Two field sites, the University of New Hampshire's Thompson Farm Research Observatory (43.1089° N, 70.9485° W) and Kingman Farm Research Observatory (43.17207° N, 70.92662° W), are located in southeastern New Hampshire, United States. The sites are separated by approximately 8 km. Both farms are actively managed and operated by the University of New Hampshire. Analyses were performed on the open field portions of the sites indicated by the hatched areas in Fig. 1.

The Thompson Farm study field is approximately 7.6 ha with elevations ranging from 10 to 35 m. The site has limited topographic relief with gentle slopes where low-lying areas within the field and clayey soils result in areas of surface water collection and icing during the winter months (Jacobs et al., 2021). Field vegetation varies between low-cut pasture grass and tall (∼ 1 m) unkempt grasses. The surrounding mixed deciduous and coniferous forest canopy is 10 to 30 m tall (Jacobs et al., 2021). Most of the field experiences two to 5 h d⁻¹ of shade with up to 8 h per day at the forest-field boundary in the north portion of the field (Cho et al., 2021). An access road, running south to north, is the high point between the two fields with the eastern field being lower than the western field. Soils at Thompson Farm are classified as silt loam, sandy loam, and loamy sand in the National Cooperative Soil Survey (NCSS).

The Kingman Farm site consists of two prominent field sections, the main field and a sheltered field, that were the focus of this study. The main field is approximately 17 ha with gentle rolling terrain and maximum elevation changes of ∼ 10 m. This field section is largely exposed with a mixed forest (deciduous/coniferous) bordering the northern edge. An access road runs east to west with the terrain on the southern side of the field. Field vegetation is primarily low-cut pasture grass with a few small sections of tall (∼ 1 m) unkempt grass along the access road. The main field soil is classified as fine sandy loam, silt loam and loamy fine sand. The sheltered field section at Kingman Farm is a smaller (∼ 0.7 ha) managed field surrounded by forest. The northern edge of the field is bordered by tall conifers (∼ 30 m), while the southern edge is bordered by deciduous trees and an access road. The sheltered field is relatively flat with elevation variations of approximately 5 m with the highest elevations near the center of the field. NCSS classifies soil in this field as fine sandy loam.

The Thompson Farm and Kingman Farm sites are located within the ephemeral snow class, which is characterized by relatively warm temperatures, shallow snow, and a total duration of snow cover lasting for an average of 70 d per year (Sturm and Liston, 2021; Johnston et al., 2023). Based on a 20-year (2002–2021) winter climatology at the co-located U.S. Climate Reference Network sites (USCRN, Thompson site ID: 1040, Kingman Site ID: 1041), January is the coldest month with an average temperature of -4.1 °C followed by February (-3.1 °C) and December (-1.2 °C). These are the only months having average temperatures below 0 °C. From December to February, monthly mean high temperatures range from 0.6 to 3.4 °C with lows ranging from -9.3 to -6.0 °C. The average monthly liquid equivalent precipitation is highest in December (122 mm) and lowest in January (78 mm). February, March, and April have approximately 100 mm of precipitation on average (98 to 108 mm). The sites' combination of sub-freezing temperatures and regular precipitation results in frequent snowfall. Conversely, because above-freezing temperatures are also common, there are intermittent melting and rain-on-snow events throughout the winter. Based on fSCA derived from MODIS observations from 2002–2021 (Hall and Riggs, 2021), snow cover occurs in November (average monthly fraction snow covered area is 0.02), December (0.36), January (0.70), February (0.73), March (0.48), and April (0.04).

To derive binary SCA from UAS LiDAR snow depth, we applied a series of snow-depth thresholds to the 13 LiDAR-derived snow depth fields that were flown coincidentally with the optical. We tested thresholds ranging from 0 to 10 cm, in increments of 2 cm, and classified pixels with snow depth exceeding the threshold as snow-covered. This sensitivity analysis was designed to evaluate the influence of threshold choice on derived fSCA, particularly under shallow and patchy snow conditions where LiDAR depth uncertainty can be comparable to observed snow depths.

