Training preparation

.flake8

@@ -9,8 +9,8 @@ exclude = .git
 max-line-length = 87
 
 per-file-ignores =
-    ./m.neural_network.preparedata/m.neural_network.preparedata.py: F821
-    ./m.neural_network.preparedata/m.neural_network.preparedata.py: E501
-    ./m.neural_network.preparedata.worker_nullcells/m.neural_network.preparedata.worker_nullcells.py: F821
-    ./m.neural_network.preparedata.worker_nullcells/m.neural_network.preparedata.worker_nullcells.py: E501
+    ./m.neural_network.preparedata/m.neural_network.preparedata.py: E501,F821
+    ./m.neural_network.preparedata.worker_nullcells/m.neural_network.preparedata.worker_nullcells.py: F821,E501
     ./m.neural_network.preparedata.worker_export/m.neural_network.preparedata.worker_export.py: E501
+    ./m.neural_network.preparetraining/m.neural_network.preparetraining.py: E501,F821
+    ./m.neural_network.preparetraining.worker/m.neural_network.preparetraining.worker.py: E501,F821

m.neural_network.html

@@ -30,7 +30,8 @@ <h2>DESCRIPTION</h2>
 <li><a href="m.neural_network.preparedata.html">m.neural_network.preparedata</a>: Prepares and exports tiles for the label process</li>
 <li><a href="m.neural_network.preparedata.worker_export.html">m.neural_network.preparedata.worker_export</a>: Worker for parallel processing for exporting for <b>m.neural_network.preparedata</b></li>
 <li><a href="m.neural_network.preparedata.worker_nullcells.html">m.neural_network.preparedata.worker_nullcells</a>: Worker to analyse the number of null cells in parallel for <b>m.neural_network.preparedata</b></li>
-<li><a href="m.neural_network.preparetraining.html">m.neural_network.preparetraining</a>: Takes and reformats labeled tiles for the neural network training</li>
+<li><a href="m.neural_network.preparetraining.html">m.neural_network.preparetraining</a>: Prepares imagery and labelled data for training and application of a neural network.</li>
+<li><a href="m.neural_network.preparetraining.worker.html">m.neural_network.preparetraining</a>: Worker to rasterize labelled data in parallel for <b>m.neural_network.preparetraining</b><</li>
 </ul>
 
 <h2>REQUIREMENTS</h2>
@@ -39,6 +40,7 @@ <h2>REQUIREMENTS</h2>
 
 <ul>
   <li>grass-gis-helpers>=2.2.0</li>
+  <li>GDAL/OGR and Python bindings</li>
 </ul>
 
 <h2>AUTHORS</h2>

m.neural_network.preparetraining.worker/Makefile

@@ -0,0 +1,7 @@
+MODULE_TOPDIR = ../..
+
+PGM = m.neural_network.preparetraining.worker
+
+include $(MODULE_TOPDIR)/include/Make/Script.make
+
+default: script

m.neural_network.preparetraining.worker/m.neural_network.preparetraining.worker.html

@@ -0,0 +1,14 @@
+<h2>DESCRIPTION</h2>
+
+<em>m.neural_network.preparetraining.worker</em> is used within <em>m.neural_network.preparetraining</em> to rasterize label data in parallel.
+<p>
+<h2>SEE ALSO</h2>
+
+<em>
+<a href="g.region.html">g.region</a>
+<a href="r.mapcalc.html">r.mapcalc</a>,
+<a href="v.to.rast.html">v.to.rast</a>,
+</em>
+
+<h2>AUTHORS</h2>
+<p>Guido Riembauer, <a href="https://www.mundialis.de/">mundialis GmbH &amp; Co. KG</a><br>

m.neural_network.preparetraining.worker/m.neural_network.preparetraining.worker.py

