raster: preserve missing NODATA in map algebra - #1100

raster/rt_pg/rtpostgis.sql.in (1)

4503-4518: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Reserve sentinel 255 using a global RGB decision, not per-band.

reservenodata is computed from each band’s own nodata, so in mixed cases (only one RGB band has NODATA) the other 8BUI bands can keep value 255. The grayscale formula can still round to 255, which then collides with the output NODATA sentinel once Line 4544 sets band NODATA to 255.

Use a pre-scan/global flag (any input band has NODATA OR non-INTERSECTION extent) and apply the 0..254 reclass to all three bands when reservation is needed.

Suggested patch sketch

-        hasnodata boolean DEFAULT FALSE;
-        reservenodata boolean;
+        hasnodata boolean DEFAULT FALSE;
+        reservenodata boolean DEFAULT FALSE;
+        _extenttype text;

@@
+        _extenttype := upper(coalesce(extenttype, 'INTERSECTION'));
+
+        -- First pass: detect whether any source band has NODATA
+        FOR idx IN 1.._NBANDS LOOP
+            rast := _set[idx].rast;
+            nband := _set[idx].nband;
+            nodata := `@extschema`@.ST_BandNoDataValue(rast, nband);
+            IF nodata IS NOT NULL THEN
+                hasnodata := TRUE;
+                EXIT;
+            END IF;
+        END LOOP;
+
+        reservenodata := hasnodata OR _extenttype != 'INTERSECTION';
+
+        -- Second pass: normalize/reclass each band
         FOR idx IN 1.._NBANDS LOOP
@@
-            reservenodata := nodata IS NOT NULL OR upper(extenttype) != 'INTERSECTION';
-
             IF `@extschema`@.ST_BandPixelType(rast, nband) != _PIXTYPE OR reservenodata THEN
@@
-        IF hasnodata OR upper(extenttype) != 'INTERSECTION' THEN
+        IF reservenodata THEN
             grayscale := `@extschema`@.ST_SetBandNoDataValue(grayscale, 1, _NODATA);
         END IF;

Also applies to: 4544-4544

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@raster/rt_pg/rtpostgis.sql.in` around lines 4503 - 4518, The `reservenodata`
flag is currently computed per-band based on each individual band's nodata
value, which causes a collision when different RGB bands have different nodata
settings and the grayscale formula rounds to 255. Create a global pre-scan flag
that checks if ANY input band has NODATA OR the extent type is not INTERSECTION,
before the per-band loop. Use this global flag to apply the 0..254 reclass (the
reclassexpr logic) consistently to all three bands when reservation is needed,
rather than making per-band decisions. This ensures that when reservation is
triggered for any reason, all bands get the same restrictive 0..254 range to
prevent collisions with the sentinel value 255 set at line 4544.

raster/rt_pg/rtpostgis.sql.in (1)

4527-4538: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Avoid stats-stretching native 8BUI bands when reserving 255.

Line 4527 now reclassifies already-8BUI bands whenever reservenodata is true, but lines 4533-4538 map stats.min..stats.max into 0..254. For native RGB input, that changes valid color values whenever the data range is narrower than 0..255; reserve the sentinel over the full 8BUI domain instead, and keep stats-based normalization for non-8BUI inputs.

🐛 Proposed fix

 		nodata double precision;
 		nodataval integer;
+		pixtype text;
 		reclassexpr text;
 		hasnodata boolean DEFAULT FALSE;
@@
-			IF `@extschema`@.ST_BandPixelType(rast, nband) != _PIXTYPE OR reservenodata THEN
-				stats := `@extschema`@.ST_SummaryStats(rast, nband);
+			pixtype := `@extschema`@.ST_BandPixelType(rast, nband);
+			IF pixtype != _PIXTYPE OR reservenodata THEN
+				IF pixtype = _PIXTYPE THEN
+					stats.min := 0;
+					stats.max := _NODATA;
+				ELSE
+					stats := `@extschema`@.ST_SummaryStats(rast, nband);
+				END IF;
 
 				IF nodata IS NOT NULL THEN
 					nodataval := _NODATA;

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@raster/rt_pg/rtpostgis.sql.in` around lines 4527 - 4538, The condition at
line 4527 incorrectly applies stats-based remapping whenever reservenodata is
true, even for native 8BUI bands, which changes valid color values when the data
range is narrower than 0..255. Modify the condition to only apply stats-based
normalization for non-8BUI inputs by adding an additional check to determine if
the band is a native 8BUI type. When the band is already 8BUI and reservenodata
is true, skip the stats calculation and reclassification in lines 4533-4538 and
instead simply set nodataval to _NODATA without remapping the data, preserving
the original color values while reserving 255 as the sentinel.