The performance was evaluated for seven classification configurations for snow-covered area mapping with 26 orthomosaics using the P4 UAS platform used in the reference dataset. The results clearly separated the tested configurations into two performance groups (Fig. 4). The MLE, RF, and SVM classifiers emerged as the top performers, achieving accuracy, balanced accuracy, and F1 score values exceeding 0.95, indicating strong agreement and classification accuracy. Their high performance was further supported by Cohen kappa values exceeding 0.89. While SVM Blue Band and Static Blue Band classifiers consistently underperformed the top three classifiers, a single blue band still yields reasonable results (balanced accuracies of 0.86 and 0.84). While the Dynamic Blue Band classifier produced high accuracy values for both thresholds (0.95 for a threshold of 90 and 0.93 for 127), the classifier had the poorest performance for the other metrics. It should be noted that for the Dynamic Blue Band classifier, the reported Cohen's Kappa represents an average across individual orthomosaics. However, for orthomosaics containing only a single class (e.g., fully snow-covered or completely snow-free scenes), Cohen's Kappa becomes unreliable, as it cannot account for agreement beyond chance in the absence of class variability. Overall, all tested classification configurations have the capacity to classify SCA with the Maximum Likelihood, RF, and SVM being the most robust classifiers.

Statistical comparison using the Friedman test revealed significant differences in both balanced accuracy (χ2= 32.82, p< 0.001) and F1 score (χ2= 31.62, p< 0.001) among the evaluated classifiers. This indicates that classifier choice had a significant impact on classification performance. Post-hoc Wilcoxon signed-rank tests on both balance accuracy and F1 score (Tables S4 and S5) indicated that MLE, RF and SVM achieved statistically comparable performance (adjusted p> 0.05), whereas both blue-band-only approaches differed significantly from the higher performing classifiers (adjusted p< 0.05). Although the SVM blue band and static blue band classifiers showed statistically comparable F1 scores, their balanced accuracies statistically differed significantly.

The disagreement maps suggested that the presence of shadows leads to the highest disagreement among models (Di> 0.4). One example of an image affected by shadows was acquired on 23 December 2020 at the TF study site. The disagreement map of this date (Fig. 5), where approximately 20 % of the area exhibited disagreement values greater than 0.4, indicates that at least three of the seven tested models disagreed with the majority vote in those locations. The models contributing to this disagreement were MLE, RF, and SVM, which correctly identified snow in shaded areas, whereas the other models classified these shaded snow pixels as snow-free.

When fSCA was high (> 90 %) and shadow effects were minimal, the models showed strong agreement across nearly the entire study area. An example of this condition is shown in Fig. 6. However, during melt conditions (fSCA < 0.4) and in scenes without shadows, disagreement between at least one model and the majority classification occurred more frequently across the study area compared to full snow-cover conditions. Figure 7 illustrates one such example.

This experiment used subsets of 6 to 25 orthomosaics from the complete set of 26 orthomosaics to train and evaluate new MLE, SVM, and RF models (maximum 500 combinations). The performance was calculated for the individual orthomosaics that were “held-out” from the dataset, and an overall average of these held-out orthomosaics is presented here. Figure 8 shows that reducing the number of orthomosaics used for training only modestly reduces the median combination's performance with mean accuracies remained consistently high (at or above 0.9) for all three classifiers. However, reducing the number of orthomosaics does lead to increased variability in performance among the combinations.

All three classifiers performed comparably when trained on larger datasets (> 20 orthomosaics) and had reduced ability on certain orthomosaics when trained on small datasets (< 12 orthomosaics). RF demonstrated the weakest generalization performance on mid-sized datasets. Even when most combinations performed well, certain combinations of training and testing datasets led to reduced performance. For models trained with 20 to 12 orthomosaics, fewer than 0.5 % of MLE and SVM models had held-out accuracies below 90 %, whereas approximately 4 % of RF models fell below the 90 % accuracy.

