All experiments throughout this study use snow data collected during the MOSAiC expedition (October 2019–September 2020) . The snow pit measurements conducted include SMP profiles, micro-computer tomography (Micro-CT) , and near-infrared (NIR) photographs . Collecting snow profiles on Arctic sea ice is especially challenging because (i) only a few hours were available to perform all measurements within one snow pit and (B) the measurements must be conducted with wind velocities up to 25 m s-1 and temperatures of -30∘C. Changing personnel, i.e., different operators, were conducting the snow pit measurements. As a result, traditional stratigraphy analysis and in situ snow grain classification from snow pits carry operator biases. could measure only 27 snow pits with stratigraphy under similar conditions.

In contrast, during the MOSAiC expedition, several thousand (3680) SMP profiles were collected. Out of the 269 snow pit events that included SMP measurements, 102 had NIR measurements and 103 had micro-CT profiles collected simultaneously. A total of 71 snow pit events had all three measurements (SMP, NIR, and micro-CT). We encountered 8 different snow types. Refer to for descriptions of the different snow types referenced here and a classification guideline for snow particles that were visually observed.

refer only to visually observed snow grains and not to SMP signals.

The main measurements collected were signal profiles from the snow micropenetrometer since it provides profiles quickly with little physical labor and independent of the person that measures them. Of the 3680, 164 profiles from the cold season (January–May 2020) were labeled and evaluated here (see Fig. ). The labels expressed by color in Fig.  indicate which snow type is found at the respective position of the profile. In this study, we focus only on profiles of cold snow that is not experiencing melt, as no standardized interpretation of SMP force profiles exists for wet snow. All profiles collected in the cold season are referred to as “MOSAiC winter data” in the following. Micro-CT and NIR data were recorded whenever possible to validate the subsequent labeling of the SMP profiles. More details on the collection methods can be found in . A comparison of these instruments can be seen in Appendix  in Fig. .

The labeling of the SMP profiles was conducted by two snow experts and is based on the properties of the force signal (magnitude, frequency, and gradient) and the signature of the SMP signal. The labeling procedure is described in detail in Appendix , building upon the notion and observations of . The first labeling phase was conducted by one expert, and in the second phase, two experts revisited the profiles to ensure consistent labeling. The labeling process involves using Micro-CT samples and NIR photography to validate the snow types identified from the force signal where possible. When assigning the labels to the SMP profiles, we lean to the abovementioned international classification guideline of seasonal snow on the ground . However, we regard the labels assigned to the SMP signals as mere approximators. During the labeling process, signal types are grouped together, and we infer from Micro-CTs which snow type matches each group best. Since we seek a language that is common to the snow community, we are using the labels provided by where possible. Since focuses on Alpine snow and does not cover all snow types on Arctic sea ice, such as different forms of depth hoar (further details are given in Appendix B), we extend those labels where necessary. The resulting labeled profiles were used during training, testing, and validation, while some unlabeled profiles were used for semi-supervised models and out-of-distribution tests. Upscaling consistent labeling of SMP profiles is exactly the type of task that ML algorithms can tackle.

We preprocessed each SMP profile and the complete labeled dataset. The surface and the ground of the profiles were detected automatically by the snowmicropyn package (https://snowmicropyn.readthedocs.io/en/latest/, last access: 3 August 2023). For each SMP profile, we replaced negative force values with 0; summarized the signal into bins (1 mm); and added mean, variance, maximum, and minimum force values for those bins. Those values were also determined for a 4 mm and 12 mm moving window. Moreover, the Poisson shot noise model of was used to extract δ, f, L, and the median force value for a 4 and 12 mm window. We added further depth-dependent information, including the distance from the ground and position within the snowpack for each data point. Refer to Table in Appendix for an overview of all features used for each SMP profile and to Fig.  to see the feature importance for each snow type.

We preprocessed the complete labeled dataset by normalizing it, removing profiles from the melting season, and merging snow classes. For example, “decomposed and fragmented precipitation particles” are merged with the class “precipitation particles” since they represent a similar type of snow. The few occurring ice formations and surface hoar instances in the MOSAiC dataset are summarized in the class “rare”. While a high classification performance cannot be expected for the rare classes, we still include them to show how the models perform on a “real-world dataset” that in most cases will also include classes with few occurrences. The data preprocessing ensures that the dataset is clean and that all necessary information, such as depth-dependent information, is available during classification.

