PRACTISE – Photo Rectification And ClassificaTIon SoftwarE (V.2.1)

Terrestrial photography combined with the recently presented Photo Rectification And ClassificaTIon SoftwarE (PRACTISE V.1.0) has proven to be a valuable source to derive snow cover maps in a high temporal and spatial resolution. The areal coverage of the used digital photographs is however strongly limited. Satellite images on the other hand can cover larger areas but do show uncertainties with respect to the accurate detection of the snow covered area. This is especially the fact if user defined thresholds are needed e.g. in case of the frequently used Normalised-Difference Snow Index (NDSI). The 5 definition of this value is often not adequately defined by either a general value from literature or over the impression of the user but not by reproducible independent information. PRACTISE V.2.1 addresses this important aspect and does show additional improvements. The Matlab based software is now able to automatically process and detect snow cover in satellite images. A simultaneously captured camera-derived snow cover map is in this case utilised as in-situ information for calibrating the NDSI threshold value. Moreover, an additional automatic snow cover classification, specifically developed to classify shadow10 affected photographs was included. The improved software was tested for photographs and Landsat 7 Enhanced Thematic Mapper (ETM+) as well as Landsat 8 Operational Land Imager (OLI) scenes in the Zugspitze massif (Germany). The results show that using terrestrial photography in combination with satellite imagery can lead to an objective, reproducible and userindependent derivation of the NDSI threshold and the resulting snow cover map. The presented method is not limited to the sensor system or the threshold used in here but offers manifold application options for other scientific branches. 15

However, station data of snow cover in alpine regions are rare except for a few well-equipped sites (Scherrer et al., 2004;Marty, 2008;Viviroli et al., 2011;Pomeroy et al., 2015). Manual in-situ measurements are often prevented for reasons of 25 remoteness and safety by the harsh environmental conditions (Klemes, 1990). Satellite remote sensing techniques are a big step forward in these data-scarce areas but it is still a challenge to achieve snow cover products with high spatial and temporal resolutions as well as a high accuracy (Klemes, 1990;Viviroli et al., 2011). The complementary use of ground and space borne measurements for observing mountainous snow cover as highlighted by Vivirioli et al. (2011) is a promising approach and a main driver of this paper. 30 Terrestrial photography is thereby utilised as "ground truth" data. This technique has been successfully applied in many applications in the context of glaciology and snow hydrology (Corripio, 2004;Rivera et al., 2008;Dumont et al., 2009;Garvelmann et al., 2013;Messerli and Grinsted, 2015;cf. Parajka et al., 2012 for an overview). The advantages of terrestrial photography are that this technique has a high accuracy, is non-invasive, and provides spatially distributed snow cover data in a high temporal and spatial resolution (Aschenwald et al., 2001;Hinkler et al., 2002;Corripio et al., 2004;Schmidt et al., 2009;Para-35 jka et al., 2012;Härer et al., 2013). The decreasing costs of digital cameras and camera lenses with no or minimal distortion, as well as the potential use of terrestrial photography in remote and hostile environments due to technical advancements in off-grid power supply and data transfer also need to be mentioned here.
The alpine snow cover patterns derived from terrestrial photography can then be used to evaluate spatially distributed (snow-) hydrological models like Alpine3D, SnowModel and others (Lehning et al., 2006;Liston and Elder, 2006;Bernhardt 40 et al., 2012). The high spatial resolution of the photograph snow cover maps is very valuable as snow cover strongly varies over time and space and the accurate description in models is difficult (Blöschl et al., 1991;Winstral and Marks, 2002;Bernhardt and Schulz, 2010). The high temporal resolution of the terrestrial camera systems, for example on an hourly basis, further enhances the probability of at least one suitable photograph per day despite the frequently occurring cloud and precipitation events at high altitudes (Härer et al., 2013). 45 To map the spatial snow cover distributions, the recorded 2-D photographs have to be classified and georectified. Corripio (2004) and Corripio et al. (2004) presented a software tool that eased the georectification process utilising the animation and rendering technique by Watt and Watt (1992). This also formed the basis for the Photo Rectification And ClassificaTIon SoftwarE (PRACTISE V.1.0; Härer et al., 2013). Though, the formulations for the calculation of the 3-D rotation and projection are slightly different to Corripio (2004) and Corripio et al. (2004). PRACTISE V.1.0 further simplifies and fastens the spatially 50 distributed monitoring of snow cover patterns in mountainous terrain as it includes in addition to the georectification module routines for the identification of camera location and orientation, the viewshed computation and the snow classification of photographs. A batch mode also allows the processing of several photographs and thus the generation of multiple snow cover maps in a single program evaluation.
The trade-off for the high spatial resolution snow cover maps from terrestrial photography is that these maps are restricted to 55 a comparatively small region. To monitor a complete catchment with an extent of several square kilometres and more, satellite imagery is more suitable. This data has a lower spatial and temporal resolution but it offers the advantage of long consistent time series and the coverage of large areas. The Normalised-Difference Snow Index (NDSI) formulated by Dozier in 1989 for Landsat data is thereby still a standard method to derive snow cover maps (cf. SNOMAP algorithm of the MODIS snow cover product; Hall et al., 2001;Hall and Riggs, 2007). Other promising methods like traditional supervised multispectral classifica-60 tions, artificial neural networks or spectral-mixture analyses are computationally highly intensive, need lots of additional input data or are dependent on the interpreter's knowledge (Hall et al., 2001). These techniques are thus difficult to automate.
The NDSI represents the space borne component in the synthesis of ground and satellite measurements in this study. The index relies on a band rationing technique with a simple but effective principle that snow is highly reflective in the visible bands (GREEN, ∼ 0.55 µm) while having a very low reflectance in the mid infrared bands (MIR, ∼ 1.6 µm, Dozier, 1989). In 65 this approach, it is assumed that snow is present within a satellite pixel if the NDSI is greater than 0.4, and the near infrared (NIR, ∼ 0.85 µm) reflectance value is above 0.11, NIR > 0.11 (Dozier, 1989;Hall et al., 1995).
The NIR condition ensures that water surfaces which can also have high NDSI values are not misclassified as snow.

