Hello, you`re right in your preprocessing workflow. When you do a image fusion you need to compatible your dataset to avoid errors and distortions. I dont understand why do you want to resample your dataset to 6.4m spatial resolution if your hyperspectral image have 3m and your multispectral have 1.6m? Resample only your multispectral image to 3m and you will minimize your spectral losses. Beyond that, you need to verify if the bands are compatible. Are they in the same wavelength? Moreover, dont forget to verify the data acquisition for both images because they may have temporal disagreement. If you apply a resample method use the nearest neighbor (I recommend ENVI for that step) and after that you need to do a coregister between the resampled image and the hyperspectral image. Then, you can finally do a classification process.

See this article for more information about data fusion:

https://www.semanticscholar.org/paper/Registration-Techniques-for-Multisensor-Remotely-S-Fonseca-Manjunath/9c1fad80ee5d2aeeef0def042c4da7f49656e167?tab=abstract

Hope this helps.

Thank you so much for your advices it is so helpful for me i will see more about it. Just one question please what do you mean by " you need to verify if the bands are compatible. Are they in the same wavelength?" how i can do it ? i have Hydice hyperspectral image and World view 2 multispectral image. Thank you

Dear Hanane Teffahi,

Look the link, please.

http://ieeexplore.ieee.org/document/1257267/

https://pdfs.semanticscholar.org/294a/2a4abccb6e4b30e6028ab792d8939278c4c3.pdf

Regards, Shafagat

Dear Hanane, I dont know Hydice hyperspectral image but if you look at Worldview-2 bands there are 9 bands. 8 multispectral and 1 PAN. In the figure you can see the wavelength of each band of wordview-2 sensor. For image fusion, you must verify the Hydice hyperspectral bands before start the data fusion process. You have to select the Hydice bands with the same wavelength of the Worldview-2 bands.

OBS: Thare are many methods to image fusion (I suggest the nearest neighbor method because it's very simple but it may not be suitable for your case). I recommend search for previous studies that already did this kind of data fusion.

Regards, Thales.

Thank you so much.

Best regards.

i will check about wavelength . thanks. I want to do feature extraction and fusion deep learning for image classification.

Dear Hanane,

It is very important that the two images be captured in the same date and both be subject to atmospheric corrections. The High spatial resolution image should cover the same wavelengths as that of hyperspectral. Moreover, some bands in Hydice must be removed if they have high noise rate or they do not have data. Another important issues is that they have to be georeferenced accurately to the same geographic reference.

I suggest that you try both frequency and spatial based data fusion schemes.

I have published many papers which cover these issues good luck.

Best regards

thank you so much Professor i will take account of your advices and explore your papers about this subject.

Best regards.

Interpolation and registration as two pillar preprocessing stages have decisive roles in different remote sensing applications; chapter 4 of following book scrutinizes these issue in the pan-sharpening problem:

https://books.google.com/books/about/Remote_Sensing_Image_Fusion.html?id=LY62NAEACAAJ

Thank you so much.

Dear Hanane

regarding the point raised by Thales Vaz Penha on wavelength, you may have a look at this article

Article Synthesis of Multispectral Images to High Spatial Resolution...

Though it deals with the fusion of PAN and MS images, it contains useful discussion on the physics that may inspire you.

Regarding the resampling, I strongly oppose the use of the nearest neighbor technique. It creates many artifacts. In particular, this will yield bad results in data fusion if you are using fusion techniques based on scale decomposition such as wavelets-based ones. Use instead cubic splines (see e.g. Article Image Fusion - The ARSIS concept and some successful impleme...

Best regards

Lucien Wald

Dear Hanane,

for quality and accurate fusion, it is important that the imagery was taken at the same time or almost the same time frame. Furthermore, it should be emphasised that the fusion process must also be based on similar spectral characteristics of the source band and the one you will use for the fusion. M. Awad has already emphasized the importance of this.

You can find a lot about this topic in the paper below. Five different fusion methods were tested in combination with three different channels used in the fusion (one is a synthesised channel made of a combination of two channels).

Article The effect of fusing Sentinel-2 bands on land-cover classification

Full-text of the paper you can freely download here: http://www.tandfonline.com/eprint/B8naGmIzAt7sDXpbtcVi/full

Complete fusion process can be made in Matlab or R.

I hope this will help you.

Best regards,

Mateo Gašparović

Dear Hanane

obviously, your data are taken in fairly similar spectral bands. I'd like to comment on the answers from M. Awad or Matteo, not to create confusion but to add to your general knowledge and theirs.

Conceptually, data fusion processes may not necessarily apply to data acquired in similar wavelengths or at similar instants. See this paper for definition

Article Some terms of reference in data fusion

I did several works fusing data acquired in very different spectral bands. For example, in this one, we fused Landsat TM6 (around 10 micrometer) at 120 m spatial resolution and another Landsat channel in near infrared at 30 m to obtain a high-resolution TM6. The physical processes behind are completely different. In TM6 are observed the emittances of the objects while TM5 observe the radiation reflected by the objects. Atmospheric effects differ too.

Article Observing air quality over the city of Nantes by means of La...

Finally, in this book

Book Data Fusion. Definitions and Architectures - Fusion of Image...

I have discussed the effect of changes in time acquisition. Obviously, this is a problem only if the properties of the objects of interest have changed between the two instants of acquisition.

Best regards

Lucien

That right , for the time acquisition but generally we can't have different data in the same dates. Some researchs working about deep fusion of hyperspectral and lidar data , the datasets haven't the same dates but we should check in the ground change for example we can fuse a place containing vegetation in the first image and it is urban in the second image. I understand that.

Thank you soooooo much for your help and your sharing i'am so glad.

For the book i can't find the free website for downoald this book and another book named Remote sensing image fusion.

Thank you so much.

Best regards.

Dear Mateo Gasparovic , thank you so much for your advices and document.

Best regards.