Hey Jie Dou,

could you specify what exactly you are interested in, how to get non-landslide sample technically (for example in GIS) or a logical procedure how to do this?

Hey, Jewgenij Torizin,

Thanks for your reply. I mean how to get non-landslide technically if the landslide inventory map is the point of landslide. As we know, maybe one point landslide, the neighbor points, maybe are also landslide? How to avoid this?

Well, a point is a very approximative representation of a landslide and will have an uncertainty, that is correct. If you have an idea about the mean size of the landslide bodies you could define the uncertainties of the landslides location represented by a point. Then it is appropriate to generate a buffer around the landslide points according to the uncertainty ( e.g. 20-50 m). After you can use the remaining area (ouside the buffers) for random sampling.Otherwise, even if you have no information of the landslide size you could still try the buffering, but then the buffer distance should be selected carefully. So you would most likely avoid to have landslide points in a non-landslide sample.

But a I have also a question, why are you using points to charcterize landslides from aerial images? What kind of images do you use (in which ground resolution)? What is the expected scale of analysis? Since you generate a sample for non-landslides points I guess you're giong to do a statistical analysis, logistic regression?

I'm asking because, if the resolution of the images allows the direct mapping of the landlide extent I would use polygons instead of points to do that. For high resolution images even a separation into source and deposition areas is recomended, since after you could gain additional information via magnitude-frequency statistics.

Best Regards,

Jewgenij

Dear Dou,

Torizin has put a valid question. Generally, the aerial images are of good resolution. So, its better to go for mapping landslides as polygons. And then no-landslide areas can also be taken as polygons for training samples. You can also consider landslide and no-landslide areas as pixels in raster GIS and can do the analysis.

Dear D. P. Kanungo,

Thanks for your answers. I read your several papers in relation to landslide hazard map and also cite them. Very good. 

By the way, if the landslide are  vector of points, but it is hard to select non-landslide points, because maybe there are also overlaid the landslide point? how to avoid this? However, polygon has no this problem.

Dear Torizin,

Thanks so much for your kind explanation. 

In my cases, there is more than thousand of landslides, if draw the boundary of landslide, it will take a long time.  Moverover, in my study area, just one scene 0.6m quickbirds covers part of the study area, there other areas no high-resolution  images, just have 20 m, 15m and 30m images, those images cannot see the clear boundary of landslides. That's why I use landslide point to replace the polygon of landslide.

As you said, it's better to do buffer, here problems come, how to select the distance of the buffer, this is uncertainty. 

First, I want to create landslide inventory map. And then, do a statistical analysis to find some rules, following use some models to produce hazard maps.

Regards,

Dou

Dear Dr. Dou,

Thanks for your appreciation. Now I got your point. You have point landslide information available. So, if you are analysing those through raster GIS, then take those landslide points as pixels. Then, map the non-landslide points from the aerial photographs or any remote sensing image by seeing the geomorphology and topography of the area. Don't forget to overlay landslide points on the image, while marking the non-landslide points. I think this is the only way out to map the non-landslide points.

Regards.

Kanungo

Further, I got to know that you have 15-30m resolution images. So, take the landslide points as pixels of say 20m size, which is the nominal size of your image resolution. By, superimposing these landslide pixels on the image, map the non-landslide pixels keeping in view the topographical and geomorphological features of the area which often guides us to identify non-landslide pixels from the landslide pixels.

Regards.

Dear Dr. D. P. Kanungo,

Many thanks for your kind explanation.

Briefly,  you mean also get non-landslide points from the satellite images or the others like topographical and geomorphological features? But this will take time. I want to create the random point from the raster without the landslide raster.This is faster. But there is a premise, raster should not include the landslides. If the landslides are polygons, use mask to extract it easily.  However, point of landslides, it is hard. Many people use points and maybe use buffer, there is some problem due to buff also including the landslide points. What do you think that?

Anyway, thanks for your suggestions.

Regards

Dou

Yaah, I agree. You can mask those pixels with the landslide points. One assumption will be that the landslide points mapped are the total landslides in the area. Then from the rest of the image, you can go for stratified random sampling to select non-landslide points.

Regards.

Kanungo

Taking a lead from all the replies, we can follow the below mentioned set of procedure:

• Calculate Euclidean Distance on Landslide Data points
• Use Raster Calculator to obtain the binary raster. In this step, we can define the buffer distance on which the Non-landslide point shouldn't fall (e.g. "LANDSLIDE_DATA" >= 100 ) so that output file will have 1 for distances greater than 100 m from the Landslide points and 0 otherwise.
• Now on this binary output file we can use "Create Spatially Balanced Point" tool in ArcGIS to generate the non-landslide points. The input binary raster defines defines the inclusion probabilities for each location in the area of interest. The location values range from 0 (low inclusion probability) to 1 (high inclusion probability).

Alternatively, one can also make the slope as a criteria to define the inclusion probability. For example there are more chances of landslides to be on localities of higher slopes and smaller the slope, the chances of landslide will be less (in general). So based the LS data point and comparing with the slope of area of interest, one can decide the minimum slope threshold and can get the binary raster for the same.

Thanks and regards,

Navdeep Agrawal

Sir, I am sharing our current article. Please give your valuable opinion.

https://www.researchgate.net/publication/360917378_Landslide_susceptibility_mapping_of_Kalimpong_in_Eastern_Himalayan_Region_using_a_Rprop_ANN_approach