70
The NDSI threshold value of 0.4 is the standard literature value (Nolin, 2010;Dietz et al., 2012) even though Hall et al. (1995) already mention that acceptable snow cover maps were found for NDSI thresholds between 0.25 and 0.45 in a study investigating six scenes in the United States and Iceland. This threshold range corresponded to changes in snow cover extent of more than ten percent in the studied scenes. In particular for local and regional applications it is thus crucial to set the NDSI threshold accurately but in a user-friendly and standardised manner. The manual adjustment of the threshold is no option in 75 most cases as it is not reproducible and offers the danger of adapting the resulting snow cover distribution to support a given hypothesis.
This paper presents a new method to monitor alpine snow cover patterns with satellite data by making use of terrestrial camera infrastructure, including webcams. The NDSI threshold value for snow is thereby calibrated to achieve an optimal agreement in the overlapping area of photograph and satellite snow cover map. Hence, an optimal NDSI-based satellite snow 80 cover map for the specific region and time is produced, for example for an alpine catchment with an extent of several square kilometres. The cameras needed for this method are often already available or can be easily installed at many sites. We focus on Landsat data in here, as the pixel dimensions of 30 m are in comparison to MODIS pixel sizes of 500 m preferable for local and regional applications, particularly as the instantaneous field of view further increases for mountainous terrain with steep slopes.

