Conference Agenda

IntroductionThe integration of machine learning (ML) in Geographic Information Systems (GIS) has opened new avenues for spatial analysis, enabling more efficient and accurate geospatial data processing. The performance of ML models is often considered highly dependent on the granularity of input data, but the effects of different granularity levels have not been fully clear. In emergency applications, granularity may significantly affect the resulting actions. For example, a better understanding of the watercourse network is crucial in many applications and even lifesaving, e.g. flood modelling and preparedness (Petty, 2016). The National Land Survey of Finland (NLS) is currently renewing its watercourse mapping from high-resolution point clouds. In this study, we present a comparative analysis of six distinct datasets, varying in point cloud density and spatial resolution, to evaluate their effectiveness as input data for ML.

Methods

This study was conducted across two 36 km² study areas in central Finland, one near Ristijärvi and the other near Heinävesi. We trained six ML models with different input datasets, varying in point cloud density and spatial resolution. The datasets consisted of elevation models generated from lidar point clouds with densities of 20 points per square meter (20p) and 5 points per square meter (5p), at three spatial resolutions: 0.25 meters, 0.50 meters, and 1.00 meter. The NLS produced the datasets.

The primary objective was to assess the performance of these datasets as input for ML models in the research of watercourse networks. We implemented various ML training runs to achieve this, systematically altering several key parameters. Specifically, we varied the tile side length, cropping strategies, and the number of training epochs to observe how these factors influenced model performance.

We recorded multiple metrics to evaluate the models during the ML runs, including F1-score, precision, recall, and runtime. These metrics provided a comprehensive view of the effectiveness and efficiency of each dataset when used in ML tasks. We also calculated [KC1] relaxed recall, precision, and F1-score, with a spatial toleration of 1 meter, to better assess the F1-scores in relation to feature detection. The results from each run were carefully documented and analyzed.

After completing the ML runs, the outcomes were further analyzed and discussed within our group of four researchers to interpret the results and identify the datasets that yielded the best performance with visual analysis. Hillshading, relative topographic position, NLS orthophotos and the NLS topographic map were used in the visual analysis. This discussion also focused on understanding trade-offs between data granularity, computational efficiency, and the accuracy of the models.

Tropical cyclones (TCs) displace millions annually, ranking second only to floods in terms of people displaced by natural hazards. While TCs impose significant hardships and pose threats to lives, their negative impacts can be mitigated through anticipatory actions, such as evacuation, emergency protection, and coordinated humanitarian aid. Impact-based forecasting offers a valuable tool for planning these actions by providing detailed information on the numbers and locations of people at risk of displacement.

We present a fully open-source, globally consistent, and regionally calibrated TC-related displacement forecasting system. This system integrates meteorological forecasts with spatial data on population exposure and vulnerability, offering a cost-effective solution for predicting displacement risks. Through a case study of TC Yasa, which struck Fiji in December 2020, we illustrate the practical application of this forecast system. We highlight the importance of accounting for uncertainties in hazard, exposure, and vulnerability through a comprehensive global uncertainty analysis, which reveals a wide range of potential outcomes.

Additionally, we conducted a sensitivity analysis on all recorded TC displacement events from 2017 to 2020 to better understand the influence of uncertain inputs on forecast accuracy. Our findings suggest that for longer lead times, decision-making should prioritize meteorological uncertainties, whereas closer to landfall, the focus should shift towards the vulnerability of local communities. Our open-source code and methodologies are easily adaptable to other hazards and impact scenarios, making them valuable tools for broader applications in anticipatory humanitarian action.

Heavy rainfall events have become more frequent in recent years and in many cases the large amount of precipitation within a very short period of time leads to severe flooding and damage. Recent studies on climate change support this observation and show increasing trends in the frequency of extreme weather events, which include these high-intensity precipitation events.

In Germany, the high intensity as well as the extent and damaging impacts of flooding gain increasing attention among the society and the federal and state governments. Attention focus shifts towards heavy rainfall risk management and prevention of catastrophic harm and damage to people and property.

Contributing to an optimal preparation for the consequences of heavy rainfall events the Federal Agency for Cartography and Geodesy (BKG) introduces a nationwide scenario-based high-resolution pluvial flood map showing simulated information. The map evolves in close partnership with federal and state authorities and is currently being successively made available to politicians, authorities and the general public as OpenData for damage prevention and civil protection (Wimmer et al. 2023).

The foundation of the hydronumerical two-dimensional modelling are high-resolution geoinformation provided by the federal and state governments. A one-meter resolution Digital Terrain Model (DTM) is modified with culverts, 3D buildings, land cover data and further geoinformation to achieve a hydrologically effective and realistic discharge.

