pikobs 3.0.415__tar.gz → 3.0.417__tar.gz

{pikobs-3.0.415 → pikobs-3.0.417}/CHANGELOG.md

@@ -1487,3 +1487,11 @@ Please use the following tags when editing this file:
 ## [3.0.415] 2026-09-17
 ### Added
 - bugs
+
+## [3.0.416] 2026-09-17
+### Added
+- bugs
+
+## [3.0.417] 2026-09-17
+### Added
+- bugs

{pikobs-3.0.415/pikobs.egg-info → pikobs-3.0.417}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: pikobs
-Version: 3.0.415
+Version: 3.0.417
 Summary: pikobs
 Home-page: https://gitlab.science.gc.ca/dlo001/pikobs
 Author: David Lobon

pikobs-3.0.417/VERSION

@@ -0,0 +1 @@
+3.0.417

{pikobs-3.0.415 → pikobs-3.0.417}/pikobs/scatter/scatter.py

@@ -143,7 +143,7 @@ tokens need it, and the list of stations that contributed is saved
 separately so the titles can name them.
 
 **Matching** (comparison runs only). For ``omp``, ``oma``, ``stdomp``,
-``stdoma`` and ``bcorr`` the two runs are compared on the *same*
+and ``stdoma`` the two runs are compared on the *same*
 observations. Each 6-h file of the control is joined with the same file of
 the experience on
 
@@ -307,11 +307,11 @@ comment on ``FONCTION`` gives the test used for each function:
    FLAGS_CRITERIA=(all assimilee)      # all, assimilee, rejets, bgckalt, ...
 
    # Functions. Values are experience minus control; the test is in brackets.
-   # omp oma stdomp stdoma bcorr use the observations common to both runs.
+   # omp oma stdomp stdoma use the observations common to both runs.
    #   omp oma        |bias exp| - |bias ctl|, units   [paired t + F-test]
    #   stdomp stdoma  100 (σ exp - σ ctl) / σ ctl      [F-test]
    #                  for these four, negative (red) = experience better
-   #   bcorr          exp - control, units             [paired t]
+   #   bcorr          change of |bias corr.|, %        [no test]
    #                  (radiance families only, skipped when the files have
    #                  no BIAS_CORR column)
    #   obs            exp - control, units             [no test]
@@ -578,19 +578,19 @@ F-test are the ones of ``saska40_v0_hs.py``, the channel statistics used in
 the IR studies, so a map and a channel plot of the same experiment tell the
 same story.
 
-+------------------------+------------------------------------+----------------------+--------------+----------------+
-| Function               | Box value (experience - control)   | Red means            | Observations | Test           |
-+========================+====================================+======================+==============+================+
-| ``omp``, ``oma``       | ``|bias_exp| - |bias_ctl|``, units | negative: exp better | common       | paired t and F |
-+------------------------+------------------------------------+----------------------+--------------+----------------+
-| ``stdomp``, ``stdoma`` | ``100 (σ_exp - σ_ctl) / σ_ctl``    | negative: exp better | common       | F              |
-+------------------------+------------------------------------+----------------------+--------------+----------------+
-| ``bcorr``              | ``bcorr_exp - bcorr_ctl``, units   | positive: larger     | common       | paired t       |
-+------------------------+------------------------------------+----------------------+--------------+----------------+
-| ``obs``                | ``obs_exp - obs_ctl``, units       | positive: higher     | all          | none           |
-+------------------------+------------------------------------+----------------------+--------------+----------------+
-| ``nobs``, ``dens``     | ``100 (N_exp - N_ctl) / N_ctl``    | positive: more obs   | all          | none           |
-+------------------------+------------------------------------+----------------------+--------------+----------------+
++------------------------+------------------------------------------+----------------------+--------------+----------------+
+| Function               | Box value (experience - control)         | Red means            | Observations | Test           |
++========================+==========================================+======================+==============+================+
+| ``omp``, ``oma``       | ``|bias_exp| - |bias_ctl|``, units       | negative: exp better | common       | paired t and F |
++------------------------+------------------------------------------+----------------------+--------------+----------------+
+| ``stdomp``, ``stdoma`` | ``100 (σ_exp - σ_ctl) / σ_ctl``          | negative: exp better | common       | F              |
++------------------------+------------------------------------------+----------------------+--------------+----------------+
+| ``bcorr``              | ``100 (|bc_exp| - |bc_ctl|) / |bc_ctl|`` | positive: larger     | all          | none           |
++------------------------+------------------------------------------+----------------------+--------------+----------------+
+| ``obs``                | ``obs_exp - obs_ctl``, units             | positive: higher     | all          | none           |
++------------------------+------------------------------------------+----------------------+--------------+----------------+
+| ``nobs``, ``dens``     | ``100 (N_exp - N_ctl) / N_ctl``          | positive: more obs   | all          | none           |
++------------------------+------------------------------------------+----------------------+--------------+----------------+
 
