This paper proposes an innovative method for classifying the physical properties of the seasonal snowpack using near-infrared (NIR) hyperspectral imagery to discriminate the optical classes of snow at different degrees of metamorphosis. This imaging system leads to fast and non-invasive assessment of snow properties. Indeed, the spectral similarity of two samples indicates the similarity of their chemical composition and physical characteristics. This can be used to distinguish, without a priori recognition, between different classes of snow solely based on spectral information. A multivariate data analysis approach was used to validate this hypothesis. A principal component analysis (PCA) was first applied to the NIR spectral data to analyze field data distribution and to select the spectral range to be exploited in the classification. Next, an unsupervised classification was performed on the NIR spectral data to select the number of classes. Finally, a confusion matrix was calculated to evaluate the accuracy of the classification. The results allowed us to distinguish three snow classes of typical shape and size (weakly, moderately, and strongly metamorphosed snow). The evaluation of the proposed approach showed that it is possible to classify snow with a success rate of 85% and a kappa index of 0.75. This illustrates the potential of NIR hyperspectral imagery to distinguish between three snow classes with satisfactory success rates. This work will open new perspectives for the modelling of physical parameters of snow using spectral data.

Improving the estimation of snow density is a key task in current snow research. Characterization of the variability of density in time and space is essential for the estimation of water equivalent, hydroelectric power production, assessment of natural hazards (avalanches, floods, etc.). Hyperspectral imaging is proving to be a promising and reliable tool for monitoring and estimating this physical property. Indeed, the spectral reflectance of snow is partly controlled by changes in its physical properties, particularly in the near-infrared (NIR) part of the spectrum. For this purpose, several models have been designed to estimate snow density from spectral information. However, none has yet achieved significant performance. One of the major difficulties is that the relationship between snow density and spectral reflectance is non-bijective (surjective). Indeed, several reflectance amplitudes can be associated with the same density and vice versa, so the correlation between density and spectral reflectance can be very poor. To resolve this issue, a hybrid snow density estimation model based on spectral data is proposed in this work. The principle behind this model is to classify the snow density prior to its estimation by means of a specific estimator corresponding to a predetermined snow density class. These additional steps eliminate the surjective relation by converting it into three bijective relations between density and spectral reflectance. The calibration step showed that the densities included within the three classes are sensitive to different spectral regions, with R2 > 0.80. The results of the cross-validation for the specific estimators were also satisfactory with R2 > 0.78 and RMSE < 36.36 kg m−3. The overall performance of the hybrid model (HM), when tested with independent data, demonstrated the effectiveness of using proximal NIR hyperspectral imagery to estimate snow density (R2 = NASH = 0.93).

Estimating the seasonal density of the snowpack has many financial and environmental benefits. Rapid assessment and daily monitoring of its evolution are therefore key to effective prevention. Traditionally, the physical characteristics of snow are measured directly in the field, which involves high costs and personnel mobilization. Hyperspectral imaging is a reliable and efficient technique to study and evaluate this physical property. The spectral reflectance of snow is partly defined by changes in its physical properties, particularly in the Near infrared (NIR) part of the spectrum. Recently, a hybrid snow density estimation model allowing retrieval of density from NIR hyperspectral data was developed, based on an a priori classification of snow samples. However, in order to obtain optimal density estimates with the Hybrid model (HM), the sources of classification and estimation error must be controlled. Following the same principle as the HM, an Ensemble-based system (EBS) was developed. This model reduces the number of misclassification errors produced by the HM. The general concept of EBS algorithms is based on the principle that obtaining more opinions before making a decision is part of human nature, especially when economic and environmental benefits are at stake. This approach has helped to reduce the risk of classification and estimation errors and to develop more robust density results. One hundred and fourteen snow samples collected during three winters (2018–2020) were used to calibrate and validate the EBS. The performance of the EBS was validated using an independent database and the results were satisfactory (R2 = 0.90, RMSE = 44.45 kg m−3, BIAS = 3.87 kg m−3 and NASH = 0.89).