Monitoring changes in diffuse pollution source risk with time-lapse photography
There are many factors within the landscape that can affect diffuse pollution source risk that vary over the year. Examples include the snow cover dynamics driving overland flow not directly connected to precipitation on that day, the changes in solar receipt resulting from cloud cover and land cover changes resulting from agricultural practices. An approach that was trialled within the first phase of DTC was the use of a time-lapse camera to create a visual record on the landscape, which could then be processed to give a continuous record of the factors. It is hoped that this extra information would provide useful context and background information for the interpretation of in-stream water quality and river flow monitoring results.
Location and Equipment
A single camera was installed in the Dacre Beck sub-catchment. This location was selected due to the suitability of the topography for the approach and local issues with snow melt that are most relevant at this site. The camera location and a typical image are shown in figure 1 below. The camera installed was a Canon 5D mk I DSLR with a 24mm f2.8 lens set to f5.0 and with a fixed ISO of 200 in aperture priority mode. This setup gave a horizontal angle of view of 73.7°. The camera was triggered to take an image every 10 minutes and has been installed since Jan 2014.
Figure 1: The location of the time-lapse camera and a typical acquired image from the site from 24/06/14. The red triangle in the map shows the camera direction and angle of view of the images.
Converting Images to Information
The images have been processed with the Python Image Library and SciPy to extract information on the changing characteristics of the landscape. The approach taken for cloud cover and snow cover followed the same steps:
Define the area of interest within the image for analysis. For the clear skies, this an area extracted from the sky and for the snow cover, this area was defined as the local foreground field.
A Gaussian blur filter was applied to the image with a high radius (200) to give a consistent colour value within the extracted area of interest image.
The red, green and blue components of the colour were extracted for analysis.
a. For ‘clear skies’, the index was defined as blue / ((green + red) / 2) and hence the lower values represent cloudy days and higher values one represent blue sky days
b. For the ‘snowiness’, the index was defined as blue / green with values towards one representing snow cover
The approach taken to find ‘interesting’ images was based on the calculated change in the scene between days. The changes in the overall scene were assessed by comparing the statistical differences between pairs of images. Two approaches were tested, the first compared the current day to the previous day and the second approach compared each day to the averaged scene from the whole year. The differences between the images were defined as the ‘Manhattan’ difference whereby the mean movement in brightness levels per pixel is calculated after the images have been converted to greyscale and normalised.
Initial Results for Blue Sky and Snow
In figure 2 below, there are the initial results for the 2014 calendar year for the clear skies and snowiness indexes along with example images. These initial results are based on the midday images from each day.
Figure 2 Changes in the clear skies and snowiness index over the 2014 calendar year.
From the results, it can be seen that the snowiness index correctly identify days with snow cover when the index value exceeds 0.87. There is a strong seasonal trend in the snowiness index values wit the lowest values in the summer. The ‘blue skies’ index shows the high level of cloud cover year the year with relatively few days being cloud free, 8.5% of days have a index value greater than 0.75. The example image in Figure 2 from 12/05/14 shows issues with condensation within the waterproof housing of the camera install. This condensation reduces the image quality for fine detail but it is still possible to extract information on the sky status.
Initial Results for Scene ‘Intrestingness’
Figure 3 below shows the changes over time for the daily differences when compared to the previous day and to the mean image. The mean image is shown in figure 4.
Figure 3 Changes over time for the daily differences when compared to the previous day and to the mean image
Figure 4, the average image from all midday images in 2014
From the results in Figure 3 it can be seen that the difference based indexes are capable of identifying the days of the year were the landscape or climate are different from the ‘normal’ conditions. The approach successfully identified the increased snow cover in February 2014, the occurrence of a change in the weather conditions from cloud to clear skies in April 2014 and the reverse in August and finally the occurrence of snowfall in December 2014.
Concluding Remarks
Time-lapse photography of the catchment and landscape has the capability to add significant contextual information to complement and support the in-stream water chemistry datasets
The approach generates significant amounts of data and hence tools are needed to identify the important times and associated images
The presented indexes in are capable of identify the properties of the environment, such as cloud cover or snow accumulation / melt.
It is also possible to identify times when there have been large changes in the landscape, either compared to the previous day or to the annual average conditions.