diff --git a/src/xray/conventions.py b/src/xray/conventions.py
index 773bd5bf188..e7df7b5785c 100644
--- a/src/xray/conventions.py
+++ b/src/xray/conventions.py
@@ -247,13 +247,13 @@ def encode_cf_variable(array):
     dimensions = array.dimensions
     data = array.data
     attributes = array.attributes.copy()
-    encoding = array.encoding
+    encoding = array.encoding.copy()
 
-    if isinstance(data, pd.DatetimeIndex):
+    if np.issubdtype(data.dtype, np.datetime64):
         # DatetimeIndex objects need to be encoded into numeric arrays
         (data, units, calendar) = utils.datetimeindex2num(data,
-                                          units=encoding.get('units', None),
-                                          calendar=encoding.get('calendar', None))
+                                          units=encoding.pop('units', None),
+                                          calendar=encoding.pop('calendar', None))
         attributes['units'] = units
         attributes['calendar'] = calendar
     elif data.dtype == np.dtype('O'):
@@ -327,7 +327,10 @@ def pop_to(source, dest, k):
     if 'dtype' in encoding:
         if var.data.dtype != encoding['dtype']:
             raise ValueError("Refused to overwrite dtype")
-    encoding['dtype'] = data.dtype
+    if not isinstance(data, pd.Index):
+        # When data is a pandas Index we assume the dtype will be
+        # inferred during encode_cf_variable.
+        encoding['dtype'] = data.dtype
     if np.issubdtype(data.dtype, (str, unicode)):
         # TODO: add some sort of check instead of just assuming that the last
         # dimension on a character array is always the string dimension