 For the departures the question is whether the experience gets closer to
 zero, so the bias is compared in absolute value:
@@ -614,6 +614,15 @@ observations, a higher observed value or a larger correction are red.
 .. math::
 
    \Delta_{N} = 100\,\frac{N_{\mathrm{exp}} - N_{\mathrm{ctl}}}{N_{\mathrm{ctl}}}
+   \qquad
+   \Delta_{\mathrm{bcorr}} = 100\,\frac{\left|b_{\mathrm{exp}}\right|
+                                    - \left|b_{\mathrm{ctl}}\right|}
+                                   {\left|b_{\mathrm{ctl}}\right|}
+
+The bias correction :math:`b` can be negative, so its size is compared. In
+regions where the control barely corrects anything the percentage gets
+large and saturates at ±100 %; read those boxes with the single-run
+``bcorr`` map at hand.
 
 Why not a percentage for the bias? It would divide by a control bias that
 is often close to zero: a box going from 0.01 K to 0.03 K is +200 % and
@@ -627,7 +636,7 @@ the last step a box takes the darkest colour.
 Three maps per comparison
 -------------------------
 
-For ``omp``, ``oma``, ``stdomp``, ``stdoma`` and ``bcorr`` every comparison
+For ``omp``, ``oma``, ``stdomp`` and ``stdoma`` every comparison
 produces three maps, switched with the *Difference / Significance*
 selector of the viewer:
 
@@ -641,7 +650,8 @@ selector of the viewer:
 | ``significant only`` | ``..._sigonly.png``     | difference map, boxes above 95 % only |
 +----------------------+-------------------------+---------------------------------------+
 
-``obs``, ``nobs`` and ``dens`` only have the difference map.
+``obs``, ``bcorr``, ``nobs`` and ``dens`` only have the difference map,
+computed on all the observations of each run.
 
 **difference** shows every box:
 
@@ -652,7 +662,10 @@ selector of the viewer:
 
 **significance** shows how sure the test is, from 0 to 100 %. Grey boxes
 stay at or below 95 % (darker when closer), orange ones are above 95 % and
-red ones above 99 %; the thick line on the colour bar marks 95 %. The box at
+red ones above 99 %; the thick line on the colour bar marks 95 %. For
+``omp`` and ``oma``, a crossed box (``xxx``) passed both tests but its bias
+and spread moved in opposite directions, so it is not counted (see *Why omp
+and oma need both tests*). The box at
 the top right recalls which test was run and counts the significant boxes:
 
 .. image:: _static/scatter_cmp_omp_significance.png
@@ -661,7 +674,8 @@ the top right recalls which test was run and counts the significant boxes:
    :width: 100%
 
 **significant only** is the difference map with the same colours and the
-same scale, where only the boxes above 95 % are drawn. It is usually the
+same scale, where only the boxes that pass are drawn (for ``omp`` / ``oma``:
+both tests, same direction). It is usually the
 one to show in a meeting:
 
 .. image:: _static/scatter_cmp_omp_sigonly.png
@@ -775,18 +789,108 @@ with :math:`p` two-sided and the confidence :math:`(1 - p) \times 100` as
 above. Because the same observation appears in both runs, O is the same and
 the difference only carries the change of the background; that is why the
 paired test sees small bias changes that an independent test would miss.
-``bcorr`` uses the same test on the bias correction.
 
