefit of ensembling, especially when using geometric averages to combine individual learners. While we only test our approach on biodiversity image datasets, our approach is generic and can be applied to any kind of images.

Plant phenology models are important components in process-based crop models, which are used to assess the impact of climate change on food production. For reliable model predictions, parameters in phenology models have to be accurately known. They are usually estimated by calibrating the model to observations. However, at regional scales in which different cultivars of a crop species may be grown, not accounting for inherent differences in phenological development between cultivars in the model and the presence of model deficits lead to inaccurate parameter estimates. To account for inherent differences between cultivars and to identify model deficits, we used a Bayesian multi-level approach to calibrate a phenology model (SPASS) to observations of silage maize grown across Germany between 2009 and 2017. We evaluated four multi-level models of increasing complexity, where we accounted for different combinations of ecological, weather, and year effects, as well as the hierarchical classification of cultivars nested within ripening groups of the maize species. We compared the calibration quality from this approach to the commonly used pooled approach in which none of these factors are considered. The pooled model led to over-confident process model parameter estimates and comparatively poor calibration quality. The mean value of the unexplained residual error standard deviation reduced from 5.5 BBCH (phenological development units) in the pooled model case (BM-0) to 5.3 BBCH when eco-region and year effects (BMM-1) were considered. Additionally accounting for weather effects (BMM-2a) resulted in a mean value of 5.2 BBCH. Calibration quality especially improved when the hierarchical classification of cultivars within ripening groups of maize was incorporated. Including the hierarchical classification with eco-region and year effects (BMM-2b) led to a mean residual error of 4.4 BBCH while additionally considering weather effects in the full model case (BMM-3) resulted in a value of 4.3 BBCH. Our findings have implications for regional model calibration and data-gathering studies, since it emphasizes that ripening group and cultivar information is essential. Furthermore, we found that if this information is not available, at least weather, eco-region and year effects should be taken into account. Accounting for only the eco-region and year effects led to parameter-compensation of the missing weather effects. Our results can facilitate model improvement studies since we identified possible model limitations related to temperature effects in the reproductive (post-flowering) phase and to soil-moisture. We demonstrate that Bayesian multi-level calibration of a phenology model facilitates the incorporation of hierarchical dependencies and the identification of model limitations. Our approach can be extended to full crop models at different spatial scales.