Semi-Automatic Fractional Snow Cover Monitoring from Near-Surface Remote Sensing in Grassland

: Snow cover is an important variable in both climatological and hydrological studies because of its relationship to environmental energy and mass ﬂux. However, variability in snow cover can confound satellite-based efforts to monitor vegetation phenology. This research explores the utility of the PhenoCam Network cameras to estimate Fractional Snow Cover (FSC) in grassland. The goal is to operationalize FSC estimates from PhenoCams to inform and improve the satellite-based determination of phenological metrics. The study site is the Oakville Prairie Biological Field Station, located near Grand Forks, North Dakota. We developed a semi-automated process to estimate FSC from PhenoCam images through Python coding. Compared with previous research employing RGB images only, our use of the monochrome RGB + NIR (near-infrared) reduced pixel misclassiﬁcation and increased accuracy. The results had an average RMSE of less than 8% FSC compared to visual estimates. Our pixel-based accuracy assessment showed that the overall accuracy of the images selected for validation was 92%. This is a promising outcome, although not every PhenoCam Network system has NIR capability.


Introduction
While the spatial and temporal dynamics of snow cover are important variables in climatological and hydrological studies because of their relationships with fluxes of environmental energy and mass, they also confound satellite-based efforts to monitor vegetation phenology, often the purpose of time-series analysis of remotely sensed data. Phenology is the science of recording natural events to identify changes in seasonal and annual cycles of plants and animals. Monitoring phenology over time can account for changes in the ecology of species in response to climate change [1].
Remote sensing makes possible frequent, repeat monitoring of vegetation canopies, especially in difficult-to-access locations. Researchers use indexes such as the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), and the Green Chromatic Coordinate (GCC) to track vegetation phenology and to quantify and monitor plant stages and plant responses to weather and climate [2][3][4][5][6][7][8][9][10].
Establishing an accurate and precise record of vegetation phenology across expansive geographic regions to validate satellite observations is the primary purpose of a system of terrestrial cameras comprising the PhenoCam Network [2,4,5,11]. Scientists have established PhenoCams to monitor vegetation in many different ecoregions, including tallgrass prairie, the focus of this study [3,12,13]. PhenoCams are fixed terrestrial webcams taking frequent, repeat oblique pictures that are used to monitor vegetation phenology and other changes in land surfaces, including snow cover dynamics [7,14]. These systems are low-cost [15], and there are 393 cameras currently located across the U.S. [16].
In prairies, Fractional Snow Cover (FSC), the percentage of an area of land covered by snow, creates challenges for accurate phenology monitoring from satellite sensors. For example, the presence of snow affects EVI by decreasing the values from the true vegetation signal [17][18][19]. In addition, snow presence affects the near-surface remote sensing GCC index; when snow starts melting; there is an increase in the GCC, but this is not related to vegetation growth itself [16].
The purpose of this study is to produce an FSC dataset from PhenoCam images that we can later use to validate and improve satellite-based measures of grassland phenology, specifically the dates defining the start and ends of annual growing seasons. We estimated FSC using PhenoCam images taken from the University of North Dakota (UND) Oakville Prairie Biological Field Station (Oakville) near Grand Forks. This research sought to answer the following questions: (1) can PhenoCam images be used to estimate FSC in a semiautomatic way?; (2) does the near-infrared band available in our PhenoCam imagery improve FSC estimation?; and (3) during five recent winters, 2015-2016 to 2019-2020, how did FSC vary in Oakville?
Below we review the use of remote sensing to monitor snow, as well as the area of study, its location, and its weather patterns. Then we describe PhenoCam data, our methodology, and the algorithm we developed to estimate FSC. We conclude that our algorithm could be used to estimate FSC with high accuracy using PhenoCam data in grassland settings.