85
The new approach to complementary use ground and space borne measurements to derive snow cover maps is fully implemented in PRACTISE V.2.1. The fast and easy-to-use processing includes the NDSI calculation from Landsat raw data as well as the use of NDSI maps produced externally in geoinformation systems. Optionally, it also allows including an existing cloud mask using for example the freely available Fmask software (Zhu et al., 2015). In addition, a newly developed snow classification algorithm for shadow-affected photographs is presented in PRACTISE V.2.1. Further improvements are bug fixes and 90 revised code of already published modules as well as increased user friendliness. This paper is supplemented with an example dataset, a manual and the associated Matlab code. The structure of the paper itself is as follows: at first, the test site and data are described.  of Landsat Level 1 data, the use of externally generated satellite NDSI maps is possible here. The spatial processing extent is user-dependent as well as if an externally produced mask for clouds, (cloud) shadows and water is used or not. Another optional input is a Landsat Look image for visualisation. Zugspitzplatt area is located in the centre of the scene and is therefore not affected by this error.
The inputs given for the georectification of the SLR and webcam photographs are presented in Table 1. Camera dependent parameters were taken from the user manual of the camera systems. The focal lengths have been adjusted according to the used image. The location and target position of the camera, as well as the GCP locations have been identified combining photographs, DEM data, topographical maps and official orthophotos with a submeter spatial resolution. Nevertheless, the 140 camera location and target position could only be estimated. The camera parameters in Table 1 except the camera sensor and photograph dimensions thus need to be optimised using GCPs. A separate estimation for each photograph in this study is further necessary as the locations and orientations of the cameras are changing in between the photographs due to either weather effects like wind, for maintenance reasons or a new camera location at the UFS.
The DEM used for the SLR photographs has a spatial resolution of 1 m in the horizontal plane and originated from an  3.1 Snow classification in partially shadow affected photographs PRACTISE V.1.0 provides two snow classification routines for terrestrial RGB photographs. The user can select between a manual routine which basically detects snow for digital numbers (DN) above user-specific snow thresholds in the red, green and blue (RGB) bands of the digital photograph and an algorithm developed by Salvatori et al. (2011). This algorithm is a threshold based procedure which automatically analyses the blue band DN frequency histogram and sets the snow threshold.

160
Both classification types of PRACTISE V.1.0 are described in detail in Härer et al. (2013).
Both algorithms are working well if the photography is evenly illuminated and in the absence of shadows (Härer et al., 2013).
However, shadow-free situations are rare in structured terrain and clouds can reason further shadowing. In the case of shaded areas, the two included classification routines tend to only identify snow surfaces which are sunlit while the classification in shaded areas does have high uncertainties. This results from similarly high blue band DN in RGB images for shaded snow 165 cover, and illuminated rock, soil or sparsely vegetated surfaces ( Fig. 3a  The second step in the classification routine is the utilisation of a PCA to detect snow cover in shaded areas. The PCA is a statistical method to analyse multivariate data sets. In our case, we use the PCA to orthogonally transform the axes of the 175 RGB space to a new principal component (PC) space where the centre of the coordinate system is shifted to the mean value of the three-dimensional data set while the axis direction of the first PC (PC1) explains the largest variance in the data set. The axis of the second PC (PC2) is orthogonal to PC1 and explains the second largest variance. The axis of PC3 is again orthogonal to PC1 and PC2. Due to the decreasing explained variance in the higher components, most information of the RGB data is stored in PC1 and PC2 while PC3 mainly represents remaining noise.

180
For the PCA, the RGB values of all visible DEM pixels (m) are standardised so that each colour column has a mean of 0 and a standard deviation of 1 (RGB s ). The PC coefficients are calculated using a singular value decomposition. The m × 3 RGB s matrix is then multiplied with the 3 × 3 PC coefficient matrix and results in the m × 3 PC score matrix (PC sc ) which represents the standardised RGB values in the PCA space. The PC sc has a decreasing explained variance from column 1 to 3 (P C sc 1 to P C sc 3) and is normalised by scaling between 0 and 1 in a last step.