@@ -0,0 +1,221 @@
+#!/usr/bin/env python3
+"""############################################################################
+#
+# MODULE:      m.neural_network.preparetraining.worker
+# AUTHOR(S):   Guido Riembauer
+# PURPOSE:     Worker module for m.neural_network.preparetraining to check
+#              and rasterize label data
+# COPYRIGHT:   (C) 2024 by mundialis GmbH & Co. KG and the GRASS Development
+#              Team.
+#
+# 		This program is free software under the GNU General Public
+# 		License (v3). Read the file COPYING that comes with GRASS
+# 		for details.
+#
+##############################################################################
+"""
+
+# %Module
+# % description: Worker module for m.neural_network.preparetraining to check and rasterize label data
+# % keyword: raster
+# % keyword: statistics
+# %end
+
+# %option G_OPT_F_INPUT
+# % required: yes
+# % multiple: no
+# % label: Path to the label vector file
+# % guisection: Input
+# %end
+
+# %option G_OPT_F_INPUT
+# % key: img_path
+# % required: yes
+# % multiple: no
+# % label: Path to the corresponding imagery raster file
+# % guisection: Input
+# %end
+
+# %option
+# % key: class_column
+# % type: string
+# % required: yes
+# % multiple: no
+# % answer: class_number
+# % label: Column of the label vector that holds the class number
+# % guisection: Parameters
+# %end
+
+# %option
+# % key: class_values
+# % type: integer
+# % required: yes
+# % multiple: yes
+# % answer: 2
+# % label: Expected and output values for the class/es of interest
+# % guisection: Parameters
+# %end
+
+# %option
+# % key: no_class_value
+# % type: integer
+# % required: yes
+# % multiple: no
+# % answer: 1
+# % label: Expected and output value for the non class of interest areas
+# % description: Can be understood as a "rest" class for a multiclass system and a "no-class" for a binary classification
+# % guisection: Parameters
+# %end
+
+# %option G_OPT_F_OUTPUT
+# % required: yes
+# % multiple: no
+# % label: Path to the output label raster file
+# % guisection: Output
+# %end
+
+# %option
+# % key: new_mapset
+# % type: string
+# % required: yes
+# % multiple: no
+# % label: Name of the new mapset to work in
+# % guisection: Parameters
+# %end
+
+import atexit
+import os
+import shutil
+
+import grass.script as grass
+from grass_gis_helpers.mapset import switch_to_new_mapset
+from osgeo import gdal
+
+NEWGISRC = None
+GISRC = None
+ID = grass.tempname(8)
+NEW_MAPSET = None
+
+
+def cleanup():
+    """Switch mapsets and deleting the new one."""
+    # switch back to original mapset
+    grass.utils.try_remove(NEWGISRC)
+    os.environ["GISRC"] = GISRC
+    # delete the new mapset (doppelt haelt besser)
+    gisenv = grass.gisenv()
+    gisdbase = gisenv["GISDBASE"]
+    location = gisenv["LOCATION_NAME"]
+    mapset_dir = os.path.join(gisdbase, location, NEW_MAPSET)
+    if os.path.isdir(mapset_dir):
+        shutil.rmtree(mapset_dir)
+
+
+def main():
+    """Run label rasterization."""
+    global NEWGISRC, GISRC, NEW_MAPSET
+    input = options["input"]
+    img_file = options["img_path"]
+    NEW_MAPSET = options["new_mapset"]
+    class_values = options["class_values"].split(",")
+    no_class_value = options["no_class_value"]
+    class_col = options["class_column"]
+    output = options["output"]
+
+    # switch to the new mapset
+    GISRC, NEWGISRC, old_mapset = switch_to_new_mapset(NEW_MAPSET)
+    # get extent from reference img file
+    info = gdal.Info(img_file, format="json")
+    south = info["cornerCoordinates"]["lowerLeft"][1]
+    west = info["cornerCoordinates"]["lowerLeft"][0]
+    north = info["cornerCoordinates"]["upperRight"][1]
+    east = info["cornerCoordinates"]["upperRight"][0]
+    cols, rows = info["size"]
+    # set the region
+    grass.run_command(
+        "g.region",
+        n=north,
+        s=south,
+        e=east,
+        w=west,
+        rows=rows,
+        cols=cols,
+        quiet=True,
+    )
+
+    # import the label dataset
+    labelvect = f"labelvect_{ID}"
+    labelrast = f"labelrast_{ID}"
+    grass.run_command("v.import", input=input, output=labelvect, quiet=True)
+
+    # check the values of the vector
+    dbselect = list(grass.parse_command("v.db.select", map=labelvect).keys())
+    colnames = dbselect[0].split("|")
+    rows = [item.split("|") for item in dbselect[1:]]
+    try:
+        idx = colnames.index(class_col)
+    except ValueError:
+        grass.fatal(_(f"File {input} has no column {class_col}"))
+    class_numbers = [item[idx] for item in rows]
+    class_num_set_ref = set([*class_values, no_class_value])
+    difference = set(class_numbers).difference(class_num_set_ref)
+    if len(difference) > 0:
+
+        grass.fatal(
+            _(
+                f"Label file {input} has features with unexpected values"
+                f" in column {class_col}: {difference}. Allowed values "
+                f"are [{','.join(class_values)}, {no_class_value}].",
+            ),
+        )
+
+    tile_empty = False
+    if len(class_numbers) == 0 or set(class_numbers) == set([no_class_value]):
+        grass.warning(
+            _(
+                f"Label file {input} contains no features with the "
+                f"expected class values {class_values} in "
+                f"column {class_col}. It is assumed that the classes "
+                "do not occur in this tile.",
+            ),
+        )
+        tile_empty = True
+
+    # rasterize
+    if tile_empty is True:
+        grass.run_command(
+            "r.mapcalc",
+            expression=f"{labelrast}={no_class_value}",
+            quiet=True,
+        )
+    else:
+        labelrast_tmp = f"{labelrast}_tmp"
+        grass.run_command(
+            "v.to.rast",
+            input=labelvect,
+            output=labelrast_tmp,
+            type="area",
+            use="attr",
+            attribute_column=class_col,
+            quiet=True,
+        )
+        # if there is any nodata left in the label, this will be assigned
+        # to the no-class class
+        exp = f"{labelrast}=if(isnull({labelrast_tmp}),{no_class_value},{labelrast_tmp})"
+        grass.run_command("r.mapcalc", expression=exp, quiet=True)
+
+    grass.run_command(
+        "r.out.gdal",
+        input=labelrast,
+        output=output,
+        type="Byte",
+        createopt="COMPRESS=LZW",  # no tiles or overviews required for the small tiles (?)
+        flags="c",
+        quiet=True,
+    )
+
+
+if __name__ == "__main__":
+    options, flags = grass.parser()
+    atexit.register(cleanup)
+    main()