-An ``omp`` or ``oma`` box is significant only when **both** tests are
-above 95 %, and the significance map shows the lower of the two. On boxes of
-10° or more the confidences are written in the box, e.g. ``t99 F97``. An
-experience that moves the bias without touching the spread therefore stays
-mostly non-significant on the ``omp`` map; its ``t`` value in the box shows
-it.
+An ``omp`` or ``oma`` box counts only when **both** tests are above 95 %
+**and** the bias and the spread moved in the same direction. The
+significance map shows the lower of the two confidences, and on boxes of
+10° or more both are written in the box, e.g. ``t99 F97``. The next section
+explains why both are needed.
+
+Why omp and oma need both tests
+-------------------------------
+
+Think of the observation as the centre of a target, ``[O]``, and of each
+O-B as a shot. The **t-test** looks at the aim: did the average shot move
+closer to the centre, or further? The **F-test** looks at the hand: are the
+shots more grouped, or more scattered?
+
+Both tests are two-sided. Passing one means *this changed more than chance
+allows*, in either direction; the sign of the change says which way. So a
+box needs both tests above 95 %, and both changes pointing the same way.
+
+**Case 1: better aim, shakier hand.** The experience brings the average
+onto the centre, but its shots are all over the place.
+
+.. code-block:: text
+
+    CONTROL                          EXPERIENCE
+    off-centre, grouped              centred, scattered
+
+                                            E
+                  CCC                              E
+         [O]      CCC                       [O]
+                  CCC                   E              E
+                                               E
+
+    t-test: passes, the bias is smaller      (better)
+    F-test: passes, the spread is larger     (worse)
+
+The average looks perfect, but the model lost its consistency. The two
+changes disagree, so the box is **not drawn** on the significant-only map
+and is crossed (``xxx``) on the significance map.
+
+**Case 2: steady hand, wrong aim.** The experience groups its shots
+beautifully, but the whole group has drifted away from the centre.
+
+.. code-block:: text
+
+    CONTROL                          EXPERIENCE
+    centred, scattered               grouped, off-centre
+
+         C
+     C       C                                      EEE
+        [O]                                 [O]     EEE
+     C       C                                      EEE
+         C
+
+    t-test: passes, the bias is larger       (worse)
+    F-test: passes, the spread is smaller    (better)
+
+The spread went down, but a systematic error came in. Again the changes
+disagree: **not drawn**, crossed on the significance map.
+
+**Case 3: better on both.** The experience corrects the bias and tightens
+the spread. This is the result you are looking for.
+
+.. code-block:: text
+
+    CONTROL                          EXPERIENCE
+    off-centre and scattered         centred and grouped
+
+            C
+                                               EEE
+    [O]                                       E[O]E
+                    C                          EEE
+            C
+
+    t-test: passes, the bias is smaller      (better)
+    F-test: passes, the spread is smaller    (better)
+
+The box is drawn in **red** on the significant-only map and counted as
+*better*.
+
+**Case 4: worse on both.** The mirror image of case 3: the experience moves
+away from the centre and scatters more. The box is drawn in **blue** and
+counted as *worse*, which is just as useful to know.
+
+A last case is the quiet one: the bias improves but the spread does not
+change at all. The t-test passes, the F-test does not, and the box stays
+out of the significant-only map. Its ``t`` value, written in the box on the
+significance map, still shows that something moved.
+
+In one line, for ``omp`` and ``oma``:
+
+.. code-block:: text
+
+    drawn in red   = bias better AND spread better, both above 95 %
+    drawn in blue  = bias worse  AND spread worse,  both above 95 %
+    not drawn      = anything else (crossed when both pass in opposite directions)
 