Even when the average of all held-out orthomosaics is high, some individual orthomosaics may have reduced performance (Fig. S5). Among the three classifiers, the RF classifier is slightly more sensitive to training data availability. RF models' median held-out accuracies declined from 0.97 to 0.90 as the number of training orthomosaics decreased, whereas MLE and SVM only decreased from 0.97 to 0.95 and 0.98 to 0.94, respectively. The RF models also had individual orthomosaic accuracies below 50 % even when 18 orthomosaics were used for training (Fig. S5). When the SVM and MLE training set included more than 14 orthomosaics, accuracy for nearly all excluded orthomosaics exceeded 80 % (Fig. S5), though the SVM models had more low accuracy orthomosaics than MLE models. When the MLE and SVM training data sets contained about 10 or fewer training orthomosaics, some of the held-out orthomosaics had accuracies below 50 %, with a few extreme cases where accuracies were as low as 0 %.

Individual orthomosaics' performance were examined. Figure 9 shows each individual orthomosaic's classification consistency when it was excluded from the training dataset plotted as the proportion of time that the classification accuracy was less than 80 % by orthomosaic. While most scenes were reliably predicted with accuracies greater than 80 % for more than 95 % of the combinations, some scenes were consistently more difficult to classify. For the six lowest accuracy orthomosaics, the MLE models performed better than either the SVM or the RF models. RF models have lower accuracies than either the MLE or SVM models except for the 13 January 2025 orthomosaic from Kingman Farm. For example, the RF models' accuracies for the 12 February 2024 orthomosaic of Kingman Farm were below 80 % for nearly 40 % of the combinations when that orthomosaic was held out from the training dataset, but for the same orthomosaic, MLE and SVM accuracies were seldom (10 %) below 80 %.

To better understand which conditions led to reduced performance, we compared fractional snow-covered area (fSCA), snow wetness and cloud cover conditions for the ten highest- and lowest-performing models for training datasets ranging from 6 to 12 orthomosaics. fSCA for each orthomosaics was estimated using the models from the base experiment (i.e., models trained on the full dataset; see Fig. S6). For each condition (i.e., fSCA, wetness, and cloud cover), we then compared its mean values from the dates included in the training sets with those of the corresponding held-out orthomosaics for the ten highest performing and up to 10 lowest-performing when less 80 % accuracy models.

For fSCA, the results for the SVM classifier are presented in Fig. 10, corresponding analyses for MLE and RF are shown in the Supplement (Figs. S7 and S8). For datasets having 10 to 12 orthomosaics, average fSCA ranges from about 0.5 to 0.8 for both the training and testing datasets. Note that this size training dataset has no models with less 80 % accuracy. As the dataset size decreases below ten, the highest-performing models still have fSCA values from 0.5 to 0.8, but the models having less 80 % accuracy used training datasets with average fSCA values higher than 0.8. In addition, the held-out orthomosaics had lower fSCA values, sometimes dropping below 50 %. This reveals that the poorest performing models used training data are dominated by snow-covered orthomosaics (often > 70 % fSCA), then used to predict conditions with lower or more variable snow coverage, Thus, there is a mismatch between the mean fSCA values of their training (included) and test (held-out) orthomosaics. We did not find any evidence that differences in cloud cover (Figs. S9, S10, S11) or snow wetness (Figs. S12, S13, S14) between training and held-out orthomosaics influenced performance.

Figure 11 compares fSCA estimated from lidar snow depth thresholding with mean fSCA derived from optical observations and the MLE, RF and SVM classifiers. This analysis indicates that during periods of deeper snowpack, when average snow depth exceeded 10 cm (Table S2), fSCA estimates derived from different snow depth thresholds closely matched the optical snow cover estimates, with differences among thresholds remaining small (< 5 %). As shown in Fig. 11, threshold-based estimates converge during periods of near-complete snow cover. In contrast, divergence among thresholds emerges during shallow or patchy snow conditions. For example, on 26 February 2021, the 0 cm threshold indicated nearly complete coverage (98 %), while progressively higher thresholds reduced estimated fSCA to ∼ 80 % (6 cm) and ∼ 30 % (10 cm). The optical-derived fSCA (95 %), used here as the reference estimate, confirms the near-complete spatial coverage of a shallow snowpack on this date.

