The accurate and efficient monitoring of snow cover is a cornerstone for understanding hydrological cycles, climate dynamics, and water resource management in cold regions. Traditional field measurements, while precise, are spatially limited and labor-intensive. Satellite remote sensing provides extensive spatial coverage but is often constrained by coarse spatial resolution, cloud contamination, and fixed revisit cycles. In this context, UAV drone-based remote sensing has emerged as a transformative technology, offering unprecedented flexibility, high spatial resolution, and the ability to operate under cloud cover. The proliferation of consumer-grade UAV drone platforms equipped with RGB (Red, Green, Blue) cameras has democratized high-resolution snow mapping. However, a significant challenge persists: in complex terrains, RGB imagery alone often struggles to reliably discriminate snow from other surfaces with similar high reflectance, such as certain rocks, ice, or man-made structures like white roofs.
This spectral ambiguity can lead to substantial errors in automated snow cover classification. To overcome this limitation, this study proposes an integrated framework that leverages the complementary strengths of multi-sensor UAV drone payloads and advanced machine learning algorithms. Specifically, we fuse high-resolution visible (RGB) and thermal infrared (TIR) imagery acquired by a UAV drone. The TIR data provides crucial information on surface radiative temperature, a property where snow, due to its high albedo and latent heat effects, typically exhibits distinct signatures compared to bare soil, vegetation, or built-up areas, even when their visual color is similar.
Beyond sensor fusion, the choice of classification algorithm is paramount. While single, powerful classifiers (e.g., Support Vector Machines – SVM) are commonly used, they may not fully capture the complex, non-linear relationships in high-dimensional feature space derived from UAV drone data. Ensemble learning methods, which combine the predictions of multiple base models, have demonstrated superior performance in various remote sensing applications by reducing variance and mitigating overfitting. Therefore, this research systematically evaluates and compares the efficacy of several state-of-the-art ensemble classifiers against traditional single classifiers for snow cover extraction.
The core objectives of this investigation are threefold: (1) To construct a rich feature dataset from co-registered UAV drone-derived RGB imagery, TIR imagery, and Digital Surface Models (DSM); (2) To rigorously assess the performance of ten different classifiers, including seven ensemble methods (e.g., Random Forest, Gradient Boosting machines) and three single classifiers, for mapping snow cover; (3) To analyze the impact of comprehensive versus optimally selected feature sets on classification accuracy and computational efficiency. This holistic approach aims to develop a robust, accurate, and practical methodology for fine-scale snow cover monitoring using versatile UAV drone platforms.
1. Materials and Methodology
1.1 Study Area and UAV Drone Data Acquisition
The research was conducted over an agricultural region characterized by flat terrain, comprising croplands, scattered buildings, and plastic film coverings. This area experiences a dry continental climate with seasonal snowfall. A mid-winter flight campaign was executed following a fresh snow event, ensuring the presence of a heterogeneous snow cover ranging from continuous patches to fragmented, melting snow interspersed with bare soil.
Data acquisition was performed using a DJI Matrice 210 RTK quadcopter UAV drone, a platform renowned for its stability and positioning accuracy. The UAV drone was equipped with a Zenmuse XT2 dual-sensor gimbal, which simultaneously captures 12-megapixel RGB images and radiometric thermal infrared images (wavelength range: 7.5–13.5 µm). The flight was planned at an altitude of 80 meters with high forward and side overlaps (85%) to ensure robust photogrammetric processing. Ground Control Points (GCPs) were surveyed using a differential GNSS receiver to georeference the outputs with high precision.
1.2 Data Processing and Feature Dataset Construction
The collected RGB and TIR images were processed separately using structure-from-motion (SfM) photogrammetry software (Pix4Dmapper). This process generated three core products: a 5-cm resolution RGB orthomosaic, a 9-cm resolution TIR orthomosaic, and a 5-cm resolution Digital Surface Model (DSM). All rasters were co-registered and resampled to a consistent 9-cm resolution for subsequent analysis.
