Automatic curve extraction for digitizing rainfall strip charts

A method has been developed that largely automates the labor-intensive extraction work for large amounts of rainfall strip charts and paper rolls. The method consists of five basic steps: (1) scanning the charts and rolls to high-resolution digital images, (2) manually and visually registering relevant meta-information from charts and rolls and pre-processing rolls to locate day transitions (3) applying automatic curve extraction software in a batch process to determine the coordinates of cumulative rainfall lines on the images, (4) post-processing the curves that were not correctly determined in step (3), and (5) aggregating the cumulative rainfall in pixel coordinates to the desired time resolution. The core of the method is in step (3). Here a color detection procedure is introduced that automatically separates the background of the charts and rolls from the foreground grid and the rainfall curve. The rainfall curve is detected by minimization of a cost function. In total 321 station-years of locations in the Netherlands have successfully been digitized and transformed to long-term rainfall time series with 5-minute resolution. In about 30% of the cases, semi-automatic postprocessing of the results was needed using a purpose-built graphical interface application.
This percentage, however, strongly depends on the quality of the recorded curves and the
charts and rolls. Although developed for rainfall, the method can be applied to other elements
as well.