The resulting dataset has the following properties. (1) There are multiple noisy and overlapping classes. (2) There is a between-class imbalance; i.e., some snow types occur much more frequently than others. (3) There is a within-class imbalance; i.e., some grain classes contain different sub-grain-classes, but some of them are more frequent than others. (4) The labeling of classes is afflicted with uncertainty; i.e., snow experts themselves are not sure to which class exactly some data points belong. The complexity of the dataset complicates classification and lowers the maximum achievable accuracy.

We compare the capabilities of different models to classify and segment the profiles of the MOSAiC winter SMP dataset. To this end, the models first classify each data point of the signal and then summarize the classified points into distinct snow layers (“first-classify-then-segment”). This task can be solved with different learning and classification techniques.

The task can be addressed via independent classification or sequence labeling. In independent classification, each individual point is classified independently, without looking at other data points. The underlying assumption is that each individual data point carries enough information to be classified solely on that basis. In contrast, sequence labeling assumes that the data are an intra-dependent sequence, where the label of each data point also depends on the preceding labels .

The models can follow either the supervised, unsupervised, or semi-supervised learning regime. In supervised learning, labels are provided to learn an input–output mapping function . In unsupervised learning, patterns and structure are found in unlabeled data ; however, no classification is possible, which is why no unsupervised models are employed here. Instead, semi-supervised models are used, which are able to find structures in sparsely labeled data and leverage this information during classification. In the following, all models employed in this work are shortly presented and put in the context of their learning and task type.

The majority vote classifier is used as the baseline for the performance comparison and simply predicts always the majority class (“rounded grains wind packed”). It satisfies the criteria that a baseline should not require much expertise, be easy to build, and be quick to evaluate .

The cluster-then-predict models employed in this study can be separated into three different semi-supervised and independent classification models. Unsupervised methods are used to find clusters in the dataset, and a supervised model is subsequently used to assign labels to the cluster . As an unsupervised model, k-means clustering , mixture model clustering (GMM) , and Bayesian Gaussian mixture models (BGMM) were used. The supervised part of the model is a simple majority vote within the clusters in order to see if the unsupervised model adds enough information to beat the majority vote baseline.

Label propagation is a graph-based, semi-supervised, and independent classification algorithm. It propagates the labels of labeled data points to unlabeled ones . Here, a modified version of this algorithm by is used (also known as “label spreading”) .

Self-trained classifiers turn a given supervised classifier into a semi-supervised independent classifier. It follows an iterative approach of training a supervised model on labeled data, predicting more data with the model, and retraining the model with the most confident predictions .

Random forests (RFs) are ensembles of diversified decision trees (supervised and independent classification). The diversification happens via tree and feature bagging, where only subsets of data or features are used during training . Decision trees are simple-to-build, explainable, white-box classifiers, and for these reasons they are among the most popular machine learning algorithms . Additionally, a balanced random forest was used with random undersampling to balance the data .

Support vector machines (SVMs) construct a hyperplane in a high-dimensional space to solve binary classification tasks (supervised and independently). When a problem is nonlinearly separable, the input data can be projected into a higher-dimensional space until the problem becomes linearly separable. The kernel trick can be used to circumvent the computationally expensive data transformation involved here. It directly extracts a nonlinear optimal hyperplane .

K-nearest neighbors (KNN) is a local, non-parametric classification method that compares samples and classifies new samples based on their k-nearest training data points (supervised and independently). The class of the prediction sample is determined via a majority vote .

Easy ensemble classifiers are ensembles of balanced adaptive boosting classifiers (supervised and independent). The method is especially helpful for imbalanced datasets since the learners are trained on different bootstrap samples, which are balanced via random undersampling .

Long short-term memories (LSTMs) are a form of artificial neural networks (ANNs) and can perform supervised sequence labeling tasks. ANNs incrementally update their decision function that describes the decision boundary between classes. ANNs have different nodes, which can be seen as representing different parts of the functions that are weighted differently. During training, the weights of the ANN are optimized by minimizing a loss function via gradient descent. A long short-term memory can handle time series data. It consists of different memory cells so the LSTM can forget information that is no longer needed, remember information that is required for future decisions, and retrieve information that is required for current decisions .

