Hello.

You must remove the shadows, since all remote sensing programs confuse as vegetation.

Hi Elijah, 

check the date of the Landsat imagery, the vegetation does not have the same NDVI level in autumn-winter than the spring-summer season, maybe one image is from winter and the another from summer.

You can also check with Google Earth if there has been an increment of forest areas around which have high NDVI values. Another possibility is a fire in the area during 2002 and now the vegetation has been recovered in 2014. 

Regards 

Raquel 

Hi

If you calculated average NDVI values per land cover type, don't your results simply mean that the areas that were greener in 2002 are those that remain in place in 2014? I.e. there was a loss in vegetated area, but the remaining vegetated patches are those that were already greener in 2002?

Hi

Double check season when images were taken.

Have you conducted any accuracy assessment?

Dear Elijah

Take all of the above responses into consideration. Troubleshoot by Inspecting your NDVI images in detail to see where the differences/anomalies are arising.

Also, could the differences be due to Landsat sensor differences (L5 2002 compared to L8 2014)?

If still not satisfied, try another index for comparison.

If using Landsat 5 (2002) & Landsat 7 (2014), you could try PPSG index, http://www.sciencedirect.com/science/article/pii/S092427160900104X

Although it is not a well known index, it is fairly insensitive to shadow and does not saturate at high levels of cover. PPSG differences due to L5 (TM) & L7 (ETM+) sensor differences are expected to be small. However PPSG has not been redesigned for Landsat 8.

in a simpliest way be sure that you examime images from the same sensor (Landsat 7 ETM+ imagery only cover all the period you refer). And don't forget of High and low gain ETM+ regimes.

In other cases you cannot compare NDVI values. 

If you need to compare the images from different sensors (Landsat 8 OLI and Landsat 5 TM I guess), you should apply special methods that allow to adjust reflectancies in images obtained by different sensors, for instance, relative radiometric normalization, or Universal Pattern Decomposition method (VUIPD based on UPDM), FAPAR etc.

Good luck,

Maria

http://fapar.jrc.ec.europa.eu/Home.php

Article A new vegetation index based on the universal pattern decomp...

NDVI and vegetation cover related each other but some of papers showed that it is not maching perfectly esspecially in unstable vegetation areas. I strongly suggest you soft classification techniques to map exact cover of the vegetation there are benefit clues in fallowing articles:

Article Mapping percentage tree cover from Envisat MERIS data using ...

Article A Simplified Method For The Determination And Monitoring Of ...

Conference Paper ESTIMATING URBAN GREEN COVER USING FUZZY (SOFT) LAND USE CLA...

A lot of very good suggestions and comments already made and I think that one, or several, of the issues mentioned is the source of your confusion.  Atmospheric and sun elevation corrections, time of year, wet or dry year / season, greened up at one time and not on the other date, and sensor differences (especially 8-bit vs 11-bit between Landsat 5 and Landsat 8) could all be contributing to the differences.  However, as mentioned you can not go by the AVERAGE of the NDVI to determine if there is more or less vegetation cover; this is more of a measure of 'HOW GREEN' the vegetation happens to be at the time of imaging.  My suggestion is to generate an image where a pixel is set to yes vegetated OR not vegetated and then count up the number of pixels in each land cover type.  This will give you the amount of AREA vegetated for each year; then you can use that to decide if there is more vegetation cover or not.

Good Luck,

Pat

Hi Elijah. NDVI values change not only between seasons but also during the whole growing season. It depends of the type of vegetation cover and its phenological phase. The 2 weeks difference between imaging could bring different values. Consider also hot and dry periods during summer. Veronika

Hi, Elijah. Just like Fulgencio, I think that we need to know if you have applied any kind of pre-processing routines to those images before applying NDVI calculations.

Have you applied any conversion of data to reflectance? Or did you calculate NDVI on raw Landsat data instead?

If your calculations were applied over raw data, that might be a problem related to the 8-bit vs. 11-bit issue, as Pat pointed out in his answer above.

Another motive to that difference, if you used raw data, is that Landsat-5 raw data are related to radiance values, while Landsat-8 raw data are related to top-of-atmosphere reflectance.

Then, I think that converting your data to reflectance values with atmospheric correction may help you to solve this problem, if it is related to illumination issues.

Converting your data to reflectance values WITHOUT atmospheric correction might also help you, but then you should take into consideration that your analyses can have errors related to backscattering effects of the radiation measured by the satellite and that these errors can affect your NDVI as well.

And as many colleagues have already pointed out, I think you should take into consideration the phenology of vegetation in your analysis. The images in both years must be taken at the same phenological period of vegetation (greenup, maturity, senescence or dormancy periods). Take a look at the figure 3 of this paper:

http://dx.doi.org/10.1016/S0034-4257(02)00135-9

For South America, we have INPE Series View, with phenological information derived from MODIS data to the whole continent. Maybe you should try to find a similar database to your study area in order to help you to choose the best periods of analysis. INPE Series View site:

https://www.dsr.inpe.br/laf/series/

Good Luck!

This may happen if images acquired in different months. For example if you acquire an image of rainy season of a place then the area is more greener and even open areas(soil) also show vegetation cover because of growth of small grasses. Similarly during summer all these small grasses and shrub becomes dry and thus expose the soil which leads to major signature differences. I would suggests to have images of near about same month for different years for better assessment. 

If above is not the case then go for atmospheric and radiometric correction before finding NDVI values. 

You can follow one more approach. go for supervised classification where you select some training areas related to vegetation which are having different signatures like dry grass, greengrass, forest canopy etc and then trained them using any classifier like SVM or Guassian NB. After classifying the image just make the image binary with two classess i.e. club all vegetation calsses to one and other class as non vegetation. 

Hope this helps.

It is really not easy comparing two or more images, even though they are of the same sensor.  Dry or wet conditions of the cover will influence much to the results due to the spectral characteristic of NIR band to water content. Maybe you can try using Blue-Green combination to calculate NDVI, instead of Red-NIR combination. But Blue and Green are sensitive to atmospheric condition, where atmospheric and radiometric correction are necessary. 

Thank you all for your insightful contributions. I was able to resolve the issues taken into account the atmospheric correction of some of the images. 

https://www.researchgate.net/publication/309589918_Driving_force_for_forest_fragmentation_explored_by_land_use_change_in_Song_watershed_India

Article Driving force for forest fragmentation explored by land use ...

For vegetation, we may understand this but in my computation of NDVI (both monthly and seasonal), the NDVI values of urban pixels has been increased in 2016 and 2019 Landsat data against 2002, even though the urban extent has increased with no question. I have done data fusion following the ESTARFM technique. I did the prediction of all maps with 8-day modis composite datasets. I checked with Aster and Sentinel data-sets too for individual day imageries, I have found a relatively similar trend. Do Landsat5 and Landsat8 does make these differences?