Similar threshold sensitivity occurred on 3 and 7 March 2021, when increasing the snow depth threshold from 0 to 10 cm reduced the estimated fSCA by more than 60 percentage points. The optical reference fSCA on these dates was 82 % and 72 %, respectively, indicating discontinuous snow cover during the melt stage. During these periods, increasing the minimum snow depth threshold systematically reduced the estimated fSCA, producing a pronounced gradient among thresholds. The spatial cross-classification maps for 7 March 2021 (Fig. 12) further showed that these reductions were concentrated along snow transition boundaries, near fragmented snow cover or in localized shallow snow regions. At low thresholds, lidar derived maps identified extensive shallow snow areas that corresponded to optically detected snow cover, whereas high thresholds increasingly classified these regions as snow free. In contrast at the 0 cm threshold, the lidar indicated localized snow presence in areas that appeared snow free in the optical imagery (Fig. 12), which is also reflected in the higher lidar derived fSCA relative to the optical reference shown in Fig. 11. One possible explanation is that small positive elevation residuals associated with residual vegetation, surface roughness, or registration uncertainty were retained within the snow class at near zero depth thresholds. Another is that in these open areas, the 1 m pixels in the snow transition boundaries had a point cloud density that resolved variations within a single pixel. This behavior became more apparent as snow free patches expanded during the melt period (Figs. 11, S15). Although no independent surface characterization was available to verify these features, the spatial distribution of these lidar false positive snow detections suggests that near zero lidar snow depth thresholds may be sensitive to non-snow surface structure, potentially leading to overestimation of lidar derived fSCA.

Our primary goal was to identify one or more classifiers that can consistently and accurately map SCA in open areas. The baseline assessment of tested classifiers established that the three classifiers (MLE, RF, and SVM) using full RGB inputs to be more effective classifiers for mapping SCA than classifiers relying solely on the blue band. Their ability to differentiate between snow-on and snow-free conditions for 3 cm pixels exceeded 95 % for accuracy, balanced accuracy, and F1 score performance metrics. In comparison to previous studies in open areas, this study's performance matches the performance reported by Eker et al.'s (2019) overall accuracy of 95 % and exceeds Belmonte et al.'s (2021) accuracy of 90.2 % using a RF classifier, Cannistra et al.'s (2021) balanced accuracy of 0.82 using a convolutional neural network-based method and Johnston et al.'s (2025) 80 % to 88 % accuracy via k-means clustering. Taken as a whole, the UAS performance appears to be similar to satellite snow cover products' performance which typically range from about 85 % to 95 % (Ault et al., 2006; Hall and Riggs, 2007; Gao et al., 2010; Nolin, 2010; Frei et al., 2012; Coll and Li, 2018) with the UAS identifying SCA at a much finer resolution (cm scale).

Classifiers relying solely on the blue band, such as SVM Blue Band, Static Blue Band Thresholding, and Dynamic Blue Band Thresholding, still achieved good results, but were consistently less accurate than their RGB-based counterparts. Of the blue band classifiers, Static Blue Band Thresholding performed the best based on its F1 Score of 0.90, balanced accuracy of 0.86 and accuracy of 0.87. While the Dynamic Blue Band Thresholding had an accuracy (0.95) comparable to the MLE, RF, and SVM, this is largely due to its strength in capturing snow when fSCA is high, its balanced accuracy of 0.78 reflects its diminished ability to capture mixed conditions. The dynamic thresholding approach generally underestimates snow-covered pixels, as indicated by a higher rate of incorrectly classifying of snow free pixels than snow covered pixels (Tables S6 and S7). While this classifier is designed to adapt to varying illumination conditions unique to each orthomosaic, its effectiveness diminishes when no local minimum is detected in the histogram (e.g., fully snow-covered days) making its performance highly dependent on the pre-defined minimum threshold. Figures S16 and S17 clearly demonstrate how the value of the pre-defined minimum threshold can influence classifier performance when the intensity histogram lacks a clear local minimum. Moreover, the dynamic thresholding approach can be memory-intensive when constructing histograms using all pixels from high-resolution orthomosaics for large study areas. As an alternative, random sampling of the orthomosaic pixels may be used to reduce memory requirements but would likely further reduce classification performance due to less representative histogram estimation. Despite its simplicity, the Static Blue Band classifier is preferable to the dynamic approach. Where reduced performance is an acceptable trade-off for increased ease of use and users are knowledgeable about conditions that may be problematic, the blue band threshold classifier appears to offer a reasonable compromise that merits additional vetting.

