Landslide is a natural disaster defined as a mass of rock, debris, or earth down a slope. Landslides are denoted by the down-slope movement of soil and rock under the direct influence of gravity. The term "landslide" encompasses five modes of slope movement: falls, topples, slides, spreads, and flows [1][2][3].


In recent years, satellite technology and remote sensing technology are developing fast. The application of satellite remote sensing to capture the Earth is rapidly increasing in number and image quality. It plays a significant role in the Earth’s surface monitoring.
![Figure 3. Landslide region from satellite image - image source from [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0-o504r5qzrYnEdWD.gif)
Our primary target is to understand and implement the methodology to detect Landslide from satellite images in this paper. From landslide regions detected at different times, a landslide scaling (landslide is developing or reducing) is presented. This article is an explanation version of my public paper under the title "Deep learning for landslide recognition in satellite architecture." The details of How to do and implements the second section of my paper is presented. Besides, I also share step by step source code of my publication. I hope I can bring a solution from a research paper to implementation in the real world. With this purpose, please share my article or refer to my publication if you are going to publish a research paper.
T. A. Bui, P. J. Lee, K. Y. Lum, C. Loh, and K. Tan, "Deep Learning for Landslide Recognition in Satellite Architecture," IEEE Access, vol. 8, pp. 143665–143678, 2020, doi: 10.1109/ACCESS.2020.3014305.
![Figure 4. Deep Learning for Landslide Recognition in Satellite Architecture - Graphic Abstract [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0x63-vmNhsKsQ2W2t.jpg)
Figure 4 shows a short process of landslide detection by the combination between a CNN model and an H-BEMD algorithm is introduced. From satellite images captured at different times (September 15, 2013, and September 18, 2014), we can announce landslide is more significantly changed and easy to detect landslide scaling (landslide direction).
1. Landslide region detection by H-BEMD (Hue – Bi-Dimensional Empirical Mode Decomposition)
Hue – Bi-dimensional empirical mode decomposition (H-BEMD) is an algorithm introduced in paper [4].
1.1. Why using a Hue channel in satellite images to detect objects?
HSV separates to luminance from chrominance, also known as image intensity. Satellite images are governed by weather, especially the lighting conditions for each shot are different. Figure 5 shows RGB satellite images of the same location under other lighting conditions.
![Figure 5. RGB, Histogram, and Hue channel [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0J4ca_sr_MuE9KsuC.gif)
Although the histogram values between three images (a,d,g) are different, the hue channels are the same. Therefore, the hue image channel is selected to detect landslide objects in satellite images.
To convert the RGB image to the Hue channel image, OpenCV is a good and easy way. Basically, OpenCV changes RGB to HSV channel by 8bit/value (np.uint8 – from 0 to 255). However, the Hue value is presented by a circle (from 0 to 360 degrees). Thus, the bellow code shows the methodology to change RGB to HSV with a full value range. In this case, an unsigned integer 16 bit is applied to Full Hue values.
1.2. H-BEMD flowchart and source code
In this part, I will present the detailed H-BEMD algorithm. Firstly, figure 6 shows the process of how to get a transformed image by H-BEMD.
![Figure 6. H-BEMD flowchart [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0y1r5gtraUDhkp957.gif)
What is H-BEMD? H-BEMD is an improved version of the BEMD algorithm. Original BEMD [5] is applied to a Grey image from RGB, which has a value range (in 8 bit/value) from 0 to 255. However, to apply on the Hue channel, BEMD doesn’t get corrected results (landslide region) because the main important key points in BEMD is detecting extrema of an image signal. Therefore, Sin and Cosin value of Hue is applied to Hue value. Then, the extrema of sin and cosine are detected. Finally, arctan2 is used to combine extrema sin and cosine. It is an extremum value of the Hue channel.
Continue, I present the detail of how H-BEMD work to get _imfs_coshue and _imf_sinhue.
What is the detail of H-BEMD?
The sin(hue) and cos(hue) values are denoted by θ and φ, respectively. H-BEMD adaptively decomposes sine and cosine of hue image (θ and φ ) through H-BEMD sifting process as flow:
- Step 1: Detect the extrema (both maxima and minima) points of θ by a morphological algorithm.
- Step 2: Connect the maxima and minima points of θ, respectively, with a Radial Basis Functions (RBFs) [23] to generate the new 2D ‘envelope.’
- Step 3: Normalize the 2D ‘envelop’ as presented in section 2.2.2.
- Step 3: Determine the local mean mθ by averaging the two envelopes.
- Step 4: Subtract out the mean from the image: ϑi=θ−mθ ;
- Step 5: Follow step 1 to 4 on φ, ωi=φ−mφ
-
Step 6: Repeat the sifting process with θ=ϑi and φ=ωi and i:=i+1
Figure 7 is the result of the above source code.
![Figure 7. Landslide region detection by H-BEMD [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/1FawPBVm-J82zZM1GZ9k2Vw.png)
2. Landslide regions segmentation
Firstly, we focus on the short review of the Hue channel. Figure 8 shows the architecture of Hue value.
![Figure 8. Hue channel [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0huKg5WBh59F5Lg4d.gif)
This article introduces to define a landslide region by value range between 330° to 90°. The detailed of python code to define landslide is:
3. Landslide size and direction detection
Two satellite images captured at different times are used to show the testing flowchart in this section.
![Figure 9. Landslide object from the satellite image. (a) Jure landslide in Nepal was captured on September 15, 2013 - time t1. (b) Jure landslide in Nepal, captured on September 18, 2014 - time t2 [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0cljL3aEwjYi8_JLU.gif)
Figure 10 is the H-BEMD result of Figure 9.
![Figure 10. H-BEMD results [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0E2Zb3y9UHcC56Sji.gif)
![Figure 11. Landslide labeling of Figure 10 [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0puT2HbAo-IzbwCwG.gif)
![Figure 12. Landslide region location. [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0CnduY-p6-YJwUcm5.gif)
From Figure 12, a centroid point of each landslide region is defined. Figure 13 and below source code show the centroid point detailed and how to get centroid point in python code.
![Figure 13. The centroid of the landslide region. [4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0h9yvvq7Jd2jRwWIK.gif)
The direction of the landslide scaling is defined by a green point and redpoint. Detailed source code below. Figure 14 shows the results of the source code.
![Figure 14. Landslide direction[4]](https://towardsdatascience.com/wp-content/uploads/2020/10/0tzOyFzeR8C9DZMid.gif)
Based on landslide direction, we can easily define what region is dangerous and easily affected by its impact. From that will accurately forecast people living in the affected areas.
I hoped this article helpful to you and made a short view of applying satellite data to life.
References: [1] https://www.usgs.gov/faqs/what-a-landslide-and-what-causes-one [2] https://www.taiwannews.com.tw/en/news/1238060 [3] https://vov.vn/en/society/massive-landslides-bury-residents-and-houses-in-lai-chau-377950.vov [4] T. A. Bui, P. J. Lee, K. Y. Lum, C. Loh, and K. Tan, "Deep Learning for Landslide Recognition in Satellite Architecture," IEEE Access, vol. 8, pp. 143665–143678, 2020, doi: 10.1109/ACCESS.2020.3014305. [5] J. C. Nunes, Y. Bouaoune, E. Delechelle, O. Niang and P. Bunel, "Image analysis by bidimensional empirical mode decomposition," Image Vis. Comput., vol. 21, no. 12, pp. 1019–1026, Nov. 2003.