Improving the assessment of ICESat water altimetry accuracy accounting for autocorrelation

Given that water resources are scarce and are strained by competing demands, it has become crucial to develop and improve techniques to observe the temporal and spatial variations in the inland water volume. Due to the lack of data and the heterogeneity of water level stations, remote sensing, and especially altimetry from space, appear as complementary techniques for water level monitoring. In addition to spatial resolution and sampling rates in space or time, one of the most relevant criteria for satellite altimetry on inland water is the accuracy of the elevation data. Here, the accuracy of ICESat LIDAR altimetry product is assessed over the Great Lakes in North America. The accuracy assessment method used in this paper emphasizes on autocorrelation in high temporal frequency ICESat measurements. It also considers uncertainties resulting from both in situ lake level reference data. A probabilistic upscaling process was developed. This process is based on several successive ICESat shots averaged in a spatial transect accounting for autocorrelation between successive shots. The method also applies pre-processing of the ICESat data with saturation correction of ICESat waveforms, spatial filtering to avoid measurement disturbance from the land–water transition effects on waveform saturation and data selection to avoid trends in water elevations across space. Initially this paper analyzes 237 collected ICESat transects, consistent with the available hydrometric ground stations for four of the Great Lakes. By adapting a geostatistical framework, a high frequency autocorrelation between successive shot elevation values was observed and then modeled for 45% of the 237 transects. The modeled autocorrelation was therefore used to estimate water elevations at the transect scale and the resulting uncertainty for the 117 transects without trend. This uncertainty was 8 times greater than the usual computed uncertainty, when no temporal correlation is taken into account. This temporal correlation, corresponding to approximately 11 consecutive ICESat shots, could be linked to low transmitted ICESat GLAS energy and to poor weather conditions. Assuming Gaussian uncertainties for both reference data and ICESat data upscaled at the transect scale, we derived GLAS deviations statistics by averaging the results at station and lake scales. An overall bias of −4.6 cm (underestimation) and an overall standard deviation of 11.6 cm were computed for all lakes. Results demonstrated the relevance of taking autocorrelation into account in satellite data uncertainty assesment.