A ground truth dataset was established through a combination of field surveys and precise visual interpretation of the orthomosaics. Samples were collected for five land cover classes: Snow, Bare Soil, Building, Shadow, and Plastic-Covered Land. This dataset was randomly split into 70% for training the classifiers and 30% for independent validation.
The strength of the proposed methodology lies in the construction of a comprehensive feature dataset that goes beyond simple spectral bands. We derived a multi-dimensional feature vector for each pixel, encompassing spectral, thermal, textural, and topographic information as summarized in Table 1.
| Feature Category | Feature Names | Description / Formula | Count |
|---|---|---|---|
| Spectral & Thermal | R, G, B, T | Original Red, Green, Blue, and Thermal Infrared band values. | 4 |
| Spectral Indices & Ratios | r, g, b | Normalized RGB values: e.g., $$r = \frac{R}{R+G+B}$$ | 3 |
| RBDI, RGDI | Red-Blue Difference Index: $$RBDI = R – B$$; Red-Green Difference Index. | 2 | |
| RGRI, RGI, RBI | Ratios and normalized difference indices (e.g., Red-Green Ratio Index). | 3 | |
| Color Space Transformations | H,S,V (HSV); H,L,S (HLS); H,S,I (HSI) | Features derived from transforming RGB to HSV, HLS, and HSI color spaces. Hue (H) is particularly useful for separating snow from white buildings. | 9 |
| Topographic Feature | DSM | Elevation from the Digital Surface Model. | 1 |
| Texture Features (GLCM) | Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, Correlation | 8 texture metrics calculated from the first principal component of the RGB bands using a Gray-Level Co-occurrence Matrix (GLCM). For example, Contrast measures local intensity variation: $$\text{Contrast} = \sum_{i,j} |i-j|^2 P(i,j)$$ where \(P(i,j)\) is the probability of co-occurrence of gray levels \(i\) and \(j$. |
8 |
| Total Features | 30 | ||
1.3 The Snow Cover Identification Framework
1.3.1 Classifiers: Single vs. Ensemble
We evaluated a suite of ten classifiers, implemented using the Scikit-learn library, to identify the most effective algorithm for snow cover mapping from UAV drone data.
Single Strong Classifiers:
- Classification and Regression Tree (CART): A simple, interpretable decision tree.
- Support Vector Machine (SVM): Finds the optimal hyperplane to separate classes.
- Multi-layer Perceptron Classifier (MLPC): A basic artificial neural network.
Ensemble Learning Classifiers: These methods build multiple models and aggregate their predictions.
- Bagging Family:
- Random Forest (RF): Builds an ensemble of decorrelated decision trees using bootstrap sampling and random feature subsets.
- Extra Trees (ET): Similar to RF but uses random thresholds for splits, increasing randomness.
- Boosting Family: Sequentially builds models, where each new model corrects errors of the previous one.
- AdaBoost with Decision Trees (AdaBoost-DT): Adaptively increases weights of misclassified samples.
- Gradient Boosting Decision Tree (GBDT): Optimizes a loss function using gradient descent.
- Extreme Gradient Boosting (XGBoost): An optimized, regularized implementation of gradient boosting.
- Light Gradient Boosting Machine (LightGBM): Uses a histogram-based algorithm for faster training and lower memory usage.
- Histogram-Based Gradient Boosting (HistGBoost): Also uses histogram binning for efficiency.
Hyperparameters for each classifier were optimized using a randomized search with 10-fold cross-validation to ensure robust performance.
1.3.2 Feature Importance and Optimal Feature Selection
Using all 30 features (“Full Feature Set”) ensures no information loss but may include redundancy, increasing computational cost and the risk of overfitting. Therefore, we implemented a feature selection strategy to create an “Optimal Feature Set” for each classifier. For tree-based ensemble methods (RF, GBDT, XGBoost, etc.), we utilized their built-in feature importance scores (e.g., Gini impurity reduction). For SVM and MLPC, which lack inherent importance metrics, univariate statistical tests were employed. This process automatically selected the most informative subset of features for each specific classifier model.
