I have three different satellite Images from three different time periods, one from Landsat 8 (30 m), another from ASTER (15m) and one from MODIS (250 m). The objective here is to do a multi-temporal land cover change detection analysis study with the help of Digital Image Classification. I plan to first do the Image Classification for the three images separately and then resample the coarser spatial resolution images to the geometry of the finer spatial resolution, i.e., MODIS (250 m) to 15m pixel size and Landsat (30m) to 15m pixel size respectively for the sake of comparison. Finally after resampling with a Nearest Neighbor Interpolation method I would go ahead with the change detection process.
I am curious to learn whether this is the right approach or are there better means of achieving the desired result?
Dear James,
check the following paper.
http://research.ijcaonline.org/volume54/number7/pxc3882317.pdf
http://research.ijcaonline.org/volume54/number7/pxc3882317.pdf
Dear James,
Image Registration is going to play a very important role in this in order to detect accurate change otherwise you will get garbage value.
Check the following paper
http://www.sciencedirect.com/science/article/pii/S1319157813000268
In your approach image registration doesn't play the most important role, because you are comparing the classifications at a rather coarse resolution, sorry Prashant that i disagree with you. I think this registration problem is related to pixel based change detections. One thing that makes me concerned about your approach is the upsampling from 250 modis pixel to 15 or 30 m aster and landsat pixel. I wouldn't do that. If the modis sensor tells you there is a certain class on the 250 m pixel, what do you do with all the landsat pixels (approx.70 Landsat Pixels!) that tell you it is not the same class? It's certainly not necessarily a change over time but rather a scale problem. Hope this somehow help you? If not, fell free to answer.
If quantification of the changes in the area of each class is important, a critical aspect of your study might be to evaluate how the magnitude of change compares to classification errors, something that is generally difficult to do.
In the simplest cases, when change can be equated to the monotonous loss/increase of just one class (e.g. in some studies of deforestation), you can assess the adequacy of your method by mapping your class for your base year, and then classifying only the remaining areas for each consecutive image in your time-series. Your method would be considered adequate if the classification error (which you would roughly assess by producing independent classifications, for example) is significantly smaller than the magnitude of change you have estimated.
If quantifying the area is important for your research, you can apply this principle, for example, to evaluate whether moderate and coarse resolution imagery can be actually combined or not.
Hello James-
Sounds interesting. Terrific era that we are in that we have access to so much data!
Post-classification change detection is not recommended. The error of any given realization (map) is carried through the change detection. Very high classification accuracies are required for each map; plus, the scale differences you would bring in from the proposed resampling would likely not lead to useful results. Please see excellent Fuller et al. paper below.
Spectral based classification techniques are more stable and offer you more options. There are literally hundreds of approaches. Typically you should select your approach based upon the information needs that you have for your outcome maps. Ideally, your information needs can be supported by Landsat, then you would have the ability to use calibrated, free, analysis ready, data that can cover both large areas and time series.
Article The characterisation and measurement of land cover change th...
I am overwhelmed by the different insightful answers I've received. Dr. Michael, Dr. Diogenes, Christoph and Prashant thank you for your reply. By and large what I could deduct from the individual suggestions and feedback is that the methodology really depends on the Application or the purpose of what I want to map and how I want to visualize and present my results, either quantitatively or qualitatively.
- As rightly pointed out by Christoph it wouldn't make sense to resample 250m into 15m given the fact that there might be some land cover class within the dimensions of 250*250 pixel sizes. The best solution which I could think of is to all together remove MODIS data from the analysis as Landsat and ASTER which represents the range of time periods for the change detection study will do the job just fine. It wouldn't matter using MODIS so much because it is of an in-between time period. Thanks Christoph for your plausible remark. I would be surprised if someone provides a better justification to keep MODIS data still in the analysis as it seems almost certain that the results would be compromised if I keep it. It's really a challenge to Integrate 250m and 15m.
- Dr. Diogenes, you have provided an experienced and judicious solution. I will be mapping vegetation mainly and therefore your formula of 'Classification Errors < Magnitude of Change' fits sufficiently well here. I will be validating my results against this standard. Thanks a ton.