185
Frequency histograms of the normalised PC score matrix (P C sc,n ) for the columns 1 to 3 are illustrated in Fig. 5a to c.
The shape of the frequency histogram of P C sc,n 1 in the PCA space ( Fig. 5a) is essentially identical to the blue band DN frequency histogram in the RGB space. Hence, P C sc,n 1 is not analysed further as the first classification step already utilises this information. But the frequency histograms of P C sc,n 2 and P C sc,n 3 are used and play a major role in the separation of shaded snow from other surfaces. Empirical analyses of numerous photographs have shown that shaded snow pixels have 190 higher P C sc,n 2 than P C sc,n 3 values ( Fig. 5b and c). For the used example, this means that shaded snow cover is grouped in the local maximum around 0.7 in the frequency histogram of P C sc,n 2 (Fig. 5b). As a consequence, pixels are classified as snow where P C sc,n 3 < P C sc,n 2 and DN b,th ≥ DN b ≥ 63.
A blue band DN (DN b ) condition is additionally included as firstly, all pixels with DN b greater or equal to the derived snow The detected rock surfaces are depicted in blue in Fig. 4.
Finally, pixels not classified in the three steps before (DN b, n ) are assigned snow probability values (P s ) from 0 for no snow 205 to 1 for snow linearly increasing from low to high DN b . P s is not a statistically derived variable but is a helpful indicator as the probability of a snow covered pixel increases with higher reflectance values in the blue spectrum. P s is calculated using, Negative P s values are set to 0, no snow, as we assume that pixels with DN b below one fourth of the DN range (63) are areas free of snow. It should be noted that the blue band threshold of 63 in Eqs. (3) and (5) can be adjusted by the user even though 210 this was not necessary for any analysed photograph throughout the development of the routine.
Results of the newly implemented snow classification routine are illustrated in Fig. 6 and can be compared to the results of V1.0 in Fig. 3b. At last, we want to mention that the new routine and in particular the PC analysis step was successfully applied in at least 95 % of our shadow-affected test photographs. For shadow-free situations, it is though still recommended to use the existing classification routines presented in Härer et al. (2013).

Threshold calibration for optimal NDSI-based snow cover maps
The new approach to automatically derive an optimal NDSI-based snow cover map is implemented in the second new module of PRACTISE V.2.1. The method utilises areas which show an overlap between a photograph snow cover map and the NDSI product of a simultaneously captured satellite scene. Then, the NDSI threshold value for snow is calibrated using the DDS optimisation algorithm (Tolson and Shoemaker, 2007) to obtain an optimal agreement of photograph and satellite snow cover (1) (Dozier, 1989;Hall et al., 1995). We want to highlight that externally produced NDSI maps from satellites like Spot, MODIS Aqua and MODIS Terra can also be directly used.

230
If the Landsat scene is partially cloud covered, an externally generated cloud mask should be used to prevent misclassifications. A direct input link for the cloud mask product of the freely available Fmask software of Zhu et al. (2015) is integrated in PRACTISE to mask clouds, cloud shadows and water. The near infrared condition of Eq.
(2) (Dozier, 1989;Hall et al., 1995) which is used to prevent water surfaces from being classified as snow is also applied here for masking strongly shaded pixels prone to misclassifications.