1.3.3 Accuracy Assessment and Fractional Snow Cover Derivation
Classifier performance was assessed using the independent validation sample set. Metrics included Overall Accuracy (OA), Kappa coefficient, Producer’s Accuracy (PA), and User’s Accuracy (UA).
Following the generation of a binary snow cover map from the best-performing model, Fractional Snow Cover (FSC) was calculated. This was achieved by applying a moving window analysis at various spatial resolutions (e.g., 1m, 10m, 50m) to compute the percentage of snow-covered pixels within each window, simulating products comparable to medium-resolution satellite sensors.
2. Results and Analysis
2.1 Classifier Performance: The Superiority of Ensemble Learning
The quantitative accuracy assessment revealed a clear and consistent advantage of ensemble learning classifiers over single classifiers for snow cover identification from UAV drone imagery. The results are summarized in Table 2.
| Classifier Type | Classifier Name | OA (%) | Kappa | Key Observation |
|---|---|---|---|---|
| Ensemble Learners | AdaBoost-DT | 92.91 | 0.90 | Top performer, excellent balance across all classes. |
| Extra Trees (ET) | 92.23 | 0.89 | Very high accuracy, comparable to top performers. | |
| LightGBM (LGBM) | 92.23 | 0.89 | Excellent accuracy with high computational efficiency. | |
| Gradient Boosting (GBDT) | 92.23 | 0.89 | Robust performance. | |
| Random Forest (RF) | 91.22 | 0.88 | Strong and reliable performance. | |
| XGBoost | 91.89 | 0.89 | Competitive accuracy. | |
| HistGBoost | 91.22 | 0.88 | Good performance. | |
| Single Classifiers | Support Vector Machine (SVM) | 89.86 | 0.86 | Best among single classifiers. |
| Multi-layer Perceptron (MLPC) | 89.53 | 0.86 | Moderate performance. | |
| Classification Tree (CART) | 87.16 | 0.82 | Lowest accuracy, prone to overfitting. |
Ensemble methods achieved OA values between 91.22% and 92.91%, significantly outperforming single classifiers (OA: 87.16% to 89.86%). The boosting algorithm AdaBoost-DT delivered the highest accuracy (OA=92.91%, Kappa=0.90). Visually, ensemble classifiers produced cleaner maps with less spurious noise and better delineation of snow patches, especially in fragmented areas and at class boundaries (e.g., between snow and building shadows).
2.2 Impact of Feature Dataset: Full vs. Optimal Set
The analysis of using the Full Feature Set versus the classifier-specific Optimal Feature Set yielded critical insights into the trade-off between accuracy and efficiency.
Accuracy: Without exception, every classifier achieved a higher OA when trained on the Full Feature Set (30 features) compared to its Optimal Feature Set. The accuracy drop for the Optimal Set ranged from 0.68 to 3.04 percentage points in OA. This demonstrates that while feature selection removes redundancy, the complete feature vector preserves subtle but valuable information from the UAV drone sensors that contributes to discerning challenging class separations, such as snow from bright buildings or shaded soil.
Efficiency: The primary benefit of the Optimal Feature Set was a marked reduction in model training and prediction time. The computational cost decreased proportionally with the number of features removed. For instance, tree-based models like RF and AdaBoost-DT saw faster execution times with minimal accuracy penalty, making them highly efficient. However, for some boosting algorithms like LightGBM and GBDT, the feature selection process sometimes removed features critical for modeling complex interactions, leading to a more noticeable accuracy decline.
2.3 Decoding Feature Importance
Analyzing the frequency with which features were selected across all classifiers provides a consensus view of their relative importance for snow identification with a UAV drone. The most frequently selected features were:
- Mean Texture (GLCM Mean): Selected by all 10 classifiers. This highlights the fundamental importance of image texture, as snow typically presents a homogeneous surface compared to rough bare soil or structured buildings.
- Thermal Infrared Band (T): Selected 9 times. This underscores the critical value of fusing thermal data with RGB. Snow’s low surface temperature is a key discriminant that RGB cameras cannot capture.
- Blue Band (B) and Red-Blue Difference Index (RBDI): Selected 8 and 7 times, respectively. Snow has high reflectance in the blue region, and indices like RBDI effectively amplify the spectral contrast between snow and common non-snow surfaces.
