Comparison of various remote sensing classification methods for landslide detection using ArcGIS

CM Escape (a,b), MK Alemania (a), PK Luzon (a), R Felix (a,b), S Salvosa (a), D Aquino (a), RN Eco (a,b), AMF Lagmay (a,b)

(a) Nationwide Operational Assessment of Hazards, University of the Philippines, Diliman, Quezon City, Philippines
(b) National Institute of Geological Sciences, University of the Philippines, Diliman, Quezon City


A comprehensive landslide inventory is vital in landslide hazard analysis. It provides statistical and spatial distributions at a given time which can be used as parameter for susceptibility and classification modelling. It is usually derived from historical data, field surveys, and manual interpretation of aerial and satellite images. However, historical data is not always available and complete, intensive field surveys are impractical for large-scale studies, and manual analysis of aerial and spectral images can be tedious and time-consuming. With the advancement of spectral remote sensing systems, different automated procedures for image classification have been developed. To test the effectiveness of various automated image classification methods, we compared several procedures utilizing spectral images taken after the Mw 7.2 Bohol (Philippines) earthquake on October 15, 2013 instead of a comprehensive landslide inventory. These procedures included: 1.) an unsupervised ISODATA clustering classification, 2.) a supervised maximum likelihood classification using raw spectral bands, 3.) another supervised classification using the Normalized Difference Vegetation Index (NDVI), and 4.) a manual reclassification of NDVI values using specific ranges. We used the fourth method to highlight the difference between using its unbiased mathematical data with supervised classification training sites that has an added human factor. We then compared each image classification with the manual inventory done to determine its accuracy. The unsupervised classification had the lowest accuracy and reliability in distinguishing the landslides. The supervised classification using raw spectral bands, though it showed clear regions of landslides, only distinguished 75% of the landslides manually inventoried. Both methods that involved NDVI were more useful for landslide identification but had different advantages. The supervised classification with NDVI was more useful in pinpointing landslide areas because of the high contrast of barren soil and earthflows to grass/forest and urban areas. It identified 88% of the previously pinpointed landslides. On the other hand, the manually reclassified NDVI showed a better delineation of the landslide area and detected 82% of the landslides.

1. Introduction

Detailed landslide inventory is important in determining and predicting landslide hazards. However, traditional methods such as interpretation of aerial photographs, manual analysis of slope and contours and direct observation from field and historical data tend to be tedious, subjective and incomplete. The limitations and errors from these methods have a direct effect on the accuracy of the landslide statistics, which in turn, could debase the hazard assessment. Through the years, various techniques had been developed for faster and more accurate landslide detection. Since land- slide events produce unique textural characteristics, surface texture analysis using High Spatial Resolution images had been developed to detect between rough and smooth surfaces and stable and unstable zones for different environments [1].Another approach is through image enhancement techniques such as color composition analysis (false and real color) and use of different indexes of image bands such as Normalized Difference Vegetation Index (NDVI). This particular form of vegetation index uses the contrast between the vegetation’s reflectance of near-infrared (NIR) light and the absorption of red light by chlorophyll [2].


where NIR=Near Infrared and VIR=Visible (Red)

Since it uses information from cluster of vegetation, areas presenting the best contrast between barren soil and closed vegetation cover show the best result. Alongside this is the development of manual and automatic detection using different classification techniques. Through these classifications, identical pixels within remotely sensed data are grouped into classes based on their comparison to a known sample or to one another. Good classification results depend on the quality of the remotely sensed data.

Consequently, various classification techniques give different result and accuracy depending on the circumstance of the study. It is therefore vital to determine what type of remote sensing data and classification should be used for a specific aim. For this particular study, four methods were used. Real color and false color compositions were subjected to an unsupervised ISODATA (Iterative Self-Organizing Data Analysis Technique) classification, a supervised maximum likelihood classification using NDVI and a manual reclassification of NDVI using specific ranges. An unsupervised ISODATA classification, assigns a pixel into a class, without using defined training classes. This iterative procedure assigns an arbitrary cluster center, classifies the pixels, calculates the cluster mean and covariance and repeats the whole process until the minimum change between iterations is achieved [3]. A supervised classification on the other hand assigns pixels based on classes predefined by manual training process. The Maximum Likelihood (ML) classification is a parametric algorithm that approximates using a normal distribution [4]. The quality of the training site dictates the quality of the supervised classification. The more training sites are identified, the better the result of the classification would be. Once the training sites are defined, the signatures – which are the statistical characterization of the sites – are created. These signatures will then be used for the classification of the image [5].

