An Automated Snow Mapper Powered by Machine Learning
Abstract
:1. Introduction
2. Materials and Methods
2.1. Study Area
2.2. Data
2.3. AutoSMILE: Automated Snow Mapper Powered by Machine Learning
2.3.1. Image Segmentation
2.3.2. Object Labelling and Feature Extraction
2.3.3. Machine Learning and Deep Learning
2.3.4. Performance Evaluation and Post-Processing
3. Results
3.1. Snow Mapping Results from Different Segmentations
3.2. Snow Mapping Results of Different Datasets
3.3. Layer Importance Analysis Based on the Best Model
4. Discussion
4.1. The Generalizability of the Trained AutoSMILE Model
4.2. Comparison of AutoSMILE and Threshold-Based Methods
5. Conclusions
- Using only 5% of the main study area for training (50.4 km2), AutoSMILE achieves extraordinarily satisfactory results of snow mapping in the rest of the main study area: object and pixel PAs, UAs, IoUs and OA values reaching 99.42% and 98.78%, 98.21% and 98.76%, 98.84% and 98.35%, and 97.23% and 98.07%, respectively. When applying the trained models to the testing zone in Qimantag Mountain region, the highest OA reaches 97.22%, indicating the excellent performance, generalizability and robustness of AutoSMILE.
- According to the parametric study of segmentation parameters, KZ and MD should be carefully determined. In AutoSMILE, a low MD is not recommended to avoid excessive oversegmentation and low computational efficiency, but a small value of KZ is recommended to better handle the transition zone between snow and non-snow areas.
- Results of alternating dataset combinations indicate that auxiliary data like multispectral image derived indices and DEM derivatives play a limited role in enhancing the performance of AutoSMILE. High-quality snow mapping can be accomplished with only the multispectral image using AutoSMILE. Based on permutation importance analysis of the best ML model, the top five important layers are red, blue, green, red edge 1 and short wave infrared 1 band layers.
- AutoSMILE outperforms the existing threshold-based methods in both regions. The optimal NDSI thresholds of the two regions vary, suggesting that the threshold method typically requires site knowledge to achieve the best performance and it is hard to find a universally applicable threshold. When investigating the performance at the sub-regions, it was found that both AutoSMILE and threshold-based methods perform well when the land cover is simple. However, when encountering complex conditions like snow mapping in transition zones, AutoSMILE outperforms the threshold-based method substantially (up to 92% of IoU increase and up to 13% of OA increase).
- Due to the inevitable segmentation loss induced by object-based analysis, transition zones between snow and non-snow areas require more attention when inspecting the final snow cover products.
Author Contributions
Funding
Data Availability Statement
Acknowledgments
Conflicts of Interest
References
- Arino, O.; Bicheron, P.; Achard, F.; Latham, J.; Witt, R.; Weber, J.L. The Most Detailed Portrait of Earth. Eur. Space Agency 2008, 136, 25–31. [Google Scholar]