- Dr. Michael, you have given a brilliant idea supported with an equally exemplary paper. No wonder you are in NRCan. It's a dream Organisation to work in. Although I am not sufficiently competent and trained enough in 'Spectral based classification' techniques yet, I'll make it a point to learn and apply one of the approaches in my case study with some expert help. I was wondering whether I can use Post-classification change detection method provided I have high classification accuracies, if not I don't know where else I can use this technique. What is the range of accuracy you would consider to call it a high accuracy? This seems so subjective isn't it? From this aspect spectral based technique seems more reliable and objective. I've sent a 'full-text request' to access the paper by Fuller et al.
- Prashant, the geo-registration part has been taken care of.
All your answers have added a totally new perspective to my study and has compelled me to reconsider the methodology. Thanks for rekindling the thought process.
Hi James,
Why don't you start with the most cited (?) paper in remote sensing: http://www.tandfonline.com/doi/abs/10.1080/01431168908903939#.VH460THF98E
Followed by http://www.tandfonline.com/doi/abs/10.1080/014311699213659#.VH47ezHF98E
http://www.tandfonline.com/doi/abs/10.1080/02757259609532305#.VH462jHF98E
http://www.tandfonline.com/doi/abs/10.1080/0143116031000139863#.VH49MTHF98E
Also there is an excellent chapter in Jensen's DIP book. All these will give you the basic idea of available change detection techniques.
Since you will be using multi-source/temporal data, along with geometric integrity, you should consider the atmospheric effects: http://www.sciencedirect.com/science/article/pii/S0034425700001693.
Now you can follow (reproduce) this work: http://www.tandfonline.com/doi/abs/10.1080/01431160801950162#.VH49ETHF98E
After you have some expertise, you can take the problem in sub-pixel or unmixing domain to deal with the different spatial resolutions. I couldn't find any particular work but I will be happy discuss on this!
Enjoy!
-Ani
Hi Aniruddha Sir. I am very delighted to hear from you after a long time. Thank you very much for providing those links to various papers. Very useful for my study. You are right I should go back to the basics with DIP. I remember the PCA class you had taken once.
Dear James
I really don't understand your problem. How long is your time series?, why don't use the same platform instead three with different spatial and radiometric resolution?. Landsat 8 hardly has a whole operating year, but the whole series, available in internet is larger than 30 years.
Even so, some answers to your question have been very interesting and mine is just to know the reasons to compare three different satellites and probably contibute then with some advice
No doubt, you have received impressive answers to your question. However, I think there are still missing gaps. That is in terms of identifying the types of changes that have taken place. Briassouli (2000), identifies 3 types of land use/ land cover changes-Coversion, Modification and Intensification. In your case since you are working on vegetation you may also have to consider the fragmentation pattern of the identified land cover.. I think handling the change detection like this will make it robust.
At the spatial resolution of Landsat/Aster please see attached articles. Beyond change detection, both include a forecast based on spatial modelling. I miss a description of your aim beyond the change detection. For my targeted change processes the once in decade image in the same season was sufficient.
To combine fine spatial resolution imagery with spatial mid-resolution, high temporal MODIS will generally not be a good is idea as Christoph has already explained for technical reasons. Why would you do so anyhow? If you want to follow phenomena with high temporal resolution (e.g. fire) MODIS is great (see attached). For vegetation phenology hyper-temporal SPOT would seem the most suitable.
Article Forecasting the pattern and pace of Fagus forest expansion i...
Article What Drives Conversion of Tropical Forest in Carrasco Provin...
Article From monospecific to mixed forest after fire?: An early fore...
Thanks Diana and Dr. Gils for your answers & for the papers.
I did not have a specific aim beyond change detection (general vegetation variation) when I framed the question initially. I was just interested in 'how to do it' question rather than 'why to do it' but after getting all these exceptional answers, I have thought to do a thorough reflection on the aims and objectives. Just a self learning exercise. That's all.
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling