diff --git a/src/data.jl b/src/data.jl
index 45c1895..c77499f 100644
--- a/src/data.jl
+++ b/src/data.jl
@@ -129,6 +129,7 @@ mutable struct NCData{T,N #=,TA=#}
     data_full::Array{T,4}
     missingmask::BitArray{3}
     meandata::Array{T,3}
+    mask::BitMatrix
     x::Array{T,5}
     isoutput::Vector{Bool}
     train::Bool
@@ -211,7 +212,9 @@ export sizey
     dd = NCData(lon,lat,time,data_full,missingmask,ndims;
                 train = false,
                 obs_err_std = 1.,
-                jitter_std = 0.05)
+                jitter_std = 0.05,
+                mask = trues(size(data_full)[1:2]),
+)
 
 Return a structure holding the data for training (`train = true`) or testing (`train = false`)
 the neural network. `obs_err_std` is the error standard deviation of the
@@ -231,6 +234,7 @@ function NCData(lon,lat,time,data_full,missingmask,ndims;
                 cycle_periods = (365.25,), # days
                 time_origin = DateTime(1970,1,1),
                 remove_mean = true,
+                mask = trues(size(data_full)[1:2]),
                 direction_obs = nothing,
 #                auxdata = (),
                 )
@@ -302,39 +306,47 @@ function NCData(lon,lat,time,data_full,missingmask,ndims;
 
     N = (is3D ? 4 : 3)
 
-    NCData{Float32,N}(Float32.(lon),Float32.(lat),time,data_full,missingmask,meandata[:,:,:,1],x,
-           isoutput,
-           train,
-           Float32.(obs_err_std),
-           Float32.(jitter_std),
-           lon_scaled,
-           lat_scaled,
-           time_cos,
-           time_sin,
-           ntime_win,
-#           auxdata,
-           direction_obs_,
-                      output_ndims,
-                      ndims,
-                      )
+    NCData{Float32,N}(
+        Float32.(lon),Float32.(lat),
+        time,
+        data_full,
+        missingmask,
+        meandata[:,:,:,1],
+        mask,
+        x,
+        isoutput,
+        train,
+        Float32.(obs_err_std),
+        Float32.(jitter_std),
+        lon_scaled,
+        lat_scaled,
+        time_cos,
+        time_sin,
+        ntime_win,
+        # auxdata,
+        direction_obs_,
+        output_ndims,
+        ndims,
+    )
 end
 
 
 getp(x,sym,default) = (hasproperty(x, sym) ? getproperty(x,sym) : default)
 
 function NCData(data; kwargs...)
-    lon,lat,datatime,data_full,missingmask,mask = DINCAE.load_gridded_nc(data)
+    lon,lat,datatime,data_full,missingmask,mask = load_gridded_nc(data)
 
     default_jitter_std = 0.05
 
     jitter_std = [getp(d,:jitter_std,default_jitter_std) for d in data]
     ndims = [getp(d,:ndims,1) for d in data]
 
-    return DINCAE.NCData(lon,lat,datatime,data_full,missingmask,ndims;
-                         obs_err_std = [d.obs_err_std for d in data],
-                         jitter_std = jitter_std,
-                         isoutput = [d.isoutput for d in data],
-                         kwargs...)
+    return NCData(lon,lat,datatime,data_full,missingmask,ndims;
+                  obs_err_std = [d.obs_err_std for d in data],
+                  jitter_std = jitter_std,
+                  isoutput = [d.isoutput for d in data],
+                  mask = mask,
+                  kwargs...)
 
 end
 
@@ -556,7 +568,10 @@ function getobs!(dd::NCData,data,index::Int)
     return data
 end
 
-function savesample(ds,varnames,xrec,meandata,ii,offset; output_ndims = 1)
+function savesample(ds,varnames,xrec,meandata,ii,offset;
+                    output_ndims = 1,
+                    mask = nothing)
+
     fill_value = -9999.
 
     function accumulate!(var,index,slice,count)
@@ -594,6 +609,7 @@ function savesample(ds,varnames,xrec,meandata,ii,offset; output_ndims = 1)
         end
     end
 
+    # typically the batch size
     nmax = size(xrec,4)
 
     if output_ndims == 1
@@ -610,6 +626,11 @@ function savesample(ds,varnames,xrec,meandata,ii,offset; output_ndims = 1)
             batch_sigma_rec[isnan.(recdata)] .= NaN
 
             for n in 1:nmax
+                if !isnothing(mask)
+                    view(recdata,:,:,n)[.!mask] .= NaN
+                    view(batch_sigma_rec,:,:,n)[.!mask] .= NaN
+                end
+
                 accumulate!(nc_batch_m_rec.var,n+offset,recdata[:,:,n],count)
                 accumulate!(nc_batch_sigma_rec.var,n+offset,batch_sigma_rec[:,:,n],count)
             end
diff --git a/src/model.jl b/src/model.jl
index cee7690..26067a4 100644
--- a/src/model.jl
+++ b/src/model.jl
@@ -328,6 +328,10 @@ at the epochs defined by `save_epochs`.
     application.
 
 
+Internally the time mean is removed (per default) from the data before it is reconstructed.
+The time mean is also added back when the file is saved.
+However, the mean is undefined for for are pixels in the data defined as valid (sea) by the mask which do not have any valid data in the training dataset.
+
 See `DINCAE.load_gridded_nc` for more information about the netCDF file.
 """
 function reconstruct(Atype,data_all,fnames_rec;
@@ -533,12 +537,13 @@ function reconstruct(Atype,data_all,fnames_rec;
 
                     offset = (ii-1)*batch_size
 
-                    DINCAE.savesample(
+                    savesample(
                         ds_,
                         output_varnames,xrec,
                         train_data.meandata[:,:,findall(train_data.isoutput)],
                         ii-1,offset,
                         output_ndims = output_ndims,
+                        mask = train_data.mask,
                     )
                 end
             end