- Dedieu, J.-P.; Carlson, B.Z.; Bigot, S.; Sirguey, P.; Vionnet, V.; Choler, P. On the Importance of High-Resolution Time Series of Optical Imagery for Quantifying the Effects of Snow Cover Duration on Alpine Plant Habitat. Remote Sens. 2016, 8, 481. [Google Scholar] [CrossRef] [Green Version]
- Biemans, H.; Siderius, C.; Lutz, A.F.; Nepal, S.; Ahmad, B.; Hassan, T.; von Bloh, W.; Wijngaard, R.R.; Wester, P.; Shrestha, A.B.; et al. Importance of Snow and Glacier Meltwater for Agriculture on the Indo-Gangetic Plain. Nat. Sustain. 2019, 2, 594–601. [Google Scholar] [CrossRef]
- Croce, P.; Formichi, P.; Landi, F.; Mercogliano, P.; Bucchignani, E.; Dosio, A.; Dimova, S. The Snow Load in Europe and the Climate Change. Clim. Risk Manag. 2018, 20, 138–154. [Google Scholar] [CrossRef]
- Zhao, J.; Shi, Y.; Huang, Y.; Fu, J. Uncertainties of Snow Cover Extraction Caused by the Nature of Topography and Underlying Surface. J. Arid Land 2015, 7, 285–295. [Google Scholar] [CrossRef]
- Zhang, H.; Zhang, F.; Zhang, G.; Che, T.; Yan, W.; Ye, M.; Ma, N. Ground-Based Evaluation of MODIS Snow Cover Product V6 across China: Implications for the Selection of NDSI Threshold. Sci. Total Environ. 2019, 651, 2712–2726. [Google Scholar] [CrossRef]
- Härer, S.; Bernhardt, M.; Siebers, M.; Schulz, K. On the Need for a Time- and Location-Dependent Estimation of the NDSI Threshold Value for Reducing Existing Uncertainties in Snow Cover Maps at Different Scales. Cryosphere 2018, 12, 1629–1642. [Google Scholar] [CrossRef]
- Hall, D.K.; Riggs, G.A.; Salomonson, V.V.; Barton, J.; Casey, K.; Chien, J.; DiGirolamo, N.; Klein, A.; Powell, H.; Tait, A. Algorithm Theoretical Basis Document (ATBD) for the MODIS Snow and Sea Ice-Mapping Algorithms; NASA Goddard Space Flight Center: Greenbelt, MD, USA, 2001. [Google Scholar]
- Gascoin, S.; Grizonnet, M.; Bouchet, M.; Salgues, G.; Hagolle, O. Theia Snow Collection: High-Resolution Operational Snow Cover Maps from Sentinel-2 and Landsat-8 Data. Earth Syst. Sci. Data 2019, 11, 493–514. [Google Scholar] [CrossRef] [Green Version]
- Hall, D.K.; Salomonson, V.V. MODIS Snow Products User Guide to Collection 5; NASA Goddard Space Flight Center: Greenbelt, MD, USA, 2006. [Google Scholar]
- Hall, D.K.; Riggs, G.A. Accuracy Assessment of the MODIS Snow Products. Hydrol. Process. 2007, 21, 1534–1547. [Google Scholar] [CrossRef]
- Stumpf, A.; Kerle, N. Object-Oriented Mapping of Landslides Using Random Forests. Remote Sens. Environ. 2011, 115, 2564–2577. [Google Scholar] [CrossRef]
- Ghorbanzadeh, O.; Blaschke, T.; Gholamnia, K.; Meena, S.R.; Tiede, D.; Aryal, J. Evaluation of Different Machine Learning Methods and Deep-Learning Convolutional Neural Networks for Landslide Detection. Remote Sens. 2019, 11, 196. [Google Scholar] [CrossRef] [Green Version]
- Keyport, R.N.; Oommen, T.; Martha, T.R.; Sajinkumar, K.S.; Gierke, J.S. A Comparative Analysis of Pixel- and Object-Based Detection of Landslides from Very High-Resolution Images. Int. J. Appl. Earth Obs. Geoinf. 2018, 64, 1–11. [Google Scholar] [CrossRef]