2. Study Area and Geology

The site for this study was the province of Bohol, Philippines. Last October 15, 2013, a Mw 7.2 earthquake shook the island. The epicenter was in between the towns of Sagbayan and Catigbian. The earthquake affected various areas prone to natural hazards such as liquefaction (coastal and riverside towns like Maribojoc, and Calape), landslides (elevated towns of Sagbayan and Catigbian) , ground subsidence (Cortes and Loon) and sinkhole formation (Tagbilaran) [8]. The major earthquake event and the other 885 earthquakes recorded the next day by the Philippine Institute of Volcanology and Seismology (PHIVOLCS) [9] had caused failure of structures and buildings, destruction of landforms (especially the Chocolate Hills) and death of hundreds of people. Previous mapping efforts by PHIVOLCS had identified one major fault line traversing the southern end of the island – the East Bohol Fault (Fig. 1). However, what triggered the earthquake was a new reverse thrust fault mechanism that strikes in the NE direction and dips SE [8]. The towns surrounding the epicenter was part of the Maribojoc Limestone, a Pliocene formation responsible for the Karst topography famously known as Chocolate Hills (Fig. 2). A five-kilometer fracture zone which was the manifestation of the “blind” fault was found in the town of Inabanga.


3. Methodology

The scope of the images used was limited to the NW-SW portion of the island, where the earthquake had occurred (Fig. 3). Satellite images taken right after the earthquake were used for manual identification of landslide and contrast analysis of the classification methods. Images from Pleiades 1A, Quickbird, Worldview-2 and Landsat 8 satellite as well as aerial photographs from unmanned aerial vehicle (UAV) were utilized.The images were pan sharpened, filtered and analyzed using ArcGIS 10. A total of 515 landslides were identified. After the manual delineation of the landslide, an unsupervised classification of both the real color images and false color images was performed. The number of classes was set to twelve so as to refine the frequency distribution of the result. The same step was performed using the supervised maximum likelihood classification. Training sites were assigned in clouds, urban areas, landslides, forest/grassland and water. Forty-five training sites were established for each images, with the most focus on landslide areas. Using the vegetation band combination, an NDVI image was produced, by dividing the difference of the NIR and red bands by the sum of the same bands. The image produced shows varying hues of blue and light green for man-made structures as well as barren soil, orange for water and light yellow for clouds. The produced images were then subjected to a new training process for a maximum likelihood classification. A separate NDVI image was created and reclassified by setting our own range for the different sites. The range was based on the average pixel value as inspected manually from ArcGIS.



4. Results and Discussion

When the images were subjected to a false color composite image (Fig. 4B) the contrast between the vegetation and the barren soil became more apparent. In the unsupervised ISO DATA classification, twelve classes were defined though there it could generally be classified into five: clouds, vegetation, barren soil, water and urban areas. It was noticeable how efficient the method was in classifying water and vegetation (Fig. 4C).However, it had problems with classifying the clouds, barren soil/landslide area and urban area. For the clouds, there was confusion with the area where the cloud shadow falls. Though it should be noted that even if the classification had grouped the cloud shadows and the water area, it had delineated the shadow and classified its border as clouds. The same went for the actual clouds though it had identified the border as barren soil/landslide. The urban areas were barely identified. Most of them were classified as barren soil/landslide though such result could be expected because most of the urban areas have exposed soil. As for the barren soil/landslide, the classification was able to even distinguish the coastline. The large landslides that were identified could easily be seen.


However, the smaller ones were barely identified (Fig. 4D). The ML supervised classifier gives better result when there is a large number of training sites with good uniformity and representation. For each image, 45 sites were identified, at least 10 of which were for landslide areas. A data that could provide good separability and characterization of the different classes also enhances the result. For this reason, the false composite image which highlights the contrast of barren soil and vegetation was very useful. The resulting image successfully classified the clouds and the cloud shadow, the water bodies and the vegetation (Fig. 5A). However, the same conflict between urban areas and landslide areas was manifested in this result. However, compared to the unsupervised ISODATA classification, smaller landslides were detected and it had better delineation of the areas.

Aside from false color rendition, NDVI (Fig. 5D) which in simple terms measures the vegetation’s greenness and photo- synthetic activity, is also effective in characterizing vegetated areas and barren soil/landslide areas. However, its application is not limited to identifying vegetation types. Since it depends on the reflection of near infrared wavelengths and visible wavelengths, different structures give different values. For example, standing water have low reflectance on both bands (NIR and VIS) and therefore, they give negative values. Soils have higher NIR reflectance compared to its VIS reflectance so its values usually range between small positive numbers. However, it also has limitations such as thin clouds and cloud shadows which could have altered values due to its combined reflectance and the reflectance of the underlying structure. The NDVI image that was subjected to ML supervised classification had, as expected, no problem in identifying vegetated regions as well as water bodies (Fig. 5B). The cloud classification has generally good results though the limitation with the thin clouds and cloud shadows was manifested in the results when they were classified as urban areas. On the other hand, the barren soil/landslide points were clearly highlighted and identified though there was still minor confusion with urban areas. As discussed earlier, different structures show NDVI values within a certain range. The last test performed was to reclassify the NDVI raster dataset, such that the landslide values would be highlighted and the range of values would be crafted based on the images. Before reclassifying, the pixels of the images were manually inspected and a range of values for each types (clouds, water bodies, barren soil/landslide, urban area and vegetation) were tailored based on the results (Table 1). The result showed a cleaner image with more defined boundaries in between different sites (Fig. 5C). However, the same conflict between barren soils/landslides and clouds and urban areas was still apparent. With the cleaner interface, the landslide areas were more defined. Clearly, the source of error of the images was the clouds and the cloud shadows. Majority of the confusion was with the clouds and the barren soil/landslides. Cloud elimination on satellite images is still an ongoing and is a developing study [10] [11]. However, a simple but pain-staking solution is to mask all the clouds and cloud shadows. It would eliminate the clouds and therefore the confusion in classifying them. However, it is tedious, time-consuming and eliminates huge data, especially for those land area under the cloud shadow. Another issue was how to distinguish between identified barren soil and landslide. Since landslide is basically barren/exposed soil, there is no way to distinguish them from each other through classification methods. However, slope and contour analysis would easily dissipate this problem. Identified areas with flat-lying zone could be eliminated since most landslides occur in areas with high gradient. However, careful inspection should be implemented for there are landslide events that defy this.c6