 Reading the tests
 -----------------
 
+* ``stdomp`` and ``stdoma`` only look at the spread, so the F-test alone
+  decides there. Only ``omp`` and ``oma`` combine both.
 * The F-test treats the two runs as independent samples. With common
   observations they are strongly correlated, which makes the F-test
   cautious: a real change of the spread needs more observations to pass.
@@ -863,8 +967,8 @@ Live examples:
 The ``scatter_*.db`` files hold the per-box sums of each run (tables
 ``moyenne`` and ``stations``), the ``scatter_pair_*.db`` files the sums of
 the common observations (table ``pairs``, columns ``n_omp``, ``sx_omp``,
-``sy_omp``, ``sxx_omp``, ``syy_omp``, ``sxy_omp`` and the same for ``oma``
-and ``bc``). Both can be queried with ``sqlite3``. ``PATHWORK/<family>``
+``sy_omp``, ``sxx_omp``, ``syy_omp``, ``sxy_omp`` and the same with
+``oma``). Both can be queried with ``sqlite3``. ``PATHWORK/<family>``
 is wiped at the start of every run that passes the input check.
 
 ----
@@ -889,7 +993,7 @@ is wiped at the start of every run that passes the input check.
 +-------------------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
 | ``WARNING: n/N ... tasks lost, KilledWorker``         | A worker ran out of memory. The other maps are still produced. Use fewer workers (``N_CPUS``) or explicit channels instead of ``join`` / ``all``.                                                |
 +-------------------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
-| No *Difference / Significance* selector               | Only with a control, and only for omp, oma, stdomp, stdoma, bcorr: pick one of them in *Function* (the page opens on the first function alphabetically).                                         |
+| No *Difference / Significance* selector               | Only with a control, and only for omp, oma, stdomp, stdoma: pick one of them in *Function* (the page opens on the first function alphabetically).                                                |
 +-------------------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
 | No bcorr maps                                         | The family has no ``BIAS_CORR`` column (non-radiance); the log says so once.                                                                                                                     |
 +-------------------------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
@@ -1002,7 +1106,7 @@ _NICE_PERCENTILE = 95     # the scale covers this percentile of |value|
 
 # Significance of a box difference: test confidence (1 - p) * 100 >= this.
 # omp/oma need both the t-test (bias) and the F-test (spread),
-# stdomp/stdoma the F-test, bcorr the t-test; obs, nobs, dens: no test.
+# stdomp/stdoma the F-test; obs, bcorr, nobs, dens: no test.
 _MIN_CONFIDENCE = 95.0
 
 _WHITE_RGBA   = (1.0, 1.0, 1.0, 1.0)
@@ -1012,20 +1116,20 @@ _WHITE_RGBA   = (1.0, 1.0, 1.0, 1.0)
 PANEL_VALUE   = 'difference'
 PANEL_CONF    = 'significance'
 PANEL_SIGONLY = 'significant only'
-_TESTED_FUNCS = ('omp', 'oma', 'stdomp', 'stdoma', 'bcorr')
+_TESTED_FUNCS = ('omp', 'oma', 'stdomp', 'stdoma')
 # Tested functions are compared on the observations common to both runs,
 # matched one by one; quantity of the pair table used by each function.
 _PAIR_Q = {'omp': 'omp', 'stdomp': 'omp',
-           'oma': 'oma', 'stdoma': 'oma',
-           'bcorr': 'bc'}
-_PAIR_QUANTITIES = ('omp', 'oma', 'bc')
+           'oma': 'oma', 'stdoma': 'oma'}
+_PAIR_QUANTITIES = ('omp', 'oma')
 # Observation key used to match control and experience: same station,
 # date, time, position (1e-4 degree), vertical coordinate / channel and
 # variable (varno separates the variables of one report).
 _PAIR_KEY = ('k_stn', 'k_date', 'k_time', 'k_lat', 'k_lon',
              'k_vcoord', 'k_varno')
 # difference maps expressed in percent
-_PERCENT_FUNCS = ('stdomp', 'stdoma', 'nobs', 'NOBSHDR', 'dens', 'dens%')
+_PERCENT_FUNCS = ('stdomp', 'stdoma', 'nobs', 'NOBSHDR', 'dens', 'dens%',
+                  'bcorr')
 