- Rastner, P.; Bolch, T.; Notarnicola, C.; Paul, F. A Comparison of Pixel- and Object-Based Glacier Classification with Optical Satellite Images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2014, 7, 853–862. [Google Scholar] [CrossRef]
- Wang, X.; Gao, X.; Zhang, X.; Wang, W.; Yang, F. An Automated Method for Surface Ice/Snow Mapping Based on Objects and Pixels from Landsat Imagery in a Mountainous Region. Remote Sens. 2020, 12, 485. [Google Scholar] [CrossRef] [Green Version]
- Wang, H.J.; Zhang, L.M.; Xiao, T. DTM and Rainfall-Based Landslide Susceptibility Analysis Using Machine Learning: A Case Study of Lantau Island, Hong Kong. In Proceedings of the The Seventh Asian-Pacific Symposium on Structural Reliability and Its Applications (APSSRA 2020), Tokyo, Japan, 4–7 October 2020. [Google Scholar]
- Wang, H.J.; Xiao, T.; Li, X.Y.; Zhang, L.L.; Zhang, L.M. A Novel Physically-Based Model for Updating Landslide Susceptibility. Eng. Geol. 2019, 251, 71–80. [Google Scholar] [CrossRef]
- Rastegarmanesh, A.; Moosavi, M.; Kalhor, A. A Data-Driven Fuzzy Model for Prediction of Rockburst. Georisk Assess. Manag. Risk Eng. Syst. Geohazards 2021, 15, 152–164. [Google Scholar] [CrossRef]
- Hosseini, S.A.A.; Mojtahedi, S.F.F.; Sadeghi, H. Optimisation of Deep Mixing Technique by Artificial Neural Network Based on Laboratory and Field Experiments. Georisk Assess. Manag. Risk Eng. Syst. Geohazards 2020, 14, 142–157. [Google Scholar] [CrossRef]
- Wang, L.; Chen, Y.; Tang, L.; Fan, R.; Yao, Y. Object-Based Convolutional Neural Networks for Cloud and Snow Detection in High-Resolution Multispectral Imagers. Water 2018, 10, 1666. [Google Scholar] [CrossRef] [Green Version]
- Liu, C.; Huang, X.; Li, X.; Liang, T. MODIS Fractional Snow Cover Mapping Using Machine Learning Technology in a Mountainous Area. Remote Sens. 2020, 12, 962. [Google Scholar] [CrossRef] [Green Version]
- Rahmati, O.; Ghorbanzadeh, O.; Teimurian, T.; Mohammadi, F.; Tiefenbacher, J.P.; Falah, F.; Pirasteh, S.; Ngo, P.T.T.; Bui, D.T. Spatial Modeling of Snow Avalanche Using Machine Learning Models and Geo-Environmental Factors: Comparison of Effectiveness in Two Mountain Regions. Remote Sens. 2019, 11, 2995. [Google Scholar] [CrossRef] [Green Version]
- Tsai, Y.-L.S.; Dietz, A.; Oppelt, N.; Kuenzer, C. Wet and Dry Snow Detection Using Sentinel-1 SAR Data for Mountainous Areas with a Machine Learning Technique. Remote Sens. 2019, 11, 895. [Google Scholar] [CrossRef] [Green Version]
- Cannistra, A.F.; Shean, D.E.; Cristea, N.C. High-Resolution Cubesat Imagery and Machine Learning for Detailed Snow-Covered Area. Remote Sens. Environ. 2021, 258, 112399. [Google Scholar] [CrossRef]
- Liu, Y.; Chen, X.; Hao, J.-S.; Li, L.-H. Snow Cover Estimation from MODIS and Sentinel-1 SAR Data Using Machine Learning Algorithms in the Western Part of the Tianshan Mountains. J. Mt. Sci. 2020, 17, 884–897. [Google Scholar] [CrossRef]
- Brodu, N. Super-Resolving Multiresolution Images with Band-Independent Geometry of Multispectral Pixels. IEEE Trans. Geosci. Remote Sens. 2017, 55, 4610–4617. [Google Scholar] [CrossRef] [Green Version]