Remote Sensing of Snow Cover
One of the most used methods in the remote sensing of snow is the Normalized Difference Snow Index (NDSI) [20]. NDSI was developed to allow for the separation of cloud cover from snow cover using satellite sensors [21]. However, the extent of snow cover is routinely under-estimated after processing satellite images [22]; for example, in the Northern Hemisphere between November and April can have an error of 8%, without considering the error caused by cloud cover [23].
Challenges in monitoring snow cover using satellite images include pixel size and temporal resolution of satellite measurements. The USGS Landsat program, the most commonly used source of remotely sensed data, has for four decades supplied images with a pixel size of 30 m × 30 m once every 16 days for most locations [24]. Landsat's relatively low temporal resolution presents a significant limitation when monitoring changes in snow cover. On the other hand, Sentinel-2, the European Space Agency (ESA) satellites constellation, have a higher spatial and temporal resolution but have less than 10 years of image availability [25]. With daily temporal resolution, Moderate Resolution Imaging Spectroradiometer (MODIS) images have coarse pixel sizes; when a MODIS pixel that is partially covered by 50% or more of snow is used for mapping it will be considered as fully snow-covered, overestimating the real value [26]. Furthermore, snow cover products derived from processed images by NASA-MODIS [27], ESA-Copernicus [28], and the Hydrological Satellite Facility project of the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) [29] are available online. Issues such as cloud presence or low temporal and spatial resolutions have been addressed using different sensors together. For example, Sentinel-2 and Landsat-8 were used to obtain high-resolution snow cover maps [30], and Li et al. summarized the different methodologies to improve the generation of cloud-free MODIS snow cover products [31].
Images from Unmanned Aircraft Systems (UASs) have been used to map snow cover with a better spatial resolution than satellite images. In one study, after performing an unsupervised classification, researchers found that the main cause for snow underestimation is shadows that fall on snow. The impact of shadows is strongest in the Red band and weakest in the Blue band [32].
Oblique terrestrial images have been used to analyze mountainous terrain [33], to estimate snow albedo [34], monitor snow cover, and make other measurements such as snow depth [35][36][37][38]. The use of this tool is suggested for snow cover because of its high reflectance values [39]. These images are also used for monitoring snow cover and its relationship with NDVI [40], the beginning of the growing season [41], and vegetation phenological stages [42]. Moreover, near-surface images show potential for monitoring events related to snow and glaciers. Daily oblique images were used to monitor events such as rainfall that cannot be seen in satellite images [43] and to monitor snow dynamics in glaciers [44].
Snow classification based on the Blue band threshold has been shown to be reliable compared with different satellite snow products, even when it has issues related to low sunlight and the presence of shadows [45][46][47][48][49][50][51][52]. Although not evaluated in this study, Pimentel et al. [53] proposed an algorithm that worked well in detecting FSC from images of rocky terrain at high altitude using a fixed terrestrial camera in Spain.

Area of Study
In 2017 in the U.S., there were an estimated 2.91 million km2 of grassland ( Figure 1). The U.S. is the country with the second-largest area of grassland after China [54], sinking an estimated 13 teragrams of carbon (Tg C) annually [55]. According to Hufkens et al. [12], grasslands in North America are projected to increase 18% in spatial extent by 2100.
Grasslands show highly dynamic growth that can vary over time and be affected by external factors [15]. Grasslands respond quickly in greening and physiological activity in response to climate drivers, especially precipitation and water availability [13,56] and, to a lesser extent, temperature [57]. Grassland productivity in the Great Plains is controlled by summer and spring precipitation [58]. reflectance values [39]. These images are also used for monitoring snow cover and its relationship with NDVI [40], the beginning of the growing season [41], and vegetation phenological stages [42]. Moreover, near-surface images show potential for monitoring events related to snow and glaciers. Daily oblique images were used to monitor events such as rainfall that cannot be seen in satellite images [43] and to monitor snow dynamics in glaciers [44]. Snow classification based on the Blue band threshold has been shown to be reliable compared with different satellite snow products, even when it has issues related to low sunlight and the presence of shadows [45][46][47][48][49][50][51][52]. Although not evaluated in this study, Pimentel et al. [53] proposed an algorithm that worked well in detecting FSC from images of rocky terrain at high altitude using a fixed terrestrial camera in Spain.

Area of Study
In 2017 in the U.S., there were an estimated 2.91 million km2 of grassland ( Figure 1). The U.S. is the country with the second-largest area of grassland after China [54], sinking an estimated 13 teragrams of carbon (Tg C) annually [55]. According to Hufkens et al. [12], grasslands in North America are projected to increase 18% in spatial extent by 2100.
Grasslands show highly dynamic growth that can vary over time and be affected by external factors [15]. Grasslands respond quickly in greening and physiological activity in response to climate drivers, especially precipitation and water availability [13,56] and, to a lesser extent, temperature [57]. Grassland productivity in the Great Plains is controlled by summer and spring precipitation [58]. Our study site is located in Grand Forks County, one of 53 counties in the State of North Dakota, part of the Northern Great Plains region that shares the U.S. portion of the area for tallgrass prairie research located 19 kilometers west of Grand Forks [59]. Oakville hosts one of the near-surface cameras of the PhenoCam Network [60]. Red River Valley of the North with Minnesota ( Figure 2). UND's Oakville is an important area for tallgrass prairie research located 19 kilometers west of Grand Forks [59]. Oakville hosts one of the near-surface cameras of the PhenoCam Network [60].  The climate of Grand Forks County is classified as Snow climate (D) fully humid (f) with warm summer (b), according to the updated Köppen-Geger climate classification [62]. Grand Forks' coldest month is January, and the warmest is July. The county typically receives snow from October to May (Table 1). Snowfall is also highly variable among years but exhibits an average trend of increasing amounts ( Figure 3). The climate of Grand Forks County is classified as Snow climate (D) fully humid (f) with warm summer (b), according to the updated Köppen-Geger climate classification [62]. Grand Forks' coldest month is January, and the warmest is July. The county typically receives snow from October to May (Table 1). Snowfall is also highly variable among years but exhibits an average trend of increasing amounts ( Figure 3).  Other important weather events in Grand Forks County are blizzards, defined as storms with wind speeds equal to or higher than 15.6 m/s [65]. During a period of 39 years (1979-2018), 100 blizzards were reported at the Grand Forks National Weather Service Forecast Office [64]. In North Dakota, the probability of a blizzard occurring annually is more than 50%, or more than one event every other year [66]. In the Red River Valley, the probability per year increases to between 61.6% to 76.4%, the highest in the U.S. [67].