Bidirectional LSTMs (BLSTMs) connect two independent LSTMs, where the first LSTM processes the inputs forward, and the second one processes the inputs backwards. The outputs of both LSTMs are connected to one output. This architecture is helpful when the dependencies of a time series go in both time directions, which is the case for snow profiles .

Encoder–decoder networks consist of an ANN encoder that compresses the time-dependent information into a vector and a decoder that uses this information to solve a supervised sequence labeling task. Additionally, the attention mechanism can be used to strengthen the ability to learn long-term dependencies by focusing only on the parts of the input sequence that are relevant for the current time step .

In this work, (1) the performance of different models is compared, (2) differences in the classification of different snow types are analyzed, and (3) the generalization capability of the best-performing model is examined. (1) The performance comparison is done by looking at the metrics of each model and the specific predictions on the test dataset. The metrics used here are accuracy, balanced accuracy, weighted precision, F1 score, area-under-the-receiver operating characteristic (AUROC), log loss, fitting, and scoring time (see Appendix  for further explanations). (2) The label-wise performance is analyzed with the help of label-wise accuracy plots and receiver operating characteristic (ROC) curves. ROC curves plot the true-positive rate versus the false-positive rate. The higher the area under the ROC curve, the clearer the model can separate between positive and negative samples. (3) The generalization capability is tested by running the best-performing model on 100 random profiles from different parts of MOSAiC winter data. These profiles are outside of the distribution of the training, validation, and testing data, and we refer to them as “out-of-distribution profiles”. Here, the out-of-distribution profiles contain the same classes as the training data, so the model still has a chance to predict the correct labels. Evaluating these three aspects ensures that users can choose a model and know (1) how it performs compared to other models, (2) what to expect from the snow-type-specific predictions, and (3) how robust a chosen model will be.

The experimental setup includes a training, validation, and testing framework: roughly 80 % of the labeled dataset is used for training and validation, while the other 20 % is set aside for testing. Validation is realized as 5-fold cross validation . The hyperparameters were tuned on the validation data and the best hyperparameters found were used during testing.

Hyperparameter tuning is the process of searching the optimal internal learning settings of an ML model. Hyperparameters control the learning process of the models, whereas parameters are learned by the model. The tuning is performed on the validation data and the hyperparameters that achieve the highest performance for their model chosen for subsequent model evaluation. Here, tuning was applied moderately and with a simple grid search. All tuning results can be found in the GitHub repository. Specifications of the machine on which the experiments were run can be found in Appendix  and descriptions of the model setup can be found in Appendix .

Overall, the results show that an automatic classification and segmentation of SMP profiles with ML algorithms is possible, even if no further information such as snow pit data or manual segmentation is provided. Category wise, all semi-supervised models were not performing particularly well (see Table ). Only the self-trainer could compete with models from other categories, but this might be the case because the self-trainer is based on a balanced random forest. The supervised models achieved mixed performances. Some models such as the random forests and the SVM are clearly performing well, whereas other models such as the KNN and the easy ensemble are underperforming. Overall, the random forest was the best model in the supervised category since it achieves the highest absolute accuracy (0.73) and F1 score (0.73). However, considering rare classes, the balanced random forest outperformed the plain random forest. All three ANNs did exceptionally well, and their category was clearly the most successful among all three categories. The encoder–decoder showed the best scores among all models in terms of absolute accuracy, precision, and F1 score, closely followed by the LSTM. We consider the LSTM to be the best model within that category since the encoder–decoder only reached its high performance after extensive hyperparameter tuning and underperformed significantly when not tuned well. In contrast, the LSTM achieved its performance more consistently, even under moderate hyperparameter tuning, and it is thus more suitable for users. The subsequent analyses compare the three models that performed best within their category: the LSTM performed best among the ANNs, the random forest among the supervised models, and the self-trainer among the semi-supervised models.

Different ML models exhibited different prediction styles in terms of smoothness and ability to predict rare classes. In Fig.  it becomes visible that the models' predictions are not far off from the labels. In general, the predictions are somewhat similar to the labeled profiles, but the models often had difficulties in determining the precise start and end of a segment. Looking at three random exemplary profiles of the test data in Fig. , one can see that the three main models seem to not only generate similar predictions but also make similar mistakes. In the medium-depth profile (middle column), all three models predicted a longer segment of depth hoar that was not present in the labeled profile. In the shallow profile, all three models predicted some intermediate “depth hoar wind packed” layers in the first third that did not exist. In the deep profile, all three models miss the narrow intermediate depth hoar layer. In summary, it becomes apparent that the different models are producing consistent predictions to a certain degree. There are of course also significant differences among the models. First, the LSTM is closest to the labeled profiles (see Fig. ). Second, the LSTM provided much smoother and less fragmented predictions than the other two models. Third, the self-trainer clearly overestimates rare classes, which hurts the overall performance. To summarize, the LSTM, random forest, and self-trainer show certain prediction similarities among each other; however, the LSTM imitates expert labeling best.