 # Significance map: greys below the threshold, colours at/above it.
 _CONF_BOUNDS = (0, 50, 80, 90, 95, 99, 100)
@@ -2636,6 +2740,21 @@ def _paired_t_confidence(mx, my, vx, vy, cxy, N) -> np.ndarray:
     return conf
 
 
+def _both_tests(d_bias, d_spread, t_conf, f_conf):
+    """omp / oma: a box counts only when the bias and the spread both changed
+    significantly AND in the same direction (both better or both worse).
+
+    Returns (values, significant, confidences, kind); confidences also
+    carries 'mixed', the boxes where both tests pass but the bias and the
+    spread moved in opposite directions.
+    """
+    both  = _is_significant(t_conf) & _is_significant(f_conf)
+    same  = np.sign(d_bias) == np.sign(d_spread)
+    mixed = both & ~same
+    return (d_bias, both & same,
+            {'t': t_conf, 'F': f_conf, 'mixed': mixed}, 'improve')
+
+
 def _paired_metric(fonction: str, N, SX, SY, SXX, SYY, SXY
                    ) -> Tuple[np.ndarray, np.ndarray,
                               Dict[str, np.ndarray], str]:
@@ -2645,14 +2764,10 @@ def _paired_metric(fonction: str, N, SX, SY, SXX, SYY, SXY
     reference, as in saska40_v0_hs.py:
       omp/oma        |bias_exp| - |bias_ref|           (negative = better)
       stdomp/stdoma  100 (s_exp - s_ref) / s_ref       (negative = better)
-      bcorr          bias_corr_exp - bias_corr_ref
     Tests: F-test (saska) on the spread, paired t-test on the mean.
     """
     mx, my, vx, vy, cxy = _paired_moments(N, SX, SY, SXX, SYY, SXY)
     with np.errstate(divide='ignore', invalid='ignore'):
-        if fonction == 'bcorr':
-            t_conf = _paired_t_confidence(mx, my, vx, vy, cxy, N)
-            return my - mx, _is_significant(t_conf), {'t': t_conf}, 'change'
 
         s_ref  = _sample_std(np.sqrt(vx), N)
         s_exp  = _sample_std(np.sqrt(vy), N)
@@ -2664,8 +2779,7 @@ def _paired_metric(fonction: str, N, SX, SY, SXX, SYY, SXY
 
         t_conf = _paired_t_confidence(mx, my, vx, vy, cxy, N)
         values = np.abs(my) - np.abs(mx)
-        return (values, _is_significant(t_conf) & _is_significant(f_conf),
-                {'t': t_conf, 'F': f_conf}, 'improve')
+        return _both_tests(values, s_exp - s_ref, t_conf, f_conf)
 
 
 def _diff_metric(fonction: str, avg1, avg2, std1, std2, bc1, bc2, n1, n2,
@@ -2687,10 +2801,11 @@ def _diff_metric(fonction: str, avg1, avg2, std1, std2, bc1, bc2, n1, n2,
             return avg2 - avg1, None, {}, 'change'
 