235
Overlapping areas of terrestrial photography and the satellite image are subsequently detected. The results of the photograph snow cover maps are used as baseline. It is a user's decision if pixels classified as "unsure" in the photograph are excluded or used in weighted form according to their probability value. The user's selection, however, only affects the NDSI threshold calibration of the satellite image while the photograph snow cover map remains unchanged. Now, the DDS optimisation routine which is also implemented in the framework of the GCP optimisation (cf. Härer et al.,240 2013) is used to optimise the NDSI threshold value. The seed is set to the threshold of 0.4 recommended by Dozier (1989) and Hall et al. (1995) and the NDSI threshold value is limited to the range of NDSI values which can be found in the overlapping area. The number of maximum iterations is user-dependent, but it was found that 150 optimisation runs are sufficient. A quality measure of Aronica et al. (2002) which was successfully used in the context of snow extent evaluation in Bernhardt and Schulz (2010) serves as objective function value F in the optimisation: n is the overall number of photo-satellite image pixel pairs whereas a represents the number of correctly identified snow pixels and d the same for snow-free pixels. F takes on values between 0 and 1 with 1 indicating a perfect agreement between the two images.
The routine is exemplarily presented for the SLR photograph of 17 November 2011 and the simultaneously captured Landsat (2) for reasons of cloud cover in the investigated area even though not visible in the Landsat Look image (Fig. 2a). Tests for several scenes of Landsat 7 and 8 have shown that masking clouds with a cloud probability of 95 % and a surrounding buffer of three pixels in Fmask is reasonable in this application. The buffer secures that the satellite pixels used are not influenced by the thin edges of 255 clouds and cloud shadows which could potentially lead to misclassifications.
The photograph and satellite snow cover maps of the SLR photograph and the Landsat 7 ETM+ image with an optimised NDSI threshold of 0.18 are illustrated in Fig. 7. The classification agreement in the overlapping area of photograph and satellite is 97 %. The snow cover extent amounts to 2.8 km 2 and the masked area due to shadows and clouds covers an area of 3.6 km 2 for this date. The areal coverages are calculated for the alpine Zugspitzplatt catchment (∼ 13.1 km 2 , Fig. 1) defined by the 260 catchment outlet at the Partnach spring.

Interactive modules, code improvements and the flow chart
In addition to the two new routines (Sects. 3.1 and 3.2), the code and the user friendliness of the existing modules in PRACTISE V.2.1 have been improved.
Interactive modes are now available in the modules, "optimisation of the camera location and orientation" and "snow clas-265 sification", which allows the user to directly interact with the software during runtime. Hence, the user can now interactively restart and refine the optimisation of the exterior and interior camera parameters without the need to restart the complete program evaluation. The interactive mode in the snow classification module allows switching between the three snow classification routines described in Sect. 3.1. The classification parameters for the different algorithms can also be adapted. The user can thus directly decide on the best classification method and parameters for each photograph.

270
PRACTISE V.2.1 is now also able to process photographs that were taken from camera locations sheltered by for example a roof and thus are assumed below ground in a DEM. In this case, surrounding DEM pixels will obstruct the view in the  3.3). Then, the viewshed is calculated for the respective camera system and the georectification procedure is executed. All visible DEM pixels are subsequently classified as snow-covered or snow-free by using the automatic blue band snow classification routine described in detail in Härer et al. (2013). Interactively switching to other classification routines and adapting the classification parameters is possible here (Sect. 3.1). In a next step, the NDSI is calculated for Landsat pixels which are not masked by the NIR condition in Eq.
(2) and an externally generated Fmask satellite image cloud mask 285 (Zhu et al., 2015). Areas which are covered by terrestrial photography and satellite are eventually used to calibrate the NDSI threshold value (Sect. 3.2). Final outputs of the described PRACTISE run are snow cover maps based on the SLR photography and Landsat Level 1 data, a Landsat NDSI map and the computed viewshed.
The runtime of PRACTISE V.2.1 for this setup with a photographed area of about 0.3 km 2 and a Landsat processing extent of 30 km 2 was about 58.6 s on an Intel Core i7-2600 CPU with 3.4 GHz utilising 1.2 GB of memory (RAM). However, interactive 290 modes were deactivated in the runtime measurement and hence the optimisation of camera parameters with 3000 iterations (∼ 0.58 s) was executed only once.

