For the single-date maps we use a single Sentinel-2 image (bands: blue,
green, red, NIR and SWIR) and add the raw LiDAR data as an additional
input to the classifier. The year-maps are based on the same bands but
include a stack computed from all the images in one year with a low
cloud coverage. A classification on multi-temporal images of Sentinel-2
data is deemed more robust as this also captures changes over time .
Classified year maps are presently available for 2000-2019 in the
viewer. For 2000-2011 Landsat-5 data are used, from 2013-2014 Landsat-8
data, and 2015-2019 Sentinel-2 data. We do not include classification
for 2012, as the data from Landsat-7 is affected by sensor failure . The
full year image stack was created by building quarterly median images
for the Landsat data, and monthly median images for Sentinel-2 data.
This decision is based on image frequency for the region. Details on the
Landsat classification method can be found at Harezlak et al. (2020).
For each year in the Sentinel mission, the monthly medians yield an
image with 12 times 5 bands (medians of blue, green, red, NIR and SWIR).
For classification we use the Random Forest classifier which is a robust
classifier that (1) handles noisy and dissimilar data well and (2)
discriminates reliably with mixed classes . As the day-maps were meant
to be classified on-the-fly after user selection of a certain single
Sentinel image, we optimised the classification time by the number
pixels in the training and testing set and the number of trees in the
random forest classifier (data not shown). Finally, we set the number of
randomly selected pixels for each class to 200 (70% training, 30%
testing) and limited the number of trees to 6. We tested classifying the
Photo-Interpretation classes first and then aggregate to the
Legger-classes versus classifying the Legger-classes directly, there was
no difference in accuracy (data not shown). We opted for classifying the
Legger-classes directly in the final product.
The post processing steps consisted of creating a change map and ranking
and color-coding the changes to their relative change in hydraulic
roughness. We used a green-yellow-red colour ramp. The largest decrease
in hydraulic roughness (bush to water) was coded green, no change
yellow, and highest increase in hydraulic roughness (water to bush) was
coloured red.
The results are presented in an easy to use web-based map interface
based Netlify and MapBox (Netlify, 2020; Mapbox, 2020) with a connection
to GEE for data and computations. In this tool data layers are
presented, and metrics are computed on-the-fly for relative vegetation
cover per land ownership polygon. Classification results can also be
downloaded for further analysis in a dedicated GIS.