Search

Spatiotemporal systems are common in the real world. Forecasting the multi-step future of these spatiotemporal systems based on past observations, or, Spatiotemporal Sequence Forecasting (STSF), is a significant and challenging problem. Due to the complex spatial and temporal relationships within the data and the potential long forecast horizon, it is challenging to design appropriate Deep Learning (DL) architectures for STSF. In this thesis, we explore DL architectures for STSF. We first define the STSF problem and classify it into three subcategories: Trajectory Forecasting of Moving Point Cloud (TF-MPC), STSF on Regular Grid (STSF-RG), and STSF on Irregular Grid (STSF-IG). We then propose architectures for STSF-RG and STSF-IG problems. For the STSF-RG problems, we proposed the Convolutional Long-Short Term Memory (ConvLSTM) and the Trajectory Gated Recurrent Unit (TrajGRU). ConvLSTM uses convolution in both the input-state and state-state transitions of LSTM and is better at capturing the spatiotemporal correlations than the Fully-connected LSTM (FC-LSTM). TrajGRU improves upon ConvLSTM by actively learning the recurrent connection structure, which achieves better prediction performance with less parameters. To better investigate the effectiveness of our proposed architectures and other DL models for STSF-RG, we chose to tackle the precipitation nowcasting problem, which is a representative STSF-RG problem with a huge real-world impact. By incorporating ConvLSTM into an Encoder-Forecaster (EF) structure, we proposed the first machine learning based solution for precipitation nowcasting that outperforms the operational algorithm. To facilitate future studies for this problem and gauge the state-of the-art methods, we proposed the first large-scale benchmark for precipitation nowcasting: HKO-7. HKO-7 has new evaluation metrics and has both the offline setting and the online settings in the evaluation protocol. We evaluated seven models in the offline and online set

Spatiotemporal systems are common in the real world. Forecasting the multi-step future of these spatiotemporal systems based on past observations, or, Spatiotemporal Sequence Forecasting (STSF), is a significant and challenging problem. Due to the complex spatial and temporal relationships within the data and the potential long forecast horizon, it is challenging to design appropriate Deep Learning (DL) architectures for STSF. In this thesis, we explore DL architectures for STSF. We first define the STSF problem and classify it into three subcategories: Trajectory Forecasting of Moving Point Cloud (TF-MPC), STSF on Regular Grid (STSF-RG), and STSF on Irregular Grid (STSF-IG). We then propose architectures for STSF-RG and STSF-IG problems. For the STSF-RG problems, we proposed the Convolutional Long-Short Term Memory (ConvLSTM) and the Trajectory Gated Recurrent Unit (TrajGRU). ConvLSTM uses convolution in both the input-state and state-state transitions of LSTM and is better at capturing the spatiotemporal correlations than the Fully-connected LSTM (FC-LSTM). TrajGRU improves upon ConvLSTM by actively learning the recurrent connection structure, which achieves better prediction performance with less parameters. To better investigate the effectiveness of our proposed architectures and other DL models for STSF-RG, we chose to tackle the precipitation nowcasting problem, which is a representative STSF-RG problem with a huge real-world impact. By incorporating ConvLSTM into an Encoder-Forecaster (EF) structure, we proposed the first machine learning based solution for precipitation nowcasting that outperforms the operational algorithm. To facilitate future studies for this problem and gauge the state-of the-art methods, we proposed the first large-scale benchmark for precipitation nowcasting: HKO-7. HKO-7 has new evaluation metrics and has both the offline setting and the online settings in the evaluation protocol. We evaluated seven models in the offline and online set

Spatiotemporal systems are common in the real world. Forecasting the multi-step future of these spatiotemporal systems based on past observations, or, Spatiotemporal Sequence Forecasting (STSF), is a significant and challenging problem. Due to the complex spatial and temporal relationships within the data and the potential long forecast horizon, it is challenging to design appropriate Deep Learning (DL) architectures for STSF. In this thesis, we explore DL architectures for STSF. We first define the STSF problem and classify it into three subcategories: Trajectory Forecasting of Moving Point Cloud (TF-MPC), STSF on Regular Grid (STSF-RG), and STSF on Irregular Grid (STSF-IG). We then propose architectures for STSF-RG and STSF-IG problems. For the STSF-RG problems, we proposed the Convolutional Long-Short Term Memory (ConvLSTM) and the Trajectory Gated Recurrent Unit (TrajGRU). ConvLSTM uses convolution in both the input-state and state-state transitions of LSTM and is better at capturing the spatiotemporal correlations than the Fully-connected LSTM (FC-LSTM). TrajGRU improves upon ConvLSTM by actively learning the recurrent connection structure, which achieves better prediction performance with less parameters. To better investigate the effectiveness of our proposed architectures and other DL models for STSF-RG, we chose to tackle the precipitation nowcasting problem, which is a representative STSF-RG problem with a huge real-world impact. By incorporating ConvLSTM into an Encoder-Forecaster (EF) structure, we proposed the first machine learning based solution for precipitation nowcasting that outperforms the operational algorithm. To facilitate future studies for this problem and gauge the state-of the-art methods, we proposed the first large-scale benchmark for precipitation nowcasting: HKO-7. HKO-7 has new evaluation metrics and has both the offline setting and the online settings in the evaluation protocol. We evaluated seven models in the offline and online set

