Location: Northwest Watershed Management Research
Title: Quantifying spatial distribution of snow depth errors from LiDAR using Random Forests Authors
|Tinkham, Wade -|
|Smith, Alistair -|
|Marshall, Hans-Peter -|
|Link, Timothy -|
|Falkowski, Michael -|
Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: October 30, 2013
Publication Date: N/A
Interpretive Summary: There is increasing need to characterize the distribution of snow in complex terrain using remote sensing approaches, especially in isolated mountainous regions. These regions often have highly heterogeneous snow distributions that strongly affect stream discharge, soil moisture patterns, vegetation structure, and ecological habitats. Light Detection and Ranging (LiDAR) remote sensing techniques are rapidly advancing observational abilities to detect these vital heterogeneities with fine-scaled details heretofore unachievable. LiDAR retrievals are often conducted by commercial companies with documented accuracy standards. Stated accuracies however are based on instrumentation characteristics and remain static across the entire collection domain. Yet it is also known that terrain and vegetation characteristics also affect retrieval quality. Field campaigns that can ground truth snow, terrain, and vegetation characteristics are necessary to properly gauge data accuracy. Associated field campaigns however have been rare. This study used a combination of intensive field measurements and modeling to determine differences between measurements and LiDAR-based estimates of both snow-off topology and snow depths. Snow depth estimates had higher root-mean-squared-errors where trees were present and were most accurate in areas with an absence of trees (e.g. low-lying shrubs and meadows). Modeling over the entire watershed found differences in modeled and LiDAR-derived snow volumes of 21 – 30%. A large percentage of these errors were due to errors in the snow-off topology being comparable in magnitude to measured snow depths and would probably be lower in regions with greater snow accumulations. Errors could be further reduced through the development of vegetation-specific classification and processing techniques. This study pointed out that LiDAR accuracies have physical dependencies and are therefore spatially variable in regions with complex terrain and vegetation. As regards LiDAR-derived snow depths these errors can potentially represent a considerable percentage of actual values.
Technical Abstract: There is increasing need to characterize the distribution of snow in complex terrain using remote sensing approaches, especially in isolated mountainous regions that are often water-limited, the principal source of terrestrial freshwater, and sensitive to climatic shifts and variations. We apply intensive topographic surveys, multi-temporal LiDAR, and Random Forest modeling to quantify snow volume and characterize associated errors across seven land cover types in a semi-arid mountainous catchment at a 1 and 4 m spatial resolution. The LiDAR-based estimates of both snow-off surface topology and snow depths were validated against ground-based measurements across the catchment. LiDAR-derived snow depths estimates were most accurate in areas of low lying vegetation such as meadow and shrub vegetation (RMSE = 0.14 m) as compared to areas consisting of tree cover (RMSE = 0.20-0.35 m). The highest errors were found along the edge of conifer forests (RMSE = 0.35 m), however a second conifer transect outside the catchment had much lower errors (RMSE = 0.21 m). This difference is attributed to the wind exposure of the first site that led to highly variable snow depths at short spatial distances. The Random Forest modeled errors deviated from the field measured errors with a RMSE of 0.09-0.34 m across the different cover types. The modeling was used to calculate a theoretical lower and upper bound of catchment snow volume error of 21-30%. Results show that snow drifts, which are important for maintaining spring and summer stream flows and establishing and sustaining water-limited plant species, contained 30 ± 5-6% of the snow volume while only occupying 10% of the catchment area similar to findings by prior physically-based modeling approaches. This study demonstrates the potential utility of combining multi-temporal LiDAR with Random Forest modeling to quantify the distribution of snow depth with a reasonable degree of accuracy.