I have built a vegetation map from a set of historical photographs from 1912. The photographs were taken from the ground (ridges and peaks of mountains), and I have developed a method of georeferencing these images and extracting gridded data of the vegetation types visible.

However, there are many portions that were not visible in this landscape due to being obscured behind ridges. These locations were not visible from other camera photo points. This is not a dissimilar problem to aerial photography with lots of obscuring clouds. The difference is there are not other photos to use to fill in the gaps. I need to use some sort of interpolation routine to fill the gaps.

I am currently using ArcGIS 10.2.2 and would like a solution that works within Arc, but if it is not possible, I will entertain using R or QGIS or other packages.

I need to interpolate the vegetation categories in the obscured areas. My hypothesis is that it is driven by aspect/slope/elevation and of course influenced by the nearby vegetation. I think this is some variant of co-kriging, but I have not been able to track anything down that deals simultaneously with categorical and polygon data. The data below in the image is in 100m square cells.

Does anyone have any suggestions of approaches to take? I would like to bootstrap the data so that I can test its predictive power, rather than just interpolate what is missing.

Attached is a map showing what I am dealing with. THe different colours indicate different vegetation categories (ie coniferous forest = dark green, deciduous = pale green, shrubs = orange, grassland = tan, open canopy forest = black). The grey areas are the obscured areas. Vegetation Map 1912


There is not enough room in comments box this is why I post it as possible solution. I think that ArcGIS maximum likelihood classification will do exactly what you want excluding vegetation neighbour.

Get elevation raster, scale it to range (0..255) using min, max and range. Do the same with slope. With aspect I'd suggest first to convert it to something less 'circular', e.g. Cos(aspect). Convert vegetation raster to polygons and obtain signatures for them using above 3 rasters for training.

It might worth to replace slope and aspect by solar radiation raster and use 2 rasters only


I am going to make 2 assumptions:

  1. It is continental climate, i.e. there is snow in winter
  2. You are dealing with natural vegetation, i.e. no human intervention like forestry or irrigation.

If above true and taking soil out of equation, in a very general terms there are 2 limiting factors for vegetation growth: air temperature and moisture availability.

Classic example are high mountains in arid areas. During the climb it goes from desert grass (not enough moisture) to forest (warm enough, enough precipitation) to alpine shrub and grass (plenty of water, short summer). In this ideal world DEM would suffice in gap filling.

However it is known, that within the same elevation band there are forested and treeless areas. Most of the time this is due to wind impact on snow, there are more snow on sheltered slopes and less on slopes exposed to wind. It would be great to obtain data from weather bureau on snow water equivalent vs elevation and aspect and convert such table to raster. It could hugely improve extrapolation results

  • Thanks for the reply, but that isn't what I am after. I have already classified the imagery (if you want an idea of what I am doing, I have just published a paper on it at APPLIED GEOGRAPHY 63:315-325 · AUGUST 2015 (link = sciencedirect.com/science/article/pii/S0143622815001770 ). the article is free to access until Sept 17. What I need is to develop a predictive algorithm to fill in all the areas that are grey. I have no imagery to classify for these areas (grey in map above), I just have the imagery in the area that is already classified (the green, tan, etc). – ChrisStockdale Aug 31 '15 at 18:10
  • I inderstood what you want very well. You didn't understand what I suggest at all. No mention of imagery in my solution, the only thing you need is DEM. – FelixIP Aug 31 '15 at 18:33
  • Okay, sorry. I was reading through the description of the Maximum Likelihood Classification and it seemed to be describing a classification procedure for a raster, not predicting what is in areas not covered by the raster. I will dig deeper and see where this leads. – ChrisStockdale Aug 31 '15 at 18:35
  • One of the most critical components of this interpolation is the relationship of the neighbouring vegetation: there is significant spatial autocorrelation in this data, and any interpolation needs to account for what is nearby in terms of vegetation. – ChrisStockdale Sep 1 '15 at 19:46
  • You might consider N more rasters each is a distance from individual vegetation class. – FelixIP Sep 1 '15 at 20:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.