Materials and Methods
Oblique terrestrial images have been widely used for monitoring the phenology of plants and snow cover. This section describes the steps followed to create an algorithm for estimating FSC in grasslands from a PhenoCam on flat terrain in eastern North Dakota. The workflow and accuracy assessment of the algorithm are also described.

Data
The PhenoCam in Oakville ( Figure 4) is a Type I (highest quality camera among the PhenoCam Network). The StarDot NetCam SC has a 5-megapixel resolution and records images every 30 minutes between 04:00 and 23:30 LST [60]. There are two types of images: 1) the Red, Green and Blue (RGB) image, and 2) the monochrome RGB + Near-Infrared (MoNIR) (Figure 4). Both types are taken by the PhenoCam every 30 minutes at nearly the same time. The entire archive of images for the Oakville Prairie, as well as all other Phe-noCam Network sites, are available online free of cost (for Oakville, see: https://phenocam.sr.unh.edu/webcam/sites/oakville/). Other important weather events in Grand Forks County are blizzards, defined as storms with wind speeds equal to or higher than 15.6 m/s [65]. During a period of 39 years (1979-2018), 100 blizzards were reported at the Grand Forks National Weather Service Forecast Office [64]. In North Dakota, the probability of a blizzard occurring annually is more than 50%, or more than one event every other year [66]. In the Red River Valley, the probability per year increases to between 61.6% to 76.4%, the highest in the U.S. [67].

Materials and Methods
Oblique terrestrial images have been widely used for monitoring the phenology of plants and snow cover. This section describes the steps followed to create an algorithm for estimating FSC in grasslands from a PhenoCam on flat terrain in eastern North Dakota. The workflow and accuracy assessment of the algorithm are also described.

Data
The PhenoCam in Oakville ( Figure 4) is a Type I (highest quality camera among the PhenoCam Network). The StarDot NetCam SC has a 5-megapixel resolution and records images every 30 min between 04:00 and 23:30 LST [60]. There are two types of images: (1) the Red, Green and Blue (RGB) image, and (2)  For each of the five winter seasons we studied (01 October to 30 April 2015-2016 to 2019-2020) we selected one PhenoCam image per day acquired between 10:30 a.m. and 12:30 p.m. LST to reduce solar illumination variations during the months that show snow presence. The majority of those we used were taken at 11:30 a.m. LST. A total of 1,712 images were downloaded: 856 RGB images and the corresponding 856 MoNIR images.

Validation of Optimal Band Selection Scenario for FSC Estimation
Previous work used the Blue band to calculate FSC. For example, [46] applied a threshold in the first local minimum above 127 in the Blue band histogram to map snow cover. In addition, they selected the Blue band after processing the three bands separately and got similar values. We also explored the utility of the MoNIR band for FSC calculation.

Validation of Optimal Band Selection Scenario for FSC Estimation
Previous work used the Blue band to calculate FSC. For example, [46] applied a threshold in the first local minimum above 127 in the Blue band histogram to map snow cover. In addition, they selected the Blue band after processing the three bands separately and got similar values. We also explored the utility of the MoNIR band for FSC calculation.
To select images that contained some snow cover, we calculated the average brightness of each band for all the images we downloaded. The oblique images from Oakville usually have a panel, a pole, information (name, date, and hour), and the sky ( Figure 5), all of which we excluded from the band brightness calculations. Calculating average brightness for each band for all images allowed us to understand which PhenoCam bands were most affected by the presence of snow cover.
To select images that contained some snow cover, we calculated the average brightness of each band for all the images we downloaded. The oblique images from Oakville usually have a panel, a pole, information (name, date, and hour), and the sky ( Figure 5), all of which we excluded from the band brightness calculations. Calculating average brightness for each band for all images allowed us to understand which PhenoCam bands were most affected by the presence of snow cover. In the first scenario, pixels with Digital Numbers (DNs) greater than or equal to 127 in the Blue band were classified as "snow", and pixels with lower Blue DNs were classified as "no-Snow" [46].

