Observed predictions

[rondon-d]

Hello,

Is there an easier way to estimate the predictive mean value at the observed sites for each point than doing interpolation?
I try to look in the predicted outcome for the exact coordinates of the data, but there no exact match.

Thank you for your time!

[PhilipMostert]

Hi,

In the predict function, you should be able to add whatever coordinate data across the study area. So if you add the data for the observed sites there, you should get predictions at those locations.

[rondon-d]

hello,
thank you for the quick reply.
so based on the example, it goes something like this:

´´´
predictions <- predict( species$eBird$geometry, #the points of the ebirds observations
modelRun, # the mode, outcome of the fitISDM
spatial = TRUE,
fun = 'linear'
)
´´´

Thank you

[PhilipMostert]

That looks correct, although maybe change the order or the data and the model.

[rondon-d]

I am trying this:

predictions <- predict( object= modelRun, # the mode, outcome of the fitISDM
                        data= species$eBird$geometry, #the points of the ebirds observations
                        spatial = TRUE,
                        fun = 'linear'
                        )

And getting the following error:
Error in NextMethod() : incorrect number of subscripts on matrix

Also by making the input of data as a matrix:

predictions <- predict( object= modelRun, # the mode, outcome of the fitISDM
                         data= st_coordinates(species$eBird$geometry), #species$eBird$geometry, #the points of the ebirds observations
                        spatial = TRUE,
                         fun = 'linear'
                         )

with the following error
Error in [<-(*tmp*, , spatCov, value = inlabru::eval_spatial(where = data, :
subscript out of bounds

Am I missing something

[PhilipMostert]

Does it work if you remove the $geometry and just predict on species$eBird?

[rondon-d]

like a chram!
here is the final code:

predictions <- predict( object= modelRun, # the mode, outcome of the fitISDM
                        data= species$eBird, #the points of the ebirds observations
                        spatial = TRUE,
                        fun = 'linear'
)

Thank you!

[rondon-d]

Hello
sorry to re open this issue.
I am trying to do the predictions for a spatio-temporal model with some covariates. The code and error is as follows:

pred_Obs <- predict(object = mod_Cov, # the mode, outcome of the fitISDM
                    data = Data_Obs, #the points of the ebirds observations
                    spatial = TRUE,
                    fun = 'linear'
                    )
Error in names(x) : 
  no slot of name "pntr" for this object of class "SpatRaster"

the class of the model are as follows:

class(mod_Cov)
[1] "modISDM" "bru"     "iinla"   "inla"   

class(Data_Obs)
[1] "sf"         "data.frame"

any clues of how to make it work?

Thank you for your time :)

[PhilipMostert]

Hi,

I'm not sure I can see the issue here. Does Data_Obs have the covariate information attached already? If not, could you try that before running predict and see if that works?

[rondon-d]

Hello,
thank you for the quick answer.

Data_Obs has date, year and location of the observation. No information on the covariates.

head(Data_Obs)
Simple feature collection with 6 features and 2 fields
Geometry type: POINT
Dimension:     XY
Bounding box:  xmin: 5021.358 ymin: 4568.844 xmax: 5145.072 ymax: 5050.525
Projected CRS: PROJCRS["unknown",
    BASEGEOGCRS["unknown",
        DATUM["Unknown based on GRS 1980 ellipsoid",
            ELLIPSOID["GRS 1980",6378137,298.257222101,     LENGTHUNIT["metre",1],                 ID["EPSG",7019]]],
        PRIMEM["Greenwich",0,             ANGLEUNIT["degree",0.0174532925199433],
            ID["EPSG",8901]]],     CONVERSION["unknown",    METHOD["Lambert Azimuthal Equal Area",
            ID["EPSG",9820]],         PARAMETER["Latitude of natural origin",52,
            ANGLEUNIT["degree",0.0174532925199433], ID["EPSG",8801]],         PARAMETER["Longitude of natural origin",10,
            ANGLEUNIT["degree",0.0174532925199433], ID["EPSG",8802]],         PARAMETER["False easting",4321,
            LENGTHUNIT["kilometre",1000],  ID["EPSG",8806]],         PARAMETER["False northing",3210,
            LENGTHUNIT["kilometre",1000], ID["EPSG",8807]]],     CS[Cartesian,2],
        AXIS["(E)",east,             ORDER[1],  LENGTHUNIT["kilometre",1000, ID["EPSG",9036]]],
        AXIS["(N)",north,             ORDER[2],    LENGTHUNIT["kilometre",1000,
                ID["EPSG",9036]]]]