The resulting map provides two flood scenarios in line with the heavy rainfall index by Schmitt et al. (2018) which classifies the hazardous character of heavy rainfall events based on the return period. It is increasingly used in risk communication towards the general public in Germany. Based on a dataset of regionalized precipitation heights as a function of precipitation duration and annularity (KOSTRA) by the German Meteorological Service (DWD), a 100-year event is simulated. Additionally, the result of a heavy rainfall event with a precipitation of 100 mm/h is published. Both scenarios are visualized with flood depths, flow velocities and flow directions (geoportal.de/map.html?map=tk_04-starkregengefahrenhinweise-nrw).

The successively published data are used as initial assessment of risk potential. Combining the map with local expertise and government action guidelines simplifies the planning of measures. The map serves as an important and nationwide tool for identifying areas at risk of pluvial flooding. Local authorities, planners and emergency services are enabled to derive appropriate measures, both preventively and immediately before the event.

Building on the initial map, the BKG is currently developing a dynamic tool being part of the Digital Twin Germany project. It aims to provide event-related information on flooding caused by heavy rainfall. By simulating the effects of changes in infrastructure on the extent of pluvial flooding, users will be able to virtually test possible measures in advance in order to successfully minimize the impact.

Stream water temperatures influence water quality with effects on aquatic biodiversity, drinking water provision, electricity production, agriculture, and recreation. More frequent challenges from changes in water temperature are expected as we undergo the effects of the human-induced climate crisis. The current evidence on observed and projected long-term stream water temperature regional warming trends is often insufficient for stakeholders that stand to benefit from an operational forecasting service providing timely information from individual streams at high temporal resolution. Here we do so with deep learning models that efficiently generate multi-horizon probabilistic forecasts across multiple stations at once.

We train and evaluate several state-of-the-art models using 10 years of data from 54 stations across Switzerland. Static catchment features, time of the year, meteorological observations from the past 64 days, and their ensemble forecasts for the following 32 days are included as predictors in the models to estimate daily maximum water temperature over the next 32 days. The considered meteorological variables are daily average near surface air temperature, precipitation, and daily fraction of sunshine duration. Gridded observations of these variables are provided by MeteoSwiss at a spatial resolution of 2 km, and their forecasts correspond to a downscaled version of the Extended-range forecasts from the European Centre for Medium-Range Weather Forecasts.

Results show that the Temporal Fusion Transformer (TFT) model performs best with a Continuous Rank Probability Score (CRPS) of 0.70 ºC averaged over all lead times, stations, and 90 forecasts distributed over 1 year. The TFT is followed by the Recurrent Neural Network Encoder – Decoder with a CRPS of 0.74 ºC, and the Neural Hierarchical Interpolation for Time Series with a CRPS of 0.75 ºC. These deep learning models outperform other simpler models trained at each station: Random Forest (CRPS = 0.80 ºC), Multi-layer Perceptron neural network (CRPS = 0.81 ºC), and Autoregressive linear model (CRPS = 0.96 ºC). The average CRPS of the TFT degrades from 0.38 ºC at lead time of 1 day to 0.90 ºC at lead time of 32 days, largely driven by the uncertainty of the meteorological ensemble forecasts.

We find that the TFT also outperforms the other models when predicting water temperature at stations on which the models were not trained and at ungauged locations. The error at new stations increases 0.11 ºC reaching an average CRPS of 0.83 ºC, whereas it increases to 1.04 ºC when local water temperature is unavailable to the model. Furthermore, the TFT outputs feature importance and attention weights that provide valuable interpretability. In our case we find a dominant role of observed water temperature and future air temperature, while including precipitation and time of the year further improve the predictive skill.

Operational probabilistic forecasts of daily maximum water temperature for the next 32 days at 54 stations across Switzerland are generated twice per week with our TFT model and made publicly available at https://www.drought.ch/de/impakt-vorhersagen-malefix/wassertemperatur-prognosen/. Overall, this study provides insights on the extended range predictability of stream water temperature, and on the applicability of deep learning models in hydrology.

Collaborative GIS is spatial groupwork software and spatial decision support systems that bring people together in a digital map-based workspace to work on a shared geography-related topic. They have been useful in several disciplines, such as maritime spatial planning (Kettunen, 2020) and urban air mobility planning (Rönneberg, 2024). Collaborative GIS has a high potential for practical impact in emergency planning and action, as human actors can better communicate and become aware of the current spatial situation and its development scenarios for the near or more distant future.