Scenario 2: Monochromatic RGB+NIR (MoNIR) threshold
The second scenario selected as snow all pixels with DNs greater than or equal to 127 in the MoNIR band. Depending on the land cover and the sunlight, the reflectance value may show differences for each band histogram ( Figure 6). Guided by our band brightness calculations, 38 images were selected from the entire dataset, representing 11 categories of FSC (0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%) for four periods (the winter seasons of 2015-2016, 2016-2017, 2017-2018, and 2018-2019). Three images with sastrugi present were also selected. We tested the images for calculating the FSC in four different scenarios: Scenario 1: Blue band threshold In the first scenario, pixels with Digital Numbers (DNs) greater than or equal to 127 in the Blue band were classified as "snow", and pixels with lower Blue DNs were classified as "no-Snow" [46]. Scenario 2: Monochromatic RGB+NIR (MoNIR) threshold The second scenario selected as snow all pixels with DNs greater than or equal to 127 in the MoNIR band. Depending on the land cover and the sunlight, the reflectance value may show differences for each band histogram ( Figure 6). Because shadowing on snow cover has been shown to confound efforts to calculate FSC from PhenoCam data, we developed two additional scenarios. The third and fourth scenarios, then, exploited the relationship between MoNIR and the Blue band calculated by: Because shadowing on snow cover has been shown to confound efforts to calculate FSC from PhenoCam data, we developed two additional scenarios. The third and fourth scenarios, then, exploited the relationship between MoNIR and the Blue band calculated by: The resulting DNs considered shadow on snow cover fell between −80 and 20. These were selected and used in the next two scenarios.
Scenario 3: Blue (B) band threshold + Shadow The selected values from Shadow were added to the Blue band selected pixels from the first scenario.
Scenario 4: MoNIR + Shadow The fourth scenario added the pixels from shadow to the pixels from scenario two.
To validate which scenario worked best for each FSC, we used the Accuracy Assessment tool in ArcGIS 10.6, which generates points for each class. For this study, we generated 50 points per image, 25 for each class (snow or no snow). Then, we visually compared the selected points to the original image to generate a confusion matrix between both classes.

Region of Interest (ROI)
We used the same area (ROI) on all downloaded PhenoCam images to calculate FSC. The ROI was delimited and measured in the field, and the resulting polygon was captured and digitized. The area of the polygon was calculated using Heron's Area formula [68]. The ROI has a real-world area of 141.45 m 2 , and its equivalent image area is 521,898 u 2 (pixels) (Figure 7).
The ROI has a real-world area of 141.45 m 2 , and its equivalent image area is 521,898 u 2 (pixels) (Figure 7).

FSC Algorithm
We developed our FSC algorithm using the Python 3 programming language. The installation of three modules was required: OpenCV, Matplotlib, and URLLIB.
First, the script we developed asks the user to enter the URL of the PhenoCam image to be processed, which allows the user to select the best image each day for FSC estima-

FSC Algorithm
We developed our FSC algorithm using the Python 3 programming language. The installation of three modules was required: OpenCV, Matplotlib, and URLLIB.
First, the script we developed asks the user to enter the URL of the PhenoCam image to be processed, which allows the user to select the best image each day for FSC estimation. Then, both images, the RGB and MoNIR, are downloaded, and the processing starts. Each band is isolated, and the average DN is calculated. The averages of the Red, Green, and Blue images are also calculated.
The Blue and MoNIR bands were normalized, and a Gaussian filter with an intensity of five is applied to their histograms. The Gaussian filter smooths the histogram curve by averaging pixels (5 × 5); this process helps to correct high or low values considered as noise (Figure 8).