Figure  shows that some snow types are easier and others are harder to classify. The label-wise accuracy seems to be influenced by the following factors: (1) choice of model, (2) the frequency of snow type in the dataset, and (3) the snow type itself. Within one snow type category, the models perform differently; however, some snow types seem to be easier, while others are more difficult to classify for all models. For example, rounded grains wind packed achieved a high accuracy among all models, whereas depth hoar wind packed achieved a low accuracy among all models. This could be partially attributed to the fact that there are fewer samples available for depth hoar wind packed. However, the snow types themselves seem to influence the classification difficulty as well: the precipitation particles class achieves high accuracy values among some models, despite the fact that it is the rarest class in the dataset. For some snow types, some models are able to access certain information, enabling a high performance for that particular snow type that is independent of its frequency. This means that the classification difficulty does not solely depend on the number of available samples. Instead, several other underlying characteristics determine the classification of difficulty of each snow type as well, most notably (1) the initial classification, which is not always consistent; (2) the underlying micro-mechanical properties, i.e., some snow types have characteristic force signals that separate them more clearly from others; and (3) the training dataset, since it does not cover all types of force signals.

Depending on the model, a higher accuracy score could lead to a lower precision score for a label (accuracy–precision trade-off). The ROC curve in Fig.  illustrates this relationship between the true-positive and false-positive rates for the different snow types and their averaged performances. It becomes apparent that both the snow type and the choice of model influence the accuracy–precision trade-off. For example, the rare class seems to be difficult to classify both accurately and precisely for all models, whereas precipitation particles show an almost perfect ROC curve. If one is interested in choosing a model that performs well for a particular snow type, these ROC curves can reveal which model is most suitable. To get even more detailed label- and model-wise insights, refer to the confusion matrices in Appendix . Both the LSTM and the random forest achieve an area under the ROC curve of 0.96. However, on average (see Fig. , dotted pink line), the LSTM outperforms the self-trainer and random forest and is thus most suitable for general classification tasks.

The prediction of the LSTM for 100 random profiles outside of the training and testing distribution is shown in Fig. . Since the labeled profiles are not yet available for these predictions, the generalization capabilities can only be evaluated on the basis of what seems “reasonable”. Melted form depth hoar appears only at the ground of the profiles, precipitation particles appears only at the top, and rounded grains wind packed appears mostly at the top and rather deep – these are all reasonable predictions. However, there are also some predictions that are not reasonable or at least unexpected: the left profile consists almost entirely of depth hoar wind packed, sometimes depth hoar wind packed appears right before melted form depth hoar, and rounded grains wind packed sometimes appears briefly in the “middle” of a profile (and not at the top). Overall, the LSTM seems to make mostly reasonable predictions; however, an in-depth expert analysis of the predictions is necessary to validate that further.

Each model category performs differently because each model takes different aspects of the data into account. Semi-supervised models try to take unlabeled data into account to improve their predictions; however, this did not work well in our context. The most likely reason for the overall underperformance of this category is that the unlabeled data contained out-of-distribution data, i.e., the unlabeled data had different underlying mechanisms than the labeled data (different parts of the winter season). Another reason might be that only a small subset of unlabeled data were included in order to limit running times. Moreover, the poor performance of the cluster-then-predict models is most likely also a result of the classifier used after clustering: a more sophisticated method than a majority vote classifier is needed here.

The simple supervised models take one data point after the other into account and do not consider time series structures within the data. The algorithms used in all previous SMP automation studies fall into this category. In contrast, ANNs are supervised models that take the underlying time sequence of the data into account. While the supervised model in general performed well, they were still clearly outperformed by the ANNs. A likely reason why the ANNs outperformed all the other models is precisely the ANNs' ability to process time-dependent – or in the case of snow profiles depth-dependent – information. ANNs are tackling the classification task as a sequence labeling task, which enables them to include information from the order and position of snow layers. The supervised models still have access to time-relevant information (time window features; see Appendix ); however, they do not have any ability to learn time-based information (what should be remembered and forgotten). Besides, the ANNs learn to imitate the training set, leading to smooth and expert-like predictions. In comparison, taking the time component of SMP signals into account has not been done in previous methods, and we argue that it adds a major information piece and boosts the overall prediction performance significantly.

