The best answer is from PCI Geomatics itself, so I suggest to write there. For now, I think you should accept that the model used in their software does require at least six control points in order to function properly.

Nevertheless, in my experience more than six GCPs are necessary at a scene level. Their number depends not only on the model, but also on local steepness of the slopes. Related to the last claim, two more problems occur: (i) to find the points measured on the ground also in the image; (ii) to obtain an acceptable error (< 0.5 pixel) at the local level over the entire scene (i.e., the global RMS Error, which is given automatically by any dedicated program, is not sufficient to assess the transformation performance; in addition, RMSE estimations at the sub-scene level are necessary, and for this stage GCPs are needed, which number depends on the slopes steepness; the sub-scenes definition depends on the landscape characterizing each area of interest).

Finally, I consider that the following article may help you with the rationale of GCPs number: "Review article: Geometric processing of remote sensing images: models, algorithms and methods" by T. Toutin, published in 2004 in Int. J. Rem. Sens., 25(10), 1893-1924 (on the web).

The best answer is from PCI Geomatics itself, so I suggest to write there. For now, I think you should accept that the model used in their software does require at least six control points in order to function properly.

Nevertheless, in my experience more than six GCPs are necessary at a scene level. Their number depends not only on the model, but also on local steepness of the slopes. Related to the last claim, two more problems occur: (i) to find the points measured on the ground also in the image; (ii) to obtain an acceptable error (< 0.5 pixel) at the local level over the entire scene (i.e., the global RMS Error, which is given automatically by any dedicated program, is not sufficient to assess the transformation performance; in addition, RMSE estimations at the sub-scene level are necessary, and for this stage GCPs are needed, which number depends on the slopes steepness; the sub-scenes definition depends on the landscape characterizing each area of interest).

Finally, I consider that the following article may help you with the rationale of GCPs number: "Review article: Geometric processing of remote sensing images: models, algorithms and methods" by T. Toutin, published in 2004 in Int. J. Rem. Sens., 25(10), 1893-1924 (on the web).

That probably depends of the model of the orthorectification that PCI is using. P.E a linear rectification model needs at least two points, other models based on polynomial aproximations needs at least 3 or four points. The rule of thumb is the more GPC you have the better is the rectification.

I have to confess I do not know the cartosat DEM product but I suspect that six GCP point for a DEM generation in order to obtain a rigorous orthorectified model will probably not be enough (it will be only good for a tutorial). P.E. In order to obtain a good orthorectified DEM, Japanese ASTER DEM data use hundreds of gcp obtained from platform data.

Related to the triangulation model runs. 6 affords at least two separate triangulation sets or model solutions within the same image. Whereas less than 6 means the triangulation always has to repeat some point.

For a stereomodel, 5 points are necessary for a unique solution. 6 or more points allows a least squares solution, the more the better.

It depends on the orientation model implemented in the software (if you are using PCI with satellite orbital modeling is the Toutin rigorous photogrammetric model) and on the specific type of image. Anyway You need a minimum number of GCPs to determine some unknown parameters (as position and attitude of satellites, but every software use different strategies and consider slightly different unknown parameters). For example If unknown parameters are twelve you need at least four points with their three- dimensional coordinate that give You the possibility to write twelve equations and solve the system.

Obviously if You use the minimum to solve the equation system (like four points in the example) You're not redundant and haven't the possibility to check if there are some errors on Your GCPs. Usually it is bettere that You use minimum 1 and half the minimum of GCPs. Note that the residual reported aren't an indication of accuracy but only of precision. To evaluate the actual accuracy You need independent control points, usually called Check Points (CP). Anyway the subject is more widely explained in the attached paper where a rigorous model implemented by my university is compared with PCI.

Best regards

Dear Dr. Valerio Baiocchi

Thank you very much for explaining it so nicely.

Thanking you very much.

I think u use a two order polynomial model and in this case we need at least six GCPs (each GCP has two equations and we need 12 equations to determine the 12 unknow coefficients of the model!) but its better to use more GCPs with conditions of precision of GCP, density and repartition in order to have a good solution! best regards

6 GCP is the good configuration for orthorectifcation. the RMSE results show the error is not much different after 6 GCP.

It depend on the software being used for georeferencing

Cartosat comes with RPC. You can use RPC 1st order adjustment which requires only a minimum of 3 GCPs distributed evenly on the image

Dear Dr. Cheng

I have Used RPC model as well as satellite Orbital model in PCI Geomatica.

As per the PCI tutorial on "Orthorectification and DEM Extraction of CARTOSAT-1 Imagery" a minimum of six GCPs per scene is required for processing cartosat -1 data using rigorous model.

But my question is why it requires minimum six GCPs.

Thanks

Mritunjay Kumar Singh