         if fonction == 'bcorr':
-            if bcs1 is None or bcs2 is None:
-                return bc2 - bc1, None, {}, 'change'
-            t_conf = _ttest_confidence(bc1, bcs1, n1, bc2, bcs2, n2)
-            return bc2 - bc1, _is_significant(t_conf), {'t': t_conf}, 'change'
+            # like nobs: change in percent over all observations, no test;
+            # magnitudes, since a correction can be negative
+            b1, b2 = np.abs(bc1), np.abs(bc2)
+            values = np.where(b1 > 0, (b2 - b1) / b1 * 100.0, np.nan)
+            return values, None, {}, 'change'
 
         s_ref  = _sample_std(std1, n1)
         s_exp  = _sample_std(std2, n2)
@@ -2703,8 +2818,7 @@ def _diff_metric(fonction: str, avg1, avg2, std1, std2, bc1, bc2, n1, n2,
 
         t_conf = _ttest_confidence(avg1, std1, n1, avg2, std2, n2)
         values = np.abs(avg2) - np.abs(avg1)
-        return (values, _is_significant(t_conf) & _is_significant(f_conf),
-                {'t': t_conf, 'F': f_conf}, 'improve')
+        return _both_tests(values, s_exp - s_ref, t_conf, f_conf)
 
 
 _CBAR_DIFF: Dict[str, str] = {
@@ -2717,7 +2831,8 @@ _CBAR_DIFF: Dict[str, str] = {
     'nobs':   'Nobs change [%], {exp} vs {ref}\nred = more observations',
     'dens':   'Density change [%], {exp} vs {ref}\nred = more observations',
     'obs':    '{var} {exp} - {ref} {units}\nred = higher value',
-    'bcorr':  'Bias corr. {exp} - {ref} {units}\nred = larger correction',
+    'bcorr':  '|Bias corr.| change [%], ({exp} - {ref}) / {ref}\n'
+              'red = larger correction',
 }
 _CBAR_SINGLE: Dict[str, str] = {
     'dens':   'Density [obs/km²/day]',
@@ -2732,28 +2847,27 @@ _CBAR_SINGLE: Dict[str, str] = {
 
 
 _TEST_TEXT = {
-    'omp':    "t-test (bias O-B) + F-test (σ O-B)\nshown: the lower confidence",
-    'oma':    "t-test (bias O-A) + F-test (σ O-A)\nshown: the lower confidence",
+    'omp':    "t-test (bias O-B) + F-test (σ O-B)\nboth > 95 %, same direction",
+    'oma':    "t-test (bias O-A) + F-test (σ O-A)\nboth > 95 %, same direction",
     'stdomp': "F-test on σ(O-B), F = σ²big / σ²small",
     'stdoma': "F-test on σ(O-A), F = σ²big / σ²small",
-    'bcorr':  "Welch t-test on the mean bias corr.",
 }
 
 
 _TEST_TEXT_PAIRED = {
-    'omp':    "paired t-test (bias) + F-test (σ), O-B\n"
-              "of common obs, lower confidence shown",
-    'oma':    "paired t-test (bias) + F-test (σ), O-A\n"
-              "of common obs, lower confidence shown",
+    'omp':    "paired t-test (bias) + F-test (σ)\n"
+              "on O-B of common obs, both > 95 %, same direction",
+    'oma':    "paired t-test (bias) + F-test (σ)\n"
+              "on O-A of common obs, both > 95 %, same direction",
     'stdomp': "F-test on σ(O-B) of common obs,\nF = σ²big / σ²small",
     'stdoma': "F-test on σ(O-A) of common obs,\nF = σ²big / σ²small",
-    'bcorr':  "paired t-test on the bias corr., common obs",
 }
 
 
 def _info_text(panel, fonction, is_diff, kind, confs, values, diff_values,
                signif, in_axes, boundaries, on_map_obs, on_map_ref, n_days,
-               avg=None, std=None, n_black=0, n_pairs=None) -> str:
+               avg=None, std=None, n_black=0, n_pairs=None,
+               n_mixed=0) -> str:
     """Text of the box at the top right of a map."""
     if not is_diff:
         if fonction in ('dens', 'nobs'):
@@ -2786,6 +2900,9 @@ def _info_text(panel, fonction, is_diff, kind, confs, values, diff_values,
         n_wor = int((box & signif & (diff_values > 0)).sum())
         counts += f'\n{n_bet} better (red), {n_wor} worse (blue)'
 