Results and discussion
We Misinterpretations in the georectification and as a result in the classification were only found for snow groomers and some 310 infrastructure not represented in the DEM and viewshed. An example of these obstacles leading to misinterpretations is an antenna in the centre of the webcam photographs (cf. Fig. 9b). As the number of pixels affected by this and similar problems is less than 0.5 % of the mapped area, the georectification quality of all camera images can be summarized as very high. to the high sun angle at this date. Hence, the automatic blue band classification algorithm was used. The resulting classification visually indicates a high quality and will not be further discussed here as the method was evaluated before in Salvatori et al. (2011) and Härer et al. (2013).
For the photographs of 7 April 2014 the PCA-based classification algorithm was applied to reduce shadow-related misclassifications ( Fig. 10c and d). The detailed visual analysis of the pixels in the two April photographs proved a high quality of the 320 new classification routine for pixels identified as "snow" and "free of snow" as well as for pixels classified as "probably snow", "highly unsure", and "probably no snow".
Misclassifications in the main classification categories "snow" and "free of snow" are rare with less than 0.3 % of classified pixels in the SLR photograph and less than 1 % in the webcam photograph. The reasons for misclassifications are however different in both photographs. In the SLR photograph, the misclassifications can mainly be attributed to the light-coloured bare 325 rock (limestone) in the Zugspitzplatt area which is mistakenly classified as "snow". This issue has already been discussed in detail in Härer et al. (2013, PRACTISE V.1.0) and is a weakness of the blue band classification method which represents one of the classification steps in the PCA-based classification routine. The misclassifications in the webcam photograph have two main reasons: A georectification problem due to infrastructure which has already been mentioned above and another problem, as shaded areas, in particular in the valley below the Zugspitzplatt, are difficult to classify as "snow" and "no snow", even with 330 the human eye.
Additionally to the two main classification categories, the three "unsure" categories need to be discussed for the April photographs. 1.9 % of classified pixels in the SLR photograph and 7.8 % in the webcam image are assigned probability values.
The low percentages emphasize that the assignment rules in the PCA-based classification routine seem to describe the RGB characteristics of the different surfaces well. In addition, most pixels classified as "unsure" in the SLR photograph are exactly 335 located at the transitional area between snow patches and snow-free areas in the photographs, and can therefore be seen as mixed pixels (Figs. 10c and 11a). The classification of the SLR photograph on 17 November 2011 (Fig. 6) has also attested this finding.
In the webcam photograph, more pixels are classified as "unsure" in particular as "probably no snow" (Figs. 10d and 11b).
The detailed analysis also shows some "no snow" misclassifications in the webcam photograph, especially in the transitional 340 zone between sunny and shaded snow. Taken together, both issues concern less than 0.5 % of the classified pixels and are only observed in the webcam photograph while the SLR photographs in TIFF-format, allowing data compression without loss, are unaffected. The finding of more "unsure" classifications and the misclassification issue could be traced back to the lower image quality and the JPEG compression of the webcam image. Hence, such uncertainties and small errors have to be expected in the context of any analysis which uses JPEG images. Overall, the new classification technique separates sunny as well as shaded 345 snow cover from other surfaces with a similar high accuracy as the blue band classification does classify equally illuminated photographs.
In a second step, the calibration of the NDSI threshold of the Landsat images was evaluated. At first, the results of the Landsat 7 ETM+ satellite image of 1 July 2013 are presented. The SLR calibrated NDSI threshold of this Landsat scene is 0.35 with 94 % of the photo snow cover map being identical to the calibrated satellite image snow cover map. The calibration 350 of the NDSI threshold using the webcam photograph results in a threshold of 0.37. Here, the classification agreement of the snow cover is with 84 % slightly lower but still high. In the Landsat 8 OLI satellite image of 7 April 2014, the NDSI threshold optimised with the SLR photograph is 0.23 (94 % agreement). An identical NDSI threshold of 0.23 (90 % agreement) was found for the simultaneously captured webcam photograph. The snow cover maps from the SLR and webcam photographs as well as from the SLR calibrated satellite images are depicted in Fig. 12a for 1 July 2013 and in Fig. 12b for 7 April 2014. The