Each model exhibits a different prediction style due to the models' intrinsic differences and thus might be suitable for specific tasks. The following aspects are listed for consideration (user's guide). A.

Time and resources for hyperparameter tuning. The LSTM and the encoder–decoder network are recommended when plenty of tuning time is available. The encoder–decoder network performs especially badly if not tuned well. The SVM and the balanced random forest need little tuning time, whereas the random forest is the go-to model in cases where (almost) no tuning time can be provided.

Snow types are difficult to classify since their categories are continuous rather than discrete. This was also observed in previous work, and in all previous work performances were reported label-wise to account for those differences . We performed t-distributed stochastic neighbor embedding (t-SNE) on the SMP dataset to visualize how separable the different classes are (see Fig. ). The precipitation particles class, for example, appears as a singled-out green grouping, which is in line with our (and previous) findings that it is easier to classify than other snow types. We conclude that some classes have features distinguishing them more strongly from other snow types. The rounded grain wind packed class behaves in a similar way . However, some classes, such as depth hoar and depth hoar indurated are completely overlapping in Fig. , and indeed our models had problems with differentiating between those two classes. Similarly, depth hoar wind packed seems to overlap largely with rounded grains wind packed and melted form of depth hoar. We theorize that the reason for their non-separability is that those snow types transform into each other during snow metamorphosis. This means many data points can not be discretized into one single category since they are on a continuous spectrum. also pointed out that they often found data points that were in transition between snow classes and attributed this to the fact that the snow is changing continuously. In conclusion, it is currently impossible to reach 100 % classification accuracy on every snow type since some snow types will always lie between categories.

Despite these difficulties, the underlying SMP signals are still characteristic enough for specific snow types to be classified successfully. The different micro-mechanical properties of the snow types are reflected in the SMP signal and are thus the driver for the classification. Some classes, such as precipitation particles, can be clearly separated from others since the bonding between the grains is so weak that the force signal is very low. As long as precipitation particles do not share this characteristic with other snow types, they can be easily classified. Refer to Appendix  to learn more about the relation between snow types and SMP signal, and refer to Appendix  to see which classes have unique signal characteristics and which classes have shared signal characteristics.

The classification difficulties also extend to the expert labeling process itself. The continuous nature of the snow types makes it particularly difficult for domain experts to agree on labeling, i.e., two different snow experts will produce two different labeled and segmented profiles for the same SMP measurement . This is another reason why a classification accuracy of 100 % cannot be reached. One might suggest supplementing the classification process with additional observational data to make the process more “objective” (as we also do here). However, each classification and segmentation of a snowpack is “subjective” in nature right now, no matter which observational data are used as the basis for the classification. When requesting a segmentation and classification of a snowpack, one is always requesting the classification of a specific expert. While the operator bias can be mitigated by using NIR, Micro-CTs, or the SMP, the classification of those measurements remains subjective. It is neither this study's goal nor its task to provide an objective classification; instead, we aim for a consistent classification.

However, difficulties in reaching 100 % accuracy do not preclude overall good performance. While experts may end up with different segmentations and classifications, they can still agree that two different analyses are both valid analyses of the same profile. Similarly, the algorithms provided here output predictions that may not always align with the expert labeling but are sensible and directly usable. Hence, we cannot evaluate the models solely based on numerical metrics, such as accuracy, but must also evaluate the performance from a qualitative perspective. This is the reason why we evaluated if an SMP user, who also labeled the training data, would (1) accept the predictions of the ML algorithms on an out-of-distribution dataset, (2) find them consistent with their own labeling, (3) and subsequently work with those predictions. In the case of the MOSAiC dataset, all those aspects were fulfilled. We find such a qualitative assessment important since these questions decide whether or not the tools provided will be used in practice.