FSC Algorithm
We developed our FSC algorithm using the Python 3 programming language. The installation of three modules was required: OpenCV, Matplotlib, and URLLIB.
First, the script we developed asks the user to enter the URL of the PhenoCam image to be processed, which allows the user to select the best image each day for FSC estimation. Then, both images, the RGB and MoNIR, are downloaded, and the processing starts. Each band is isolated, and the average DN is calculated. The averages of the Red, Green, and Blue images are also calculated.
The Blue and MoNIR bands were normalized, and a Gaussian filter with an intensity of five is applied to their histograms. The Gaussian filter smooths the histogram curve by averaging pixels (5 × 5); this process helps to correct high or low values considered as noise (Figure 8).  The Blue and MoNIR bands are used to calculate the shadow part of the image. The values between −80 and 20 are considered shadow. The next steps are based on decisions that the algorithm takes using knowledge from validation of the scenarios described. The workflow, also shown in Figure 9, is: i.
Images with a MoNIR average (M) greater than 135 are classified using the MoNIR band. The threshold value is the lowest between RGB average, M and 100. ii.
The Red band has higher values when there is no snow and the lowest values with snow presence, and this is inverse with the Blue band. A ratio between B/R is calculated. When the value is lower than one and B is lower than 90, a threshold of 127 is applied to the Blue band. iii.
When i and ii are false, a threshold is applied to the Blue band, whose value is the lowest among RGB average, Blue and 100. iv.
Steps i-iii produce an image with values of 0 (no snow) or 1 (snow). The new image is masked with the region of interest (ROI), and the FSC is calculated. v.
Sometimes, grassland has reflectance values as high as snow reflectance values. When the FSC value is lower than 10, the FSC is recalculated using a higher threshold value of 165. vi.
Images with FSC between 10 and 60 do not have to pass to any other classification. vii.
The values calculated with the shadow formula are added to the images with FSC higher than 60, which is the final FSC value.
Finally, the FSC is saved with the name of the image as a text file.  Figure 9. Image processing workflow. Figure 9. Image processing workflow.

FSC Algorithm Accuracy Assessment
The accuracy of the algorithm was calculated using the root-mean-squared-error (RMSE), which compared the estimated FSC values with FSC values determined by two specialists in image interpretation [42]. The RMSE was calculated with the following formula: The RMSE was calculated for two periods: 2017-2018 and 2018-2019 in both cases from 1 October to 30 April. The first period has a total of 177 images, and the second period 92 images. All the classified images were acquired at 11:30 LST except for two images with fog, which were changed to images acquired at 12:00 and 12:30 LST of the same day after the fog had lifted.
Gathering observed FSC values required help from visual image interpretation and PhenoCam experts outside the study. Both experts visually estimated FSC over the entire ground surface imaged, and not just within the ROI, for the winters of 2017-2018 and 2018-2019. The visual estimates of FSC were provided as values between 0% and 100%. It is important to mention the experts were instructed to estimate FSC at 10% intervals (0%, 10%, 20%...100%).
In addition, we selected two images for each of the five winters we studied (10 images total) to validate the classification. The selected images had an FSC estimated by our algorithm at between 40% and 60%. On each selected image, we randomly selected 50 points per class (two classes, "snow" vs. "no snow"; 100 points per image), resulting in 1,000 validation points. For each validation point, we visually compared the snow/nosnow result with the corresponding pixel from each PhenoCam source image, noting when our algorithm correctly and incorrectly identified "snow" or "no snow" pixels [40].

Results
Red and Blue reflectances are highly variable because of snow presence. When the ROI is 0% snow covered, it is 100% covered by grassland. The Red band had higher values when the images show vegetation and lower when there is snow present. This is in contrast to the Blue band that had lower values when there was no snow (Figure 10).

FSC Algorithm Accuracy Assessment
The accuracy of the algorithm was calculated using the root-mean-squared-error (RMSE), which compared the estimated FSC values with FSC values determined by two specialists in image interpretation [42]. The RMSE was calculated with the following formula: The RMSE was calculated for two periods: 2017-2018 and 2018-2019 in both cases from 1 October to 30 April. The first period has a total of 177 images, and the second period 92 images. All the classified images were acquired at 11:30 LST except for two images with fog, which were changed to images acquired at 12:00 and 12:30 LST of the same day after the fog had lifted.
Gathering observed FSC values required help from visual image interpretation and PhenoCam experts outside the study. Both experts visually estimated FSC over the entire ground surface imaged, and not just within the ROI, for the winters of 2017-2018 and 2018-2019. The visual estimates of FSC were provided as values between 0% and 100%. It is important to mention the experts were instructed to estimate FSC at 10% intervals (0%, 10%, 20%...100%).
In addition, we selected two images for each of the five winters we studied (10 images total) to validate the classification. The selected images had an FSC estimated by our algorithm at between 40% and 60%. On each selected image, we randomly selected 50 points per class (two classes, "snow" vs. "no snow"; 100 points per image), resulting in 1,000 validation points. For each validation point, we visually compared the snow/no-snow result with the corresponding pixel from each PhenoCam source image, noting when our algorithm correctly and incorrectly identified "snow" or "no snow" pixels [40].

