This is an interesting question. Would you kindly share the Scene IDs/Product IDs of these two Landsat Scenes? I would like to take a look at these Landsat scenes before making any specific comments.

It would be also helpful to know what image processing steps were carried out exactly and whether any ancillary information was used to take into account the factors affecting NDVI. You already mentioned atmospheric correction and use of rainfall data. But which specific atmospheric correction algorithm was used? For example, the NDVI product that can be ordered from USGS for free for a Landsat scene uses a different atmospheric correction algorithm compared to the one you may have used. Did you also use any image masks to remove optically thick clouds, cirrus clouds, shadow, snow/ice, terrain occlusion etc.

Why I'm interested in knowing the steps you have followed to calculate NDVI for these two scenes and its subsequent interpretation is because there are a number of factors that may effect NDVI including, 'atmospheric effects, clouds, soil effects, anisotropic effects, spectral effects'. Please go through the wikipedia page of NDVI, especially the section on 'Performance and Limitations' which provides a reasonable overview of each factors.

https://en.wikipedia.org/wiki/Normalized_Difference_Vegetation_Index

Thank you for answer. I think, the geographic coordinates of validation points will be more useful (see the other file, CRS WGS84 zone UTM34N), the dates of scenes you can find in previous file.

Our procedure was simple: 1) request for corrected scenes from USGS; 2) calculation of NDVI values using "raster calculator" of QGIS 2.18; 3) attribution of NDVI values to chosen points using "point sampling tool" of QGIS 2.18.

I must explain one issue. We work mainly on winter wheat, however, we want to check the comparability of NDVI obtained in different dates. Thus we chosen different points supposed to have constant NDVI, i.e. without vegetation, especially buildings, and river, sand isles on the river and the forest for comparison. It is clear, that the state of vegetation (NDVI of forest of wheat) must be different at different dates (with very few exceptions), however, why the roof of building differs in NDVI at these dates?

Michał, thanks for sharing the scene info and validation points. Alright, so let me explain what I did in order to analyse the problem. I ordered NDVI of two (13/6/2013 & 6/5/2014) landsat 8 scenes as well as the pixel_qa bands for both scenes through the ESPA USGS Interface (link attached). If you have never used the pixel_qa band before, I've attached the USGS link for your reference. Basically, pixel_qa band helps pin point optically thick clouds, cirrus clouds, snow/ice, terrain occlusion, water and clear terrain pixels within a landsat scene. I've also attached the link to download the ArcGIS toolbox to convert the bit packed pixel_qa band values into meaningful terms. The reason why I used the pixel_qa band in the NDVI analysis is because it helped to identify and subsequently mask out pixels that are cloud contaminated, which, if not avoided from NDVI calculation will result in anomalous values, and worse, would lead to incorrect interpretation. Therefore, my suggestion for you is to use uncontaminated pixels (clear pixels) for validation in both scenes.

The validation points that you shared were overlayed with both NDVI Landsat 8 Images. All the 8 validation points are present over the 'Lotnisko Chopina' apron ('airport tarmac'). Right now the problem as it stands is that, these 8 validation points fall over cirrus cloud contaminated pixels over the airport apron from the 13th June 2013 Landsat 8 scene. Optically thick clouds are obvious to identify but cirrus clouds can be tricky. Therefore, overlay 13th June 2013 Landsat 8 Band 9 (Cirrus Band) over the 2013 NDVI Image to get an idea and visualize the extent of cirrus cloud contamination. The validation points over the 6th May 2014 scene seems reasonable. Let me know if you also see the same problem that I do.

https://espa.cr.usgs.gov/ordering/new/

https://landsat.usgs.gov/landsat-surface-reflectance-quality-assessment

https://github.com/USGS-EROS/landsat-qa-arcgis-toolbox

What about other indices? I think you should consider EVI which is designed to reduce atmosphere influences. But, when there are clouds...

Thank you James for explanation. Perhaps next week we will check the clouds and other factors.

Any way, you can see another set of testing points, prepared by my colleague with results - see the file. But be aware, the coordinates are in "normal" WGS 84 (in degrees, DD).

Piot, for now we are working on NDVI, due to lack of time. Anyway, may be later we will check other indices.

Regards,

Michał

Michał, let me add some screenshots to emphasize my words. In order to test your assumption of observing 'constant' NDVI values over artificial and built-up surfaces in multi-temporal satellite images, you need to make sure there are no clouds and other constantly changing attributes that may adversely affect your interpretation. Remember, even though NDVI is derived from atmospherically corrected data, this does not remove clouds. For removing clouds from the scenes, an image mask is necessary. Here I've attached some screenshots with the following description:

With that said, I want to add that even if you find multi-temporal landsat images that are nearly similar in metadata information, still, I doubt whether you are going to find exactly the same NDVI values for 'non vegetated surfaces' because NDVI was not originally designed for this purpose. Morevoer, there are other factors that will affect the analysis, like, atmospheric correction algorithms are never 'perfect'. The solution is already provided in the wikipedia page, such as by using more frequent composite images, alternative indices etc. which Piotr also suggested in his comment. However, I would really like to see the results of your assumption being tested on NDVI, as I've never done it myself.

Apologies for the long explanation, but I hope it helps you in someway.

James, thank you very much for great explanations. We were selecting images without clouds, but as you noted, the cirrus are not so clear to distinguish.

Pozdrawiam,

Michał

Michał, thanks for confirming.

I've only used clouds as one of the examples but you need to also consider other attributes that need masking such as shadows, terrain occlusion etc. before you begin your NDVI analysis. You may want to make use of the Quality Assessment (pixel_qa) bands in order to see the influence of each factor that may distort your interpretation. I've added a screenshot of the 'cloud shadow' mask highlighted in blue color in addition to the cloud masks in dark color.

James, that is right, simply I have little knowledge in this subject, so I need some help.

What do you think about little collaboration in this subject?