We also want to point out that the algorithms themselves are entirely agnostic to the question of “subjectivity”. The algorithms are merely reproducing what they have been trained on. If we can provide the algorithms with a dataset that can be considered “fully objective”, and the community agrees on that as ground truth data, the algorithms could reproduce those hypothetical objective labels. Alternatively, signals could also be grouped first, and some abstract classes could be assigned to them. Nevertheless, even this would rely on human expertise since the parameters to separate those groups would be subject to discussion (see Fig. ; the groups are not simply separable from each other, and the clustering would depend on parameter choices). In general, we provide a methodological framework here to classify and segment SMP profiles. Which classification patterns are reproduced depends on the user's choice.

The benefits of using an automatic classification are that the SMP user can (1) save valuable time, (2) receive consistent labeling, and (3) perform statistical analysis on their SMP dataset. In the case of the MOSAiC dataset, manual labeling would have meant labeling over 3000 profiles, which can easily take up to a year to classify (next to other obligations of domain experts). In terms of consistency, we already experienced how some of the models' predictions helped us – to our surprise – to detect human mistakes and inconsistencies during the first labeling round. Furthermore, such an upscaled classification enables, for the first time, the statistical analysis of an SMP dataset. One of the initial research questions for MOSAiC was “Is depth hoar in Arctic snowpacks mostly present at the bottom and rounded grains wind packed at the top?”. With the help of snowdragon, the MOSAiC dataset could be consistent and accurately labeled enough to answer such a question with “Yes, this is indeed the case.”.

The LSTM can generalize to other winter profiles with the same snow types since the underlying classification and segmentation rules stay the same. However, the LSTM's generalization capability does not extend to other seasons or regions when and where other snow types are found, such as melted forms or regional snow types. As mentioned before, the models do not generalize regarding the different classification styles of experts. The models used in this work are still generalizable in that they can be used on any desired dataset as long as they are retrained on the chosen dataset. This would not have been possible in previous works, such as , since knowledge rules for one snow region and season do not transfer to other regions or seasons. For greater generalization capability, the LSTM – or any other model – must be either trained with a more general dataset or specifically retrained for an individual dataset.

As previously discussed, the uncertainty in expert labeling is a general limitation of this particular study. While this uncertainty might be partially mitigated further by using a dataset for which many additional in situ observations exist, it would still remain an issue. One approach for future work would be to quantify the uncertainty that is inflicted upon the labeled profiles. Subsequently, a machine learning model could be trained to not only classify snow types but also provide a probabilistic classification.

This work does not address the task setting of a first-segment-then-classify algorithm because this would require a completely different set of methods. In a first-segment-then-classify setting, the SMP signal could first be segmented with techniques used in audio segmentation . The resulting time series pieces could subsequently be classified as a whole . Future work could experiment with this problem formulation and analyze if performance further increases in this setting.

The ANNs used here are off-the-shelf networks and are not adapted to the specific underlying task in order to ensure a fair comparison between the different models. However, one could look into adapting the loss functions to include similarity measurements between snow samples. Results from clustering, performed on t-SNE data, could then be leveraged during classification to increase classification performance. Adapting the loss function of the ANNs could increase prediction performance greatly; however, such a loss function must be carefully constructed and evaluated on different datasets.

As mentioned in Sect. , the models cannot generalize to completely different settings in terms of seasons and regions. To ensure generalization capability one could train a large model on a dataset that includes snow types from different regions and seasons. Such a dataset would need to be newly compiled because common SMP datasets are usually limited to one region . In theory, a large enough model trained on a large enough dataset could be able to produce direct predictions for any SMP users. Thus, it would be interesting to train an ML model on a generalized dataset and validate its performance on the specialized MOSAiC SMP dataset. This would shed new light on the spatiotemporal transferability of the ML models presented here.

Alternatively, SMP users can simply retrain a chosen model for their particular dataset. They need to provide a set of SMP profiles for their region, season, and classification style, but the overall time savings are still immense. To summarize, the generalization capabilities may be enhanced by using a more general dataset, or one can bypass this problem by retraining to specific datasets. The snowdragon repository addresses the needs of the latter.

An immediate consequence of this study is the further analysis of the unlabeled part of the MOSAiC dataset. Domain experts can use the LSTM or other models to create predictions for the remaining 3516 profiles. A previously almost impossible task to classify and segment those thousands of profiles became feasible by providing just a set of 164 labeled profiles. The results of these predictions and their impacts on the cryospheric analysis of snow coverage in the Arctic will become apparent in future publications.