Results
Red and Blue reflectances are highly variable because of snow presence. When the ROI is 0% snow covered, it is 100% covered by grassland. The Red band had higher values when the images show vegetation and lower when there is snow present. This is in contrast to the Blue band that had lower values when there was no snow (Figure 10).  The Red and Blue bands had different behaviors when there was no snow, but with snow present, they had similar curves. We assumed that the ratio (B/R) could be used to differentiate between snow and no-snow and used different threshold values in the Blue band image. The ratio was compared with the experts' FSC estimations. Figure 11 shows that in both cases, the R 2 was higher than 0.9, showing a strong relation between them. In most of the cases, when FSC = 0, the ratio value was lower than 1.0. The Red and Blue bands had different behaviors when there was no snow, but with snow present, they had similar curves. We assumed that the ratio (B/R) could be used to differentiate between snow and no-snow and used different threshold values in the Blue band image. The ratio was compared with the experts' FSC estimations. Figure 11 shows that in both cases, the R 2 was higher than 0.9, showing a strong relation between them. In most of the cases, when FSC = 0, the ratio value was lower than 1.0.

Validation of Optimal Band Selection Scenario for FSC Estimation
The Blue band yielded accurate results when FSC is between 0% and 20%. Grassland presence was reflected in the low Blue band amount. Between 60% and 100%, MoNIR + shadow had higher accuracy values, but they were not as high as expected. In the algorithm, we decided to apply shadow when the FSC was greater than 60% because that was when it showed better accuracy.
Pixel misclassification was caused by many factors. At the beginning of the season, grassland reflectance values were high in the Red and MoNIR bands. During snow presence, there was no constancy in the level of brightness. When it was bright, there were more shadow areas, and MoNIR and the Blue band had different behaviors. Sastrugi also caused misclassification because of the different shadows that it produced. Table 2 summarizes the accuracy of the 44 images.

Validation of Optimal Band Selection Scenario for FSC Estimation
The Blue band yielded accurate results when FSC is between 0% and 20%. Grassland presence was reflected in the low Blue band amount. Between 60% and 100%, MoNIR + shadow had higher accuracy values, but they were not as high as expected. In the algorithm, we decided to apply shadow when the FSC was greater than 60% because that was when it showed better accuracy.
Pixel misclassification was caused by many factors. At the beginning of the season, grassland reflectance values were high in the Red and MoNIR bands. During snow presence, there was no constancy in the level of brightness. When it was bright, there were more shadow areas, and MoNIR and the Blue band had different behaviors. Sastrugi also caused misclassification because of the different shadows that it produced. Table 2 summarizes the accuracy of the 44 images.

FSC Algorithm Accuracy Assessment
A total of 288 processed images were compared with visual estimations. These images are part of the winter of 2017-2018 and 2018-2019. For the first winter, 184 images were processed, and a Root Mean Square Error (RMSE) of 3.43% and 6.66% FSC were obtained, compared with Expert 1 and Expert 2, respectively. The second winter had 104 images, and the RMSEs were 7.01% and 14.71% FSC ( Table 3). The estimated FSC was closer to the experts' values, as they both had similar curves (Figure 12). Our pixel-based accuracy assessment showed that the overall accuracy of the images selected for validation was 92%.

FSC Algorithm Accuracy Assessment
A total of 288 processed images were compared with visual estimations. These images are part of the winter of 2017-2018 and 2018-2019. For the first winter, 184 images were processed, and a Root Mean Square Error (RMSE) of 3.43% and 6.66% FSC were obtained, compared with Expert 1 and Expert 2, respectively. The second winter had 104 images, and the RMSEs were 7.01% and 14.71% FSC ( Table 3). The estimated FSC was closer to the experts' values, as they both had similar curves (Figure 12). Our pixel-based accuracy assessment showed that the overall accuracy of the images selected for validation was 92%..   Figure 13 shows the five years of the estimated FSC from our PhenoCam. There was no distinct pattern in FSC in Oakville. Except for the winter of 2017-2018, the majority of winters showed snowfall events at the beginning and at the end that remained for a couple of days, followed by rapid snow melting.  images. This winter was different from the other winters in the study period because it had two snow periods, as was common in other regions (e.g. [69]). The first period was from 3 November to 23 November. Then, the snow melted for 10 days, and on 5 December, Oakville was again covered by snow for 132 consecutive days before it took just three days to melt entirely. The winter of 2018-2019 had one snowfall at the beginning and one at the end of the season. For this winter, a total of 104 images were processed. The first snowfall This winter was different from the other winters in the study period because it had two snow periods, as was common in other regions (e.g. [69]). The first period was from 3 November to 23 November. Then, the snow melted for 10 days, and on 5 December, Oakville was again covered by snow for 132 consecutive days before it took just three days to melt entirely. The winter of 2018-2019 had one snowfall at the beginning and one at the end of the season. For this winter, a total of 104 images were processed. The first snowfall was 10 October then, it melted, and the next event was after 22 days. The snow melted on 1 April, but there were two days with snow cover on 12 and 13 April. The winter of 2019-2020 has 190 images. This period had 64 snow-free days and three snowfall events before it was covered with snow for 110 consecutive days. At the end of the season, on 3 April, a snowfall remained on the grassland for seven days. Figures 14 and 15 show the obtained FSC compared with temperature, snowfall, and rainfall for the study period [63].
Remote Sens. 2021, 13, x FOR PEER REVIEW 16 of 23 was 10 October then, it melted, and the next event was after 22 days. The snow melted on 1 April, but there were two days with snow cover on 12 and 13 April. The winter of 2019-2020 has 190 images. This period had 64 snow-free days and three snowfall events before it was covered with snow for 110 consecutive days. At the end of the season, on 3 April, a snowfall remained on the grassland for seven days. Figure 14 and Figure 15 show the obtained FSC compared with temperature, snowfall, and rainfall for the study period [63].