       date Year                  geometry
1 2019-05-22 2019  POINT (5119.24 4537.01)
2 2018-12-27 2018 POINT (5111.33 4724.35)
3 2019-03-28 2019 POINT (5121.54 4246.87)
4 2019-04-30 2019 POINT (5011.38 5550.55)
5 2020-12-28 2020 POINT (5115.02 4265.75)
6 2021-04-15 2021 POINT (5061.76 4468.84)

To include the covariate information I should add a new column to each observation with the respective value of the covariate? When I did the prediction with the example code modelRun did not have covariate information, nor the predict() function seems to need it. Why is it required now?

Thank you again.

[PhilipMostert]

Yes, just something like terra::extract to get the covariate information at each point, and I assume it's a covariate issue since the error has something to do with terra. What's the traceback() for this error? Or how does this model differ from the other one, did the other one not have a temporal aspect?

[rondon-d]

I will try including the covariates.
yes, the main difference is that now (the model that is giving the error) there is a temporal aspect.
Here is the traceback of the error:

Error in names(x) : 
  no slot of name "pntr" for this object of class "SpatRaster"

Enter a frame number, or 0 to exit   

 1: predict(object = mod_Cov, data = Data_Obs, spatial = TRUE, fun = "linear")
 2: predict(object = mod_Cov, data = Data_Obs, spatial = TRUE, fun = "linear")
 3: predict.modISDM(object = mod_Cov, data = Data_Obs, spatial = TRUE, fun = "linear")
 4: inlabru::eval_spatial(where = data, data = get("spatialcovariates", envir = object
 5: get("spatialcovariates", envir = object$spatCov$env)[spatCov]
 6: get("spatialcovariates", envir = object$spatCov$env)[spatCov]
 7: .local(x, i, j, ..., drop = drop)
 8: grep(i, names(x))
 9: is.factor(x)
10: names(x)
11: names(x)

If i start the debugging process in action 3 I get a similar error from an inlabru function:

 data[, spatCov] <- inlabru::eval_spatial(where = data, 
        data = get("spatialcovariates", envir = object$spatCov$env)[spatCov], 
        layer = spatCov)

Error during wrapup: no slot of name "pntr" for this object of class "SpatRaster"
Error: no more error handlers available (recursive errors?); invoking 'abort' restart

In this case the used object are as follows: data is the same as Data_Obs, while object$spatCov$env is <environment: 0x00000215bead3460>

hopefully this is useful.

[PhilipMostert]

Adding covariates yourself will probably fix it, but strange that it's not working. Are your covariates in list form, I think that would cause an issue?

[rondon-d]

Okay, I think I figure it out.
I was trying to use the predict() function on an environment that has not run the fitISDM() function. I was uploading the object from a previous run session.
I think I need to run all the model and then use the predict function. I am working on this right now. hopefully it will work.

:)

[rondon-d]

Hello,
I try two approaches, both are giving me errors, those are:

1. including the covariates infomration via terra::extract()
2. Run the model, create the environment and run the prediction. In this case covariates are a object of class SpatRaster . This second one is giving me the following error:

 predict(object=mod_Cov, 
         data= Data_Obs, 
         spatial = TRUE, 
         fun="linear"
         )

Error in Matrix::sparseMatrix(i = iii, j = jjj, x = xxx, dims = c(n, ncol(M2) *  : 
  invalid 'dims'
In addition: Warning messages:
1: In fmesher::fm_cprod(data, data.frame(time_data)) :
  Block information mismatch resulting in NA weights, and 'na.rm' was not supplied. These rows will be removed.
2: In max(repl) : no non-missing arguments to max; returning -Inf

Is there something that I am missing, or just making the prediction when years is a variable is not that straight forward.

Thank you in advance!