The mitigation of debris flow (DF) risk can be achieved by the means of reliable warning systems. At present time, operational warning systems are mostly based on critical rainfall thresholds (CRT) derived through the elaboration of past event records. The prediction ability of these systems strictly depends on the quality and availability of historical data. In order to overcome these limitations, in this work the CRT curves are derived from mathematical and numerical simulations carried out using a coupled hydrological-stability model. The CRT curves thus obtained are combined with rainfall nowcasts based on the observations of meteorological radar network. A case study in central Switzerland is presented. The performances of the proposed system are tested through a playback analysis of a past event. It results that the use of nowcasting allows for a crucial increment of the advance time interval., Peer Reviewed, Postprint (published version)

The objective of this study is to assess the skill of a precipitation nowcasting (very short range forecasting) system, with particular emphasis on hig-impact Heavy Precipitation Events (HPE). A threshold amount of 20 mm in 30 minutes is adopted, according to the first warning level issued for high rainfall rates by the Meteorological Service of Catalonia (NE Spain). These type of events are relatively frequent during autumn in the Mediterranean area and may produce flash-floods and substantial socio-economic losses when affecting densely populated areas in small catchments near the coast. Moreover, they can also produce local damage due to severe weather induced by winds of convective origin such as tornadoes or microbursts. Therefore, under these circumstances nowcasting systems producing very-short range (<3h) quantitative precipitation forecasts (QPF) are particularly necessary. Recent studies have pointed out the benefits of combining or blending the information from radar-based nowcasting systems and high resolution Numerical Weather Prediction (NWP) models to fill the gap between the lead-times of forecasts provided by each type of system - see for example Atencia et al. (2010), Berenguer et al. (2012) or the review by Pierce et al. (2012). Here we focus on assessing the quality of precipitation forecasts to examine realistic lead-times provided in four selected cases (see Table 1) for different precipitation amounts. The evaluation is done in two ways. Firstly forecasts are compared against radar based Quantitative Precipitation Estimates (QPE). Secondly the QPF fields are compared against actual raingauge data. Results are presented for all events together and for individual cases., Peer Reviewed, Postprint (published version)

Quantitative Precipitation Nowcasting (QPN) is one of the main applications of radar observations. On one hand, one of the most used nowcasting algorithms is Lagrangian extrapolation. It shows skill in specifying the timing and location of precipitation over short time periods, but shows low skill when using past precipitation trends to predict changes in precipitation intensity. On the other hand, Numerical Weather Prediction (NWP) models have roof skill at predicting the precise timing and location of precipitation, although they provide useful information about the intensity trends. It is therefore to investigate whether this additional information provided by NWP could be used to improve QPN. SBMcast (Berenguer et al., 2011) is an ensemble nowcasting algorithm based on Lagrangian extrapolation of recent radar observations. It generates a set of future rainfall scenarios (ensemble members) compatible with observations and preserving the spatial and temporal structure of the rainfall field according to the String of Beads model (Pegram and Clothier, 2001). This study shows the first results obtained with a methodology to constrain the spread of SBMcast ensembles with the additional information provided by NWP., Peer Reviewed, Postprint (author’s final draft)

Tactical Environmental Processor (TEP) is a through-the-sensor technique that converts radar returns from the AN/SPY-1 into environmental information known as spectral moments. TEP was installed aboard the USS Normandy (CG 60) in May 2000 to support a Limited Objective Experiment during Joint Task Force Exercise (JTFEX) 00-2. On 15 May, TEP observed severe weather associated with a line of passing thunderstorms. These weather events proved serious enough to suspend mid-cycle flight operations for the USS George Washington (CVN 73) during its simulated wartime scenario. TEP is a significant benefit to nowcast weather forecasting and supports at-sea METOC and warfighters in two primary areas: improved situational awareness and optimization of sensors, weapons and tactics. Results from this case study demonstrate the importance of TEP as a Doppler at-sea weather radar in support of naval operations., http://archive.org/details/utilityoftactica109455850, Lieutenant, United States Navy, Approved for public release; distribution is unlimited.