- Moore, I.D.; Grayson, R.B.; Ladson, A.R. Digital Terrain Modelling: A Review of Hydrological, Geomorphological, and Biological Applications. Hydrol. Process. 1991, 5, 3–30. [Google Scholar] [CrossRef]
- Wilson, E.H.; Sader, S.A. Detection of Forest Harvest Type Using Multiple Dates of Landsat TM Imagery. Remote Sens. Environ. 2002, 80, 385–396. [Google Scholar] [CrossRef]
- Xu, H.Q. Modification of Normalised Difference Water Index (NDWI) to Enhance Open Water Features in Remotely Sensed Imagery. Int. J. Remote Sens. 2006, 27, 3025–3033. [Google Scholar] [CrossRef]
- Vedaldi, A.; Soatto, S. Quick Shift and Kernel Methods for Mode Seeking. In Computer Vision—ECCV 2008, Proceedings of the European Conference on Computer Vision, Marseille, France, 23–28 August 2008; Springer: Berlin/Heidelberg, Germany, 2008; pp. 705–718. [Google Scholar]
- Haralick, R.M. Statistical and Structural Approaches to Texture. Proc. IEEE 1979, 67, 786–804. [Google Scholar] [CrossRef]
- Wang, H.; Zhang, L.; Yin, K.; Luo, H.; Li, J. Landslide Identification Using Machine Learning. Geosci. Front. 2021, 12, 351–364. [Google Scholar] [CrossRef]
- Wang, H.J.; Zhang, L.M.; Luo, H.Y.; He, J.; Cheung, R.W.M. AI-Powered Landslide Susceptibility Assessment in Hong Kong. Eng. Geol. 2021, 288, 106103. [Google Scholar] [CrossRef]
- James, G.; Witten, D.; Hastie, T.; Tibshirani, R. An Introduction to Statistical Learning; Springer: New York, NY, USA, 2013. [Google Scholar]
- Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: London, UK, 2016. [Google Scholar]
Band ID | Name | Original Resolution | Super Resolution | Statistics of Super-Resolution Products | |||
---|---|---|---|---|---|---|---|
Minimum | Maximum | Mean | Standard Deviation | ||||
1 | Coastal Aerosol | 60 | - | - | - | - | - |
2 | Blue | 10 | 10 | 0 | 1.87 | 0.75 | 0.53 |
3 | Green | 10 | 10 | 0 | 1.77 | 0.72 | 0.48 |
4 | Red | 10 | 10 | 0 | 1.69 | 0.67 | 0.43 |
5 | Red Edge 1 | 20 | 10 | 0 | 1.52 | 0.67 | 0.41 |
6 | Red Edge 2 | 20 | 10 | 0 | 1.3 | 0.64 | 0.37 |
7 | Red Edge 3 | 20 | 10 | 0 | 1.2 | 0.61 | 0.33 |
8 | Near Infra-Red (NIR) | 10 | 10 | 0 | 1.5 | 0.63 | 0.33 |
8a | Narrow NIR | 20 | 10 | 0 | 1.11 | 0.57 | 0.29 |
9 | Water Vapor | 60 | - | - | - | - | - |
10 | Cirrus | 60 | - | - | - | - | - |
11 | Short Wave Infrared 1 (SWIR1) | 20 | 10 | 0 | 1.44 | 0.11 | 0.09 |
12 | Short Wave Infrared 2 (SWIR2) | 20 | 10 | 0 | 1.73 | 0.1 | 0.07 |
Kernel Size | Maximum Distance | ||||
---|---|---|---|---|---|
2 | 4 | 8 | 16 | 32 | |
2 | 3,484,716 | 703,033 | 77,848 | 35,329 | 30,263 |
4 | 3,567,314 | 674,536 | 31,822 | 9333 | 7391 |
8 | 3,601,859 | 711,186 | 17,106 | 3737 | 1989 |
16 | 3,618,718 | 749,797 | 17,466 | 1210 | 845 |
32 | 3,629,014 | 782,430 | 23,571 | 798 | 306 |
Kernel Size | Maximum Distance | Segmentation Loss (%) | Object PA (%) | Object UA (%) | Object IoU (%) | Object OA (%) | Pixel PA (%) | Pixel UA (%) | Pixel IoU (%) | Pixel OA (%) |
---|---|---|---|---|---|---|---|---|---|---|
2 | 8 | 1.17 | 99.42 | 98.78 | 98.21 | 98.76 | 98.84 | 98.35 | 97.23 | 98.07 |