- Digital Surface Model (DSM) and Hue from HSV: Also frequently selected. Elevation can correlate with snow persistence, while Hue is effective in separating white snow from white/gray man-made materials.
Features like Entropy (texture) and certain normalized band ratios were rarely selected, indicating lower discriminative power for this specific task.
2.4 From Classification to Fractional Snow Cover Mapping
Using the binary snow cover map generated by the best model (AdaBoost-DT with Full Features), we demonstrated the derivation of Fractional Snow Cover (FSC). The FSC maps generated at different aggregation scales (1m, 10m, 50m) clearly showed how spatial resolution impacts the representation of snow cover heterogeneity. High-resolution (1m) FSC maps preserved fine details of melting snow patterns, while coarser resolutions (50m) provided a spatially averaged overview, similar to what a satellite sensor like MODIS would see but derived from the foundational accuracy of the UAV drone classification. This bridges the gap between very high-resolution mapping and operational snow monitoring products.
3. Discussion and Conclusion
This study systematically investigated a robust methodology for snow cover monitoring by integrating multi-sensor UAV drone imagery with ensemble learning. The key findings and their implications are synthesized below.
The Ensemble Advantage is Decisive. The empirical results leave little doubt that ensemble learning classifiers are superior for processing the complex, high-dimensional feature space derived from modern UAV drone platforms. Algorithms like AdaBoost-DT, Random Forest, and LightGBM consistently outperformed strong single classifiers like SVM by 2-6% in Overall Accuracy. This performance gain translates into more reliable maps with fewer errors in ecologically or hydrologically critical transition zones. The robustness of ensembles comes from their ability to reduce model variance and generalize better from the training data, which is essential when working with the high-resolution but potentially noisy data from a UAV drone.
Sensor Fusion is not Optional; it’s Essential. The high importance ranking of the Thermal Infrared (T) band is a pivotal finding. It conclusively validates the hypothesis that RGB imagery alone is insufficient for accurate snow mapping in complex scenes. The thermal data provides a physically-based separation criterion that is largely independent of illumination and color. Therefore, investing in a dual RGB-TIR payload for a UAV drone, or planning concurrent flights with different sensors, is highly recommended for serious snow monitoring applications.
The Full Feature Trade-off: Accuracy vs. Efficiency. While feature selection is a standard step to improve efficiency, our results caution against its automatic use when maximum accuracy is the priority. The Full Feature Set, encompassing spectral, thermal, textural, and topographic features, delivered the best possible classification results. The marginal accuracy loss from using an optimal set might be acceptable for rapid, large-area processing with an efficient classifier like Random Forest. However, for critical applications requiring the highest fidelity, utilizing the full suite of UAV drone-derived features is justified.
Operational Recommendations. Based on our comprehensive analysis, we propose the following workflow for optimal snow cover mapping using a UAV drone:
- Platform & Sensor: Use a UAV drone capable of carrying both an RGB camera and a radiometric thermal infrared camera.
- Data Processing: Generate co-registered orthomosaics for RGB, TIR, and DSM.
- Feature Extraction: Compute a comprehensive feature set including original bands, key indices (e.g., RBDI), texture (GLCM Mean), and transformed color spaces (HSV Hue).
- Classifier Selection: Employ an ensemble learning classifier. AdaBoost-DT or Random Forest are excellent, robust choices. For very large datasets, LightGBM offers a speed advantage.
- Model Training: Train the selected classifier using the Full Feature Set for maximum accuracy, or with an optimally selected subset if processing speed is a greater concern than the last 1-2% of accuracy.
In conclusion, this research demonstrates that the synergistic use of visible-thermal UAV drone data and ensemble machine learning algorithms forms a powerful paradigm for high-accuracy, fine-scale snow cover monitoring. This approach effectively solves the long-standing challenge of spectral confusion in optical snow mapping. The methodology is adaptable and holds great promise for integration into operational monitoring schemes, contributing to improved hydrological modeling, climate research, and water resource management in snow-dominated regions.