355
SLR derived snow covered area in the Zugspitzplatt catchment amounts to 6.5 km 2 for the July date, respectively 9.9 km 2 for the April date. Masked areas are 1.3 km 2 on 1 July 2013 and 0.9 km 2 on 7 April 2014 due to shadows and clouds.
We want to emphasize here that the percentage of pixels identically classified in photograph and satellite image map is enormously high having in mind the different horizontal resolutions of photograph map (SLR: 1 m, webcam: 5 m) and satellite image map (30 m). The resolution effect becomes more pronounced for patchier snow cover, in this case in the lower 360 Zugspitzplatt area, which also explains the slightly lower agreement between webcam photograph and satellite image.
Another important finding is that the calibration of the NDSI threshold using SLR and webcam result in almost identical NDSI thresholds. As the differences are insignificant the NDSI threshold calibration seems to be robust in the Zugspitzplatt area independently of the used camera system and field of view.  The presented values and findings underline that the strong temporal variations found in NDSI thresholds transfer to large uncertainties in the derivation of snow cover extents and studies relying on these snow cover products. A spatial and temporal adjustment of NDSI thresholds is therefore important to ensure optimum results in the snow cover mapping of specific areas, for example of the studied alpine catchment. ability of this study to other areas. Obviously when using freely available webcam infrastructure, the processing of PRACTISE needs an increased attention for any problems that may arise in the snow mapping due to image quality, lens distortion and obstacles in the field of view.
Our next step will be to apply PRACTISE and the integrated new approach to the complete available time series of photographs and satellite images in the Zugspitzplatt area. In addition, we will process another long-term time series of photographs 405 in the alpine Vernagtferner area, Austria, which is located in the same Landsat scene as the Zugspitzplatt. We think that this experimental setup will be a first step towards understanding the temporal variability of the calibrated NDSI thresholds in alpine areas. Furthermore, the setup will also allow for testing spatial representativeness of the optimal NDSI threshold on the regional scale as this is another topic of ongoing discussion. This will be especially important as the spatio-temporal extrapolation possibilities and limits of the presented method are as yet unknown. Further research will also be necessary to verify if 410 the synthesis of terrestrial photograph and satellite image is applicable in a modified form to other research fields like thermal photography and satellite imagery.

Code availability
The source code of PRACTISE V. for that is simply that no precompiled 64 bit-version of Octave 4.0 is available for Windows yet which is though necessary to process large arrays in PRACTISE.
The Supplement related to this article is available online at doi:10.5281/zenodo.35646.       are depicted in blue. The pixels classified as "probably snow" (orange), "probably no snow" (light blue) and "highly unsure" (yellow) are enlarged for the sake of clarity. Only about 3.6 % of all classified pixels fall within one of the three probability categories and hence are assumed as "unsure".    Here, snow and snow-free pixels are again displayed in red and blue but additionally unsure classification results are illustrated in light blue, yellow and orange for the categories "probably no snow", "highly unsure", and "probably snow". Black rectangle boxes in (c) and (d) are depicted for detailed analyses of the classification accuracy ( Fig. 11a and b). In the webcam photograph, more pixels are classified in the three "unsure" categories. Moreover, some "no snow" misclassifications are found at the transitional zone between sunny and shaded snow cover. . Snow cover in the satellite data is illustrated with white crosses, masked areas with black crosses mainly due to clouds in (a) and shadows in (b). Pixels not superimposed with crosses are areas classified as "free of snow" in the satellite images. The photograph snow cover maps display "snow" and "no snow" in red and blue for the SLR and in light red and light blue for the webcam. "Unsure" snow classification results only occur for the photographs on 7 April 2014 (b) as the PCA-based classification routine is applied and are only shown for the webcam (yellow) as the percentage of "unsure" snow classifications in the SLR photograph ( Figs. 10c and 11a) is negligible.