16 | 2.04 | 99.50 | 99.25 | 98.76 | 99.15 | 98.08 | 98.45 | 96.58 | 97.62 | |
32 | 2.64 | 99.30 | 99.61 | 98.91 | 99.26 | 97.39 | 98.43 | 95.90 | 97.15 | |
4 | 8 | 1.74 | 99.23 | 99.41 | 98.65 | 99.07 | 98.04 | 98.66 | 96.76 | 97.75 |
16 | 2.61 | 99.52 | 99.65 | 99.18 | 99.44 | 97.29 | 98.60 | 95.96 | 97.19 | |
32 | 3.09 | 99.39 | 99.87 | 99.26 | 99.50 | 96.71 | 98.61 | 95.41 | 96.81 | |
8 | 8 | 2.81 | 99.54 | 99.68 | 99.23 | 99.48 | 97.06 | 98.57 | 95.71 | 97.02 |
16 | 3.12 | 99.70 | 99.84 | 99.54 | 99.69 | 96.76 | 98.58 | 95.43 | 96.82 | |
32 | 3.34 | 99.64 | 100.00 | 99.63 | 99.75 | 96.41 | 98.68 | 95.18 | 96.65 | |
16 | 8 | 3.38 | 99.54 | 99.82 | 99.36 | 99.57 | 96.29 | 98.56 | 94.95 | 96.49 |
16 | 3.63 | 99.71 | 98.53 | 98.25 | 98.81 | 96.43 | 97.48 | 94.09 | 95.85 | |
32 | 3.69 | 100.00 | 98.81 | 98.81 | 99.20 | 96.47 | 97.66 | 94.29 | 96.00 | |
32 | 8 | 4.16 | 99.73 | 99.53 | 99.27 | 99.50 | 96.15 | 97.61 | 93.94 | 95.75 |
16 | 5.30 | 99.94 | 99.72 | 99.67 | 99.78 | 95.13 | 97.14 | 92.54 | 94.74 | |
32 | 5.43 | 99.79 | 99.29 | 99.08 | 99.38 | 95.08 | 96.73 | 92.12 | 94.42 |
Dataset + Algorithm | Feature Number | Object PA (%) | Object UA (%) | Object IoU (%) | Object OA (%) | Pixel PA (%) | Pixel UA (%) | Pixel IoU (%) | Pixel OA (%) |
---|---|---|---|---|---|---|---|---|---|
MSID 1 + CNN 2 | 120 | 98.69 | 98.86 | 97.58 | 98.32 | 98.15 | 98.47 | 96.67 | 97.68 |
MSID + RF 3 | 120 | 99.42 | 98.78 | 98.21 | 98.76 | 98.84 | 98.35 | 97.23 | 98.07 |
MSID + MSIDD 4 + CNN | 138 | 98.83 | 98.72 | 97.59 | 98.33 | 98.29 | 98.33 | 96.68 | 97.69 |
MSID + MSIDD + RF | 138 | 99.41 | 98.77 | 98.20 | 98.75 | 98.84 | 98.34 | 97.21 | 98.06 |
MSID + DEMD 5 + CNN | 150 | 98.85 | 98.68 | 97.56 | 98.31 | 98.31 | 98.29 | 96.66 | 97.67 |
MSID + DEMD + RF | 150 | 99.10 | 98.85 | 97.98 | 98.60 | 98.53 | 98.43 | 97.01 | 97.92 |
MSID + MSIDD + DEMD + CNN | 168 | 99.24 | 98.33 | 97.60 | 98.33 | 98.70 | 97.94 | 96.69 | 97.69 |
MSID + MSIDD + DEMD + RF | 168 | 98.94 | 98.90 | 97.87 | 98.52 | 98.38 | 98.49 | 96.92 | 97.85 |
ML Model | Object PA (%) | Object UA (%) | Object IoU (%) | Object OA (%) | Pixel PA (%) | Pixel UA (%) | Pixel IoU (%) | Pixel OA (%) |
---|---|---|---|---|---|---|---|---|
CNN trained with MSID, KZ = 2 and MD = 8 in Bome County | 96.06 | 91.48 | 88.17 | 97.02 | 93.86 | 90.65 | 85.58 | 96.29 |
RF trained with MSID, KZ = 2 and MD = 8 in Bome County | 99.12 | 92.54 | 91.78 | 97.95 | 96.89 | 91.73 | 89.11 | 97.22 |
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |
© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
Wang, H.; Zhang, L.; Wang, L.; He, J.; Luo, H. An Automated Snow Mapper Powered by Machine Learning. Remote Sens. 2021, 13, 4826. https://doi.org/10.3390/rs13234826
Wang H, Zhang L, Wang L, He J, Luo H. An Automated Snow Mapper Powered by Machine Learning. Remote Sensing. 2021; 13(23):4826. https://doi.org/10.3390/rs13234826
Chicago/Turabian StyleWang, Haojie, Limin Zhang, Lin Wang, Jian He, and Hongyu Luo. 2021. "An Automated Snow Mapper Powered by Machine Learning" Remote Sensing 13, no. 23: 4826. https://doi.org/10.3390/rs13234826