+    mixed_line = (f'\nmixed: {n_mixed} boxes, bias and σ changed\n'
+                  f'in opposite directions (not counted)'
+                  if n_mixed else '')
     black_line = (f'\nblack: {n_black} boxes not compared, no\n'
                   + ('common observation in control and experience'
                      if n_pairs is not None else
@@ -2795,17 +2912,19 @@ def _info_text(panel, fonction, is_diff, kind, confs, values, diff_values,
     if panel == PANEL_SIGONLY:
         return (f'Only boxes with confidence > {_MIN_CONFIDENCE:.0f} %\n'
                 f'({tests.get(fonction, "").splitlines()[0]})\n'
-                f'{counts}\n{nobs_line}{black_line}')
+                f'{counts}\n{nobs_line}{mixed_line}{black_line}')
 
     if panel == PANEL_CONF:
         return (f'{tests.get(fonction, "")}\n'
                 f'(1 - p) x 100, > {_MIN_CONFIDENCE:.0f} % = significant\n'
-                f'{counts}\n{nobs_line}{black_line}')
+                f'{counts}\n{nobs_line}'
+                + (mixed_line.replace('boxes,', 'boxes (xxx),')
+                   if n_mixed else '') + black_line)
 
     text = nobs_line
     if confs:
         text += f'\n{counts.replace("boxes", f"boxes at {_MIN_CONFIDENCE:.0f} %")}'
-    return text + black_line
+    return text + mixed_line + black_line
 
 
 def _render_tile_map(task: Dict[str, Any],
@@ -2957,6 +3076,10 @@ def _render_tile_map_impl(task: Dict[str, Any],
     if signif is None:
         signif = np.ones(len(values), dtype=bool)
 
+    mixed = confs.pop('mixed', None) if confs else None
+    if mixed is None:
+        mixed = np.zeros(len(values), dtype=bool)
+
     # combined confidence: the weakest test decides (omp/oma need both)
     conf = None
     if confs:
@@ -3038,6 +3161,7 @@ def _render_tile_map_impl(task: Dict[str, Any],
         on_map_ref = float(n1[in_axes].sum()) if is_diff else 0.0
 
     patches, patch_colors = [], []
+    patches_mixed = []
     for i in np.flatnonzero(in_axes):
         if panel == PANEL_SIGONLY and not signif[i]:
             continue
@@ -3058,12 +3182,21 @@ def _render_tile_map_impl(task: Dict[str, Any],
         patches.append(plt.Polygon(
             _project_polygon(proj_obj, lat[i], lon[i], deltax, deltay, pc)))
         patch_colors.append(facecolors[i])
+        if panel == PANEL_CONF and mixed[i]:
+            patches_mixed.append(plt.Polygon(
+                _project_polygon(proj_obj, lat[i], lon[i], deltax, deltay, pc)))
 
     if patches:
         ax.add_collection(PatchCollection(
             patches, facecolors=patch_colors,
             edgecolors='black', linewidths=0.2, alpha=ALPHA_TILES, zorder=4))
 