Via visual inspection, the unsupervised ISODATA classification identified the least number of landslides, whether it was using the natural color composite or the false color composite. The supervised ML classification showed a better result. Though the classification between the barren soil/landslide areas and urban areas wasn’t well established, 385 landslide

5. Conclusion

There are various ways to produce a comprehensive land- slide inventory. Traditional expert-based inventory are still widely employed, however, with the latest advancements in remote sensing and computer statistics, new ways are being developed and employed. Combining different techniques such as production of composite images via band manipulation or index production (such as NDVI) with different types of classification (such as the unsupervised ISODATA and supervised maximum likelihood classification) could result to multivariate results. It is all a matter of determining what datasets would be used and how to use them. Automatic detection of landslides would lessen the time and effort invested in making landslide inventories. By combining the two types of classification and the manipulation of the satellite image, different results had been achieved. Clearly the most effective of them was the use of NDVI image and maximum likelihood. The NDVI serve to heighten the differences between clouds, water bodies, urban areas, vegetation and barren soil/landslide. The supervised maximum likelihood classification automates the process with very little sacrifice on the accuracy. The study had given favorable results but these could further be enhanced by other analysis such as cloud detection and elimination and/or object-oriented analysis. After this analysis, further study should be done in order to create a more accurate but more time-efficient process for landslide inventory.

6. References

[1] T. Fernandez, J. Jimenez, P. Fernandez, R. E. Hamdouni, F. Card, J. Delgado, C. Irigaray, J. Chacon. Automatic detection of landslid features with remote sensing techniques in the Betic Cordilleras (Granada, Southern Spain), The International Archives of the Photogrammetry. Remote Sensing and Spatial Information Sciences.

[2] D. De Alwis, Z. Easton, H. Dahlke, W. Philpot, T. Steenhuis. Unsupervised classification of saturated areas using a time series of remotely sensed images, Hydrology and Earth System Sciences Discussions (2007) 1663 – 1696.

[3] A. Ahmad, S. F. Sufahani, Analysis of landsat 5 tm data of Malaysian land covers using isodata clustering technique, IEEE Asia-Pacific Conference on Applied Electromagnetics (APACE 2012) (2012) 92 – 97.

[4] F. Guzzetti, A. C. Mondini, M. Cardinali, F. Fiorucci, M. Santangelo, K.-T. Chang. Landslide inventory maps: New tools for an old problem. Earth-Science Reviews (112) (2012) 42–66.

[5] K. Perumal, R. Bhaskaran, Supervised classification performance of multispectral images, Journal of Computing 2 (2) (2010) 124 – 129.

[6] P. I. of Volcanology, Seismology, Distribution of active faults and trenches in the Philippines, Map (2000).

[7] P. Planning, D. O. of Bohol, Geologic map of the Philippines, Online (2013). URL

[8] M. Aurelio, J. M. Rimando, K. J. Taguibao, J. D. Dianala, A. E. Berador, Seismotectonics of the magnitude 7.2 Bohol earthquake of 15 October 2013 from onshore earthquake and offshore data: A key to discovering other buried active thrust faults?, Geological Society of the Philippines, 2013.

[9] P. I. of Volcanology, Philippines 15 October 2013 magnitude 7.2 Bohol earthquake Philippine Institute of Volcanology and Seismology, Online (2009). URL

[10] C.-H. Lin, P.-H. Tsai, K. hua Lai, J.-Y. Chen, Cloud removal from multitemporal satellite images using information cloning, IEEE Transactions on geoscience and remote sensing 20 (20) (2013) 1–10.

[11] S. Martinuzzi, W. Gould, O. M. Gonzalez, Creating cloud-free landsat etm+ data sets in tropical landscapes: Cloud and cloud-shadow removal, Tech. rep., U.S. Department of Agriculture, Forest Service, International Institute of Tropical Forestry (2007).

Leave a Reply

Your email address will not be published. Required fields are marked *

+ twelve = thirteen