Discussion
We presented a semi-automatic procedure to obtain the FSC in grassland cover, highlighting the potential of the PhenoCam in Oakville. All the findings were exclusive for our grassland cover and for winter weather conditions. Despite the use of just the Blue band in previous research, we focused on the other spectral bands that could improve the FSC estimation. Assuming that there is no change in elevation, we also did not use a Digital Elevation Model (DEM) to georeference the image and calculate the total area.
One of the main reasons for the need for this high temporal resolution data is the variability of snow cover. It takes one day to go from 0 to 100 FSC and sometimes less than five days for snow to completely melt ( Figure 16).

Discussion
We presented a semi-automatic procedure to obtain the FSC in grassland cover, highlighting the potential of the PhenoCam in Oakville. All the findings were exclusive for our grassland cover and for winter weather conditions. Despite the use of just the Blue band in previous research, we focused on the other spectral bands that could improve the FSC estimation. Assuming that there is no change in elevation, we also did not use a Digital Elevation Model (DEM) to georeference the image and calculate the total area.
One of the main reasons for the need for this high temporal resolution data is the variability of snow cover. It takes one day to go from 0 to 100 FSC and sometimes less than five days for snow to completely melt ( Figure 16).

Discussion
We presented a semi-automatic procedure to obtain the FSC in grassland cover, highlighting the potential of the PhenoCam in Oakville. All the findings were exclusive for our grassland cover and for winter weather conditions. Despite the use of just the Blue band in previous research, we focused on the other spectral bands that could improve the FSC estimation. Assuming that there is no change in elevation, we also did not use a Digital Elevation Model (DEM) to georeference the image and calculate the total area.
One of the main reasons for the need for this high temporal resolution data is the variability of snow cover. It takes one day to go from 0 to 100 FSC and sometimes less than five days for snow to completely melt ( Figure 16). Other studies have used the Blue band because the three bands have similar behaviors in mountainous regions [19]. Figure 17 shows that in separate cases of 100% FSC, the three bands have different histograms that could be related to sunlight. We highlighted that previous studies did not have MoNIR band, and implementing it in our study allowed us to have fewer cases of misclassification because of shadows. Other studies have used the Blue band because the three bands have similar behaviors in mountainous regions [19]. Figure 17 shows that in separate cases of 100% FSC, the three bands have different histograms that could be related to sunlight. We highlighted that previous studies did not have MoNIR band, and implementing it in our study allowed us to have fewer cases of misclassification because of shadows. Furthermore, previous studies have used 127 as the minimum value in the Blue band because they found that areas with snow cover in this band have a value greater or equal to 127. In this research, we used the Blue/MoNIR average, and if these values were greater than 100, the threshold value was set at 100. Figure 18 shows that in this region, the Blue band value, when there is snow present, was greater than 80. The difference between these values shows that depending on the region (mountainous or a valley) and the land cover (forest, grassland, etc.), the Blue band will have different behavior.  Furthermore, previous studies have used 127 as the minimum value in the Blue band because they found that areas with snow cover in this band have a value greater or equal to 127. In this research, we used the Blue/MoNIR average, and if these values were greater than 100, the threshold value was set at 100. Figure 18 shows that in this region, the Blue band value, when there is snow present, was greater than 80. The difference between these values shows that depending on the region (mountainous or a valley) and the land cover (forest, grassland, etc.), the Blue band will have different behavior. Other studies have used the Blue band because the three bands have similar behaviors in mountainous regions [19]. Figure 17 shows that in separate cases of 100% FSC, the three bands have different histograms that could be related to sunlight. We highlighted that previous studies did not have MoNIR band, and implementing it in our study allowed us to have fewer cases of misclassification because of shadows. Furthermore, previous studies have used 127 as the minimum value in the Blue band because they found that areas with snow cover in this band have a value greater or equal to 127. In this research, we used the Blue/MoNIR average, and if these values were greater than 100, the threshold value was set at 100. Figure 18 shows that in this region, the Blue band value, when there is snow present, was greater than 80. The difference between these values shows that depending on the region (mountainous or a valley) and the land cover (forest, grassland, etc.), the Blue band will have different behavior.  One of the biggest issues related to the FSC from near-surface remote sensing is shadows. These can create misclassifications and significantly decrease accuracy. The use of the MoNIR band helps to improve the classification (Figure 19). One of the biggest issues related to the FSC from near-surface remote sensing is shadows. These can create misclassifications and significantly decrease accuracy. The use of the MoNIR band helps to improve the classification (Figure 19).

