Dear Febri,

your job is a difficult one but in my case I would go to different possibilities first I would choose areas where they do not undergo substantial change such as forests, water bodies...etc. In this case I can use the ground survey of 2014 on the previously selected areas to verify part of 2002 and 1994 accuracy. Another possibility is Google earth. You may find some high resolution images of 2002.

In general if the accuracies of the classification of both 2014 and 2002 are high there is a large possibility that 1994 accuracy is high too.

Again I can say that verified classes created from 2014 image can be used for other images similar to 2014 landsat image. However, images for 2002 and especially 1994 can be different from 2014 landsat image (8 or 7 or 5).

Dear Febri,

You may also find this paper useful: Fortier J, Rogan J, Woodcock C, Runfola DM (2011) Utilizing temporally invariant calibration sites to classify multiple dates of satellite imagery. Photogrammetric Engineering & Remote Sensing 77 ((2)) 181–189.

Best regards.

Hi bro Febri,

your question is kind of a common problem for those who want to do land use change analysis. In this case, I will say that you are not necessarily performing accuracy assessment for your 2002 and 1994 if you do not have any supporting data for ground validation. You may use your validated land use/cover map from 2014 as your data ancillary for retrieving land use/cover from prior dates with an assumption that there are no additional in land use/cover class in the prior years (i.e. changes are expected for only changes in the area extent and land use/cover class reduction). As a result, your land use/cover maps for 1994 and 2002 will be dependent to your 2014 map. Well I guess this is the best solution for your case.

Good luck!        

Agree with all answers that I read. Hero Marhaento answer is reasonable. My suggestion is that you need to carry out accuracy assessment for the most recent date (2014). Calculate accuracy for each class and its percentage, transfer results to weighted accuracy. Compare changes among dates and adjust accuracy to percentage land use/cover, i.e. expected accuracy according to dominant classes at each date.

Good luck

Dear Febri

It does not matter for the year if it is the same sensor (TM or ETM+  30m)

To assess the classifications you can use the confusion matrix.

This must be greater than at least 70% that the result to be reliable

There needs to be a good samples (rois tools) by selecting pixels from the ZOOM window.

After, you have to test the global index and precision index (0-1). If the result approaches 1 kindly note you consider that your classification is good

Good luck

Dear Febri,

Your question is an interesting one and it is a very important aspect of land use science as the maps produced from Imageries are frequently assumed to be good enough as a basis for decision-making on land use issues. However, one point you have raised is the fact that you are not able to go back in time to collect the data necessary for validation. I agree with Jawad's point about using the current data. But there is a lot of argument about this among practitioners.

Please look at the following, though we used the Intensity Analysis model in this work.

Good luck

Article Influence of classification errors on Intensity Analysis of ...

Dear all,

Thank you very much for all valuable responses and also for the papers. I will try to understand each suggestion and decide which to apply in my research. Good luck for all your works.

Best regards

Febri

Dear Febri,

There are several ways to solve your problem,

1) If you find images better spatial resolution then landsat from your study area like ikonos spot aerial photos or google earth, just classify using eCognition classifier (definies software you can find the applying object based classification it is providing you to fix the your classification object by object so that accuracy is increased. You may use these higher resolution classification results as reference.

2) You are finally classifing the same area and classify the 2014 images using several classification methods and test it to find the best techniques. Apply best method to others. It may acceptible for the same area which have been had the same characteristics.

3) Try to find the old maps in archives. Generally, countries are produced topographic maps (pysical maps in 1/25000 scale) you can use old maps for accuracy assessments.

I hope it would be helpful.

Dear Febri,

Following paper may help you to certain extent. This is a very common method in practice. 

http://dx.doi.org/10.4236/jgis.2012.43024

I think Onur's suggestions, signed 1 and 3 are very good suggestions, though, as for my experience , the same techniques can't be applied to  TM ETM+ OLI sensors data . They need to be adjusted for certain image type. So you certainly need any information for 2002and 1994. You can use ETM+ images for 2014 and 2002 but....it also can't be completely comparable, you know. So you should be careful in your work. 

Best regards,

Maria

Thanks for Lars answer, I would like to add an idea. As you know Gain/Offset of multi-temporal data is different, so you could also try to rescale any  of old datasets from digital number to radiance firstly and  then to assess correlation between results of classification of them and new GT map.

 This paper is about the upadate of Land use database, they assess the accuracy of the dynamic spot through re-visual interpretation of other specilist for the sampling unit.

A 2010 update of National Land Use/Cover Database of China at 1:100000

scale using medium spatial resolution satellite images

Zengxiang Zhang, XiaoWang ⁎, Xiaoli Zhao, Bin Liu, Lin Yi, Lijun Zuo, QingkeWen, Fang Liu, Jinyong Xu, Shunguang Hu

 You could use Google earth Historical data if available for the study area during the same time or near by. You can also use some high resolution images from other satellites or already generated authentic maps. you can also use any other sot of work to validate your data.

Dear everyone, I say thank you very much for all the attentions. I have received a lot valuable information from you all.

By the way, I have just also found a paper dealing with the accuracy of temporal image classification by assessing the rationality of changes. I think it also can enrich choices dealing with such topic. 

Best Regards

Febri

Article Accuracy analysis of remote sensing change detection by rule...

Very interesting. I also have a similar change detection hurdle to cross. In my own case, i am trying to assess changes using 2002 and 2014 Landsat images with 1990 land use map as the base map. Any guidance or better ideas please?

I used different evaluation approaches in the following publications:

Article Automatic building extraction in dense urban areas through G...

Conference Paper AN AUTOMATIC REGION-BASED SEGMENTATION ALGORITHM FOR BUILDIN...

You can validate the classification results by Kappa coefficient or RMSE error