+    if patches_mixed:
+        # both tests pass, but bias and spread moved in opposite directions
+        ax.add_collection(PatchCollection(
+            patches_mixed, facecolors='none', edgecolors='#3a3a3a',
+            linewidths=0.3, hatch='xxx', zorder=5))
+
     if exclusive:
         ax.add_collection(PatchCollection(
             [plt.Polygon(_project_polygon(proj_obj, lt, ln, deltax, deltay, pc))
@@ -3132,7 +3265,8 @@ def _render_tile_map_impl(task: Dict[str, Any],
                          avg if not is_diff else None,
                          std if not is_diff else None,
                          n_black=len(exclusive),
-                         n_pairs=n_pairs_on_map)
+                         n_pairs=n_pairs_on_map,
+                         n_mixed=int((mixed & in_axes).sum()))
     # top right, next to the title block
     info_txt = ax.text(1.0, 1.02, textstr, transform=ax.transAxes,
             fontsize=11, ha='right', va='bottom',
@@ -3246,7 +3380,7 @@ _INTRO_DIFF = [
     "Each map compares one experience with the control, box by box, over "
     "{ds} to {de} ({bx}° x {by}° boxes). <b>These are not the same maps as "
     "the single-experience ones</b>: every value is experience minus control, "
-    "and for omp, oma, stdomp, stdoma and bcorr it is computed on the "
+    "and for omp, oma, stdomp and stdoma it is computed on the "
     "observations common to both runs (same station, date, time, position, "
     "channel and variable).",
 
@@ -3255,18 +3389,24 @@ _INTRO_DIFF = [
     "σ control. For these four a <b>negative value is an improvement and is "
     "drawn in red</b>; positive values (experience worse) are blue. "
     "<b>nobs, dens</b>: change of the number of observations in percent, "
-    "red = more observations. <b>obs, bcorr</b>: experience - control, "
+    "red = more observations. <b>bcorr</b>: change of the size of the bias "
+    "correction in percent, red = larger correction. <b>obs</b>: experience "
+    "- control, "
     "red = higher value.",
 
     "The <b>Difference / Significance</b> selector offers three maps for "
-    "omp, oma, stdomp, stdoma and bcorr. <b>difference</b> shows every box. "
+    "omp, oma, stdomp and stdoma. <b>difference</b> shows every box. "
     "<b>significance</b> shows the confidence (1 - p) x 100 of the test in "
     "each box: grey up to 95 %, orange above 95 % and red above 99 %. "
     "<b>significant only</b> is the difference map where only the boxes "
     "above 95 % are drawn. The spread is tested with the F-test "
-    "F = σ²larger / σ²smaller, two-sided; the bias and bcorr with a paired "
-    "t-test; omp and oma need both, and the map shows the lower confidence. "
-    "obs, nobs and dens have no test; bcorr exists for radiances only.",
+    "F = σ²larger / σ²smaller, two-sided; the bias with a paired "
+    "t-test. omp and oma need both tests above 95 % and the bias and the "
+    "spread moving the same way (both better: red, both worse: blue); when "
+    "they move in opposite directions the box is crossed on the significance "
+    "map and left out. The map shows the lower confidence. "
+    "obs, bcorr, nobs and dens have no test; bcorr exists for radiances "
+    "only.",
 
     "The scale is symmetric and adapts to each map (steps like 0.1, 0.2, 0.3, "
     "0.5, 1, 2); the step around zero is white. With thousands of boxes, "

{pikobs-3.0.415 → pikobs-3.0.417}/pikobs/script/run_scatter.sh

@@ -46,7 +46,7 @@ CONTROL_NAME="Control"
 # Output directory. WARNING: wiped on each run to prevent ghost files.
 # Will contain one subdir per family with the .db files and .png maps,
 # plus an interactive HTML viewer at pikobs_scatter_viewer.html.
-PATHWORK="/home/dlo001/sites8/pikobs_scatter_exp"
+PATHWORK="/home/dlo001/sites8/pikobs_scatter_exp2"
 
 # ---- Time window ---------------------------------------------------------
 # Processing window in YYYYMMDDHH format (UTC). Inclusive on both ends,

{pikobs-3.0.415 → pikobs-3.0.417/pikobs.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: pikobs
-Version: 3.0.415
+Version: 3.0.417
 Summary: pikobs
 Home-page: https://gitlab.science.gc.ca/dlo001/pikobs
 Author: David Lobon

pikobs-3.0.415/VERSION

@@ -1 +0,0 @@
-3.0.415