Geostatistical analysis of water quantity and quality spatiotemporal data

Regulatory environmental reporting on groundwater levels and various substance concentrations in ground water involves statistical analysis of a large number of data. However, statistical calculations have to take into account seasonality or temporal correlation; usual statistical calculations are not valid. Many questions arises as regards to automation of this processes: specific definition of variables, calculation simplification and robustness. We consider three main aspects: 1. Outliers or singularities detection for each time series, for instance the intermittent pumping presence on a groundwater level observation. Firstly the temporal nature of the measures is determined. Abnormality detection algorithms are proposed (interquantile deviation, multimodality of values or of their increments,...). 2. Accuracy of environmental indicators (mean, quantile) and characterization of their interannual variations. Approximations in the variance estimation calculations are required for automated processing, so that validity their validity is discussed. Furthermore due to the presence of quantization threshold, "micropolluant" concentration can be censured. What is the consequence in practice? 3. Detection of trends and breaks for time series. After a detailed comparison between different methods, BRGM selected the Mann-Kendall test for trend characterization. However, other procedure can be considered, for example mean comparison (or quantile) calculated on temporal support in order to introduce time scale aspect. Then, we study possible definition of trends and their breaks. These different approaches for trends and breaks are studied, they are compared and then combined. Finally, spatialization of indicators is discussed: definition, simplified calculation and precision estimation