The clear difference between using classifiers with three bands as compared to the single blue-channel is somewhat unexpected because the blue-channel is a widely used classifier. As reviewed by Webster and Jonas (2018), the use of the single blue-channel is the favored method for RGB imagery from field cameras because blue light is reduced in the canopy view fraction (e.g. Reid and Essery, 2013; Frazer et al., 1999) and saturates over most snow-covered scenes causing loss of sun-lit/shadow contrast (Wolter et al., 2012; Dozier, 1989). Using UAS RGB imagery, Jacques-Hamilton et al. (2025) reporting satisfactory results by thresholding the blue-channel values following Salvatori et al.'s (2011) approach for mapping snow cover using field camera images. Their blue pixel intensity threshold value of 90 was determined using a qualitative approach in which they adjusted the threshold until pixels classified as snow cover agreed with what they saw in their photos. This is considerably lower than the threshold value of 127 used for field cameras (e.g. Salvatori et al., 2011; Portenier et al., 2020; Caparoì Bellido and Rundquist, 2021). Our results suggest that while the blue band is highly sensitive to snow due to its strong reflectance properties (Nolin, 2010), combining it with the red and green bands provides more reliable snow classification, particularly when mapping shadowed snow, dirty snow, and wet snow. Figure S18 shows examples of SCA maps on a partially snow-covered day and a fully snow-covered. It appears that the reduced blue band reflectance caused by shadows, impurities, or melting led to misclassification of snow-covered surfaces leading to a reduction in the field's fSCA values by 15 % to 25 % in comparison to fSCA that was properly captured by MLE, RF, and SVM.

While MLE, SVM and RF classifiers produced comparable baseline performance, the assessment of their generalization performance revealed that RF models may be more sensitive to some surface conditions, possibly due to its reliance on multiple decision trees that are impacted by the absence of key features in outlier orthomosaics. Conversely, SVM and MLE demonstrated relatively more stable behavior, indicating that they may be less prone to overfitting and could offer better generalization for snow detection when the training set is not representative of a new orthomosaic's conditions. MLE, while simple, seems to perform consistently in the majority of cases.

Our generalization experiment offered insights into how both the number and characteristics of orthomosaics influence the development of a robust model, highlighting the challenges posed by data availability and diverse snow cover conditions when mapping snow using UAS RGB imagery and machine learning classifiers. For dynamic winter conditions similar to those in our study area (characterized by an ephemeral snowpack) at least 12 orthomosaics spanning a range of snow cover conditions was enough to result a reasonable classification accuracy (exceeding 80 %) when employing robust classifiers such as MLE, SVM, and RF. However, beyond just the number of orthomosaics, such a dataset should be representative of a diverse range of snow cover conditions that pose challenges for snow mapping. Key examples of challenging conditions include:

Thin snow: The two orthomosaics with the highest frequency of accuracies below 80 % (13 January 2025, KF and 6 February 2024, KF) in the generalization experiment also exhibited the lowest median accuracy across all models (Fig. S19). Although both orthomosaics were predominantly snow-covered, with fSCA exceeding 60 %, the observed snow depth was relatively shallow (3 to 4 cm). This consistently poor generalization performance, regardless of classifier, underscores the inherent difficulty of detecting shallow snow cover using only optical imagery. These findings are not surprising for these optically thin snowpacks. In the visible wavelengths, where there is little absorption and scattering is dominant, light penetrates snow by 10's of centimeters beneath the surface (deeper for blue than red). If snow is shallower than this then almost certainly the observed signal is being darkened by the ground below. In contrast, methods that use the NIR, where ice absorption/grain size dominates, would be expected to perform better because light does not penetrate below 1–2 cm of snow. That said, our experiment used numerous other orthomosaics taken when snowpacks are equally thin. In most cases, these thin snowpacks were successfully classified, suggesting that even if the ground is making the snow appear darker than it actually is, it still is brighter than most other surface types in this mapping area. Special cases such as shallow and wet snow or snow that is shallow and shadowed require additional study (McKenzie Skiles, personnel communication 19 March 2025).

