4.8 Spatial Analysis

 The spatial analysis tools have been developed to generate and assess regional forecast distributions and skill (Figure 11). The tools are presented as a single interface with many options for generating and comparing a wide range of forecasting variable in multiple mapping windows. Zooming and scrolling is synchronized across multiple maps, and textural outputs are also available. The analysis can be used to compare different output types, predictive systems, forecasting rule-sets, and seasons. Key features of the analysis includes:

  • Approximately thirty output types can currently be displayed, and more are planned in the future. Outputs include training, testing and missing data counts, threshold values, forecasting probabilities, LEPS scores, percent-consistent scores, ROC scores, and p-value outputs for both tercile and above/below median forecasts. Multiple outputs can be compared directly in separate maps (Figure 11d);
  • Stations are included in the analysis by using the mouse (in combination with the CTRL and SHIFT keys) to select multiple regions and/or points on the chart.
  • Individual station results are displayed as coloured bubble-plots that are sized according to either a user-defined metric, or scaled according to the training data size. Colour schemes vary according to the output type. The user can select metadata to display above each station, including station name, training data size, and the numerical value of the plotted data (Figure 11a,d);
  • Double-clicking on station activates the chart overlays with navigation buttons provided to scroll through possible ‘browser’ and ‘station’ analysis outputs.
  • Contoured overlays can be generated using inbuilt kriging algorithms, which use landmass polygons to define the contouring boundaries. Bubble-plots and chart outputs can be further overlaid on these;
  • Plotting of results at stations can be filtered by selecting any single output as a filter-control. For example, the user can display the LEPS scores for stations that have a LEPS p-value of greater than 0.95 (providing a crude way of determining the LEPS value for climatology for the current predictive system). A summary of the regional average results can be displayed as an overlay in each map (Figure 11b,d,f);
  • Temporal forecast characteristics can be analysed though generating outputs for the twelve starting months of the year (Figure 11b);
  • Analyses are interactive with hot-tracking of the cursor to display station information; and
  • Different presentation options are provide in a right-hand side tools-panel.

 

The spatial analyses can be quite processor intensive. For this reason, they run multi-threaded in the background (currently only on a single core) to allow the user to interact with the interface while calculations are progressing. Changing settings midway though the calculations will restart the analysis with relevant results being updated. Computational ‘flags’ are embedding in the output data object to ensure that results are not recalculated unnecessarily. Results are stored in memory using complex multi-dimensional arrays for each predictand, and configured for output type, starting period, predictor, and ruleset.

Figure 11: Examples of spatial analysis outputs including (a) simple station output with station name and ‘count-sized’ points; (b) 12month output; (c) simple output with probability overlays; (d) Multiple output types; (e) point-based output with contoured underlay and printed result; and (f) multiple predictor analysis with result filtered on LEPS p-values.
Figure 11: Examples of spatial analysis outputs including (a) simple station output with station name and ‘count-sized’ points; (b) 12month output; (c) simple output with probability overlays; (d) Multiple output types; (e) point-based output with contoured underlay and printed result; and (f) multiple predictor analysis with result filtered on LEPS p-values.