Conclusions
This research used PhenoCam images in a semi-automatic way to estimate FSC at a site within Oakville, showing that the capabilities of the images are not only limited to monitor vegetation phenology. The results validated with two experts had average RMSEs of 5.02% and 10.32%. Our pixel-based accuracy assessment showed that the overall accuracy of the images selected for validation was 92%. The number of cameras of the PhenoCam Network makes it possible to provide localized FSC estimates over the entire country, generating a database that can be used to validate and complement satellite data.
Most of the images are classified using the Blue band, but thanks to the MoNIR band, the algorithm recognizes darker areas (shadows), and also, some days when the Blue band has a low reflectance value, the MoNIR band showed that it could be used to classify snow cover. Not all the cameras in the PhenoCam Network include the MoNIR band, and these sites could use the Blue band to estimate the FSC.
During the study period (2015-2020), there was not a consistent pattern in the FSC in this area. Some years the first snowfall was on 1 December and other years it was on 9 October. The same happened with the snowmelt; in 2016, the last day with snow present was in March, while other years had their last snowfall in April. FSC is a parameter that depends on snowfall, temperature, energy flux, and their anomalies.
Some limitations of this research were the field of view, the oblique view, and the vignetting effect; all of these related to the sensor. The field of view, which determines the area in the image, compared with other sensors, was small. The oblique view generates pixels that represent different real sizes; the closer the pixel, the smaller the area of ground represented. Vignetting happens when the borders of the image have darker colors compared with the center of the image.

Conclusions
This research used PhenoCam images in a semi-automatic way to estimate FSC at a site within Oakville, showing that the capabilities of the images are not only limited to monitor vegetation phenology. The results validated with two experts had average RMSEs of 5.02% and 10.32%. Our pixel-based accuracy assessment showed that the overall accuracy of the images selected for validation was 92%. The number of cameras of the PhenoCam Network makes it possible to provide localized FSC estimates over the entire country, generating a database that can be used to validate and complement satellite data.
Most of the images are classified using the Blue band, but thanks to the MoNIR band, the algorithm recognizes darker areas (shadows), and also, some days when the Blue band has a low reflectance value, the MoNIR band showed that it could be used to classify snow cover. Not all the cameras in the PhenoCam Network include the MoNIR band, and these sites could use the Blue band to estimate the FSC.
During the study period (2015-2020), there was not a consistent pattern in the FSC in this area. Some years the first snowfall was on 1 December and other years it was on 9 October. The same happened with the snowmelt; in 2016, the last day with snow present was in March, while other years had their last snowfall in April. FSC is a parameter that depends on snowfall, temperature, energy flux, and their anomalies.
Some limitations of this research were the field of view, the oblique view, and the vignetting effect; all of these related to the sensor. The field of view, which determines the area in the image, compared with other sensors, was small. The oblique view generates pixels that represent different real sizes; the closer the pixel, the smaller the area of ground represented. Vignetting happens when the borders of the image have darker colors compared with the center of the image. A significant limitation was the subjectivity introduced into the accuracy assessment because we relied on the visual estimates of FSC made by two experts. Although we generally followed the accuracy assessment methods described by other research [50], we also added a pixel-based assessment, following methods described by [41]. In future studies, our goal is to tie our PhenoCam observations to corresponding data gathered by satellite sensors. We support the idea that datasets should be combined to achieve higher quality and less uncertainty in the remote sensing of snow cover [70].
Other limitations are external factors like weather conditions, land cover, and the presence of objects out of the study. On a sunny day, the scene is brighter, but there are also more shadows, and on cloudy days, the images have relatively low contrast. Blizzards and fog decrease visibility, and strong winds generate sastrugi. The land cover at Oakville is grassland, which during the senescence period has a golden color and sometimes is as bright as snow, creating misclassification.
FSC is an important parameter that has to be monitored. The use of PhenoCam images and Python language, both free, allow researchers to keep developing scripts and to improve the existing work. This research was focused on one grassland site; future work should explore extending the findings to additional PhenoCam Network sites, working with other land covers and at different latitudes. Furthermore, future work should include the relationship between FSC and other parameters, such as positive and negative feedbacks related to climate change, the beginning and end of the growing season, and soil temperature and moisture.