Finally, several considerations should be taken into account when applying RGB classifiers for operational applications. First, for transfer of trained classifiers across sensors, attention should be given to differences in sensor radiometric characteristics. We note that in this study, the RGB reflectance of the P4 sensor spanned the full dynamic range (i.e., 0–255). In cases where a sensor exhibits reduced sensitivity in a spectral channel or produces compressed or biased RGB values, it is critical to apply bias correction to the intensity values prior to classification to ensure comparability and preserve model performance. Second, it should be noted that SCA mapping using UAS RGB imagery faces limitations under continuous, deep snow cover conditions. Under such conditions, orthomosaic generation from visible imagery can become challenging because structure-from-motion algorithms rely on identifying stable visual features to establish tie points between overlapping images. Uniform snow surfaces often lack sufficient texture and contrast, which can reduce reconstruction quality and, in some cases, prevent reliable orthomosaic generation. This limitation primarily affects image reconstruction rather than the snow-classification methods evaluated here. Future drone-based snow-mapping campaigns targeting continuous snow cover may therefore require alternative acquisition strategies, such as the inclusion of artificial ground control features, lower flight altitudes, or the integration of complementary sensing technologies.

Our analysis of SCA mapping using lidar data highlighted that in ephemeral and transitional snow regimes, where shallow but spatially continuous snow cover frequently occurs, small changes in the minimum depth threshold can therefore produce large differences in estimated snow-covered area under such conditions. This sensitivity is especially important when snow depths approach the measurement uncertainty of the lidar product (∼ 2 cm in our case; Harder et al., 2020; Jacobs et al., 2021), as residual errors may be incorrectly classified as snow and lead to overestimation of SCA. Consequently, threshold selection should be guided by the uncertainty characteristics of the snow-depth product, spatial resolution, and the prevailing snow regime, with particular care in environments characterized by shallow, discontinuous, or rapidly evolving snow cover.

Although this study focuses on open-field environments to enable a controlled comparison of SCA mapping approaches, extending the analysis to forested environments remains an important direction for future work. To explore this possibility, we masked the forest canopy and applied the RF, SVM, and MLE models developed in the open-field analysis to canopy gaps within the forested portions of the Kingman Farm study site orthomosaics (See Supplement for methodological details, Fig. S21). When applied to canopy gaps, the models' accuracy, balanced accuracy, and F1 scores were comparable to those obtained in the open areas, while Cohen's kappa values decreased modestly from ∼ 0.90 in open terrain to ∼ 0.80 within forest openings (Fig. 13).

Although the models perform well in canopy gaps, reliable SCA mapping in forested environments depends on the availability of a consistent and accurate canopy masking approach. This is particularly challenging in landscapes dominated by deciduous forests. Using a k-medoids clustering algorithm, we generated canopy masks when the forest floor was snow-covered; however, we encountered several limitations that are likely common to most unsupervised classification approaches. In particular, these methods cannot reliably distinguish branches lying on the ground from branches that remain elevated within the canopy. In deciduous forests, branch structures are sufficiently fine that modest shifts in canopy position due to wind or snow loading, as well as minor alignment differences between UAS orthomosaics, prevent masks generated for one date from being reliably applied to other dates. These issues highlight an important limitation of relying solely on optical imagery for SCA mapping in forested environments.

Additional challenges arise from branches and litter present on the snow surface, which can reduce classification reliability and complicate identification of true canopy gaps where snow accumulates on the ground. These effects are particularly pronounced in deciduous forests under leaf-off conditions. Furthermore, snow intercepted and retained on branches cannot be easily distinguished from snow on the ground using optical imagery alone. Together, these limitations highlight the challenges of relying on optical data for snow mapping in forested landscapes and motivate the exploration of existing satellite fSCA and NDSI methods which use canopy cover information and integration of complementary observations, such as lidar, which can provide additional information about sub-canopy snow distribution.