m.neural_network.preparetraining/Makefile

@@ -0,0 +1,7 @@
+MODULE_TOPDIR = ../..
+
+PGM = m.neural_network.preparetraining
+
+include $(MODULE_TOPDIR)/include/Make/Script.make
+
+default: script

m.neural_network.preparetraining/m.neural_network.preparetraining.html

@@ -0,0 +1,63 @@
+<h2>DESCRIPTION</h2>
+
+<em>m.neural_network.preparetraining</em> prepares imagery and labelled data for training and application of a neural network.
+<p>While <a href="m.neural_network.preparedata">m.neural_network.preparedata</a> initially provides a setup for labelling tiles of imagery,
+<em>m.neural_network.preparetraining</em> rasterizes the vector labels and restructures the imagery data.
+
+<h2>Notes</h2>
+It is expected that all data lie in the directory structure and naming format as created by <a href="m.neural_network.preparedata">m.neural_network.preparedata</a>.
+This data is provided to <em>m.neural_network.preparetraining</em> via the <em>input_traindir</em> and <em>input_applydir</em> parameters.
+<em>m.neural_network.preparetraining</em> creates a new directory with the two directories <em>train</em> and <em>apply</em>. Each of these contains
+the following directories/data:
+
+ <ul>
+  <li><em>train_images:</em>: contains tilewise multiband .vrt-files with all imagery bands and an ndsm band to be used for training. This directory is empty in the <em>apply</em> dir.</li>
+  <li><em>train_masks:</em>: contains tilewise rasterized .tif label files to be used for training. This directory is empty in the <em>apply</em> dir.</li>
+  <li><em>val_images:</em>: contains tilewise multiband .vrt-files with all imagery bands and an ndsm band to be used for validation. This directory holds data both in the <em>train</em> and <em>apply</em> dirs. In the <em>train</em> dir, this data is used for validation during training, while in the <em>apply</em> dir, this directory holds all imagery used for prediction.</li>
+  <li><em>val_masks:</em>: contains tilewise rasterized .tif label files to be used for training. This directory is empty in the <em>apply</em> dir.</li>
+  <li><em>singleband_vrts:</em>: contains singleband .vrts for each imagery band of each tile. They are stored here as a basis to create the tilewise multiband .vrts.</li>
+  <li><em>tile_XX_YY.vrt:</em> (only in the <em>train</em> dir): One multiband tile .vrt is stored here for the NN model to read in the number of bands.</li>
+ </ul>
+<p>
+In order to save diskspace, all imagery is stored as .vrts, so the original datasets (created by <a href="m.neural_network.preparedata">m.neural_network.preparedata</a>) should
+  not be moved (or <em>m.neural_network.preparetraining</em> should be run again afterwards).
+</p>
+<p>
+  The user can indicate what percentage of the training tiles are used for validation (during training) with the <em>val_percentage</em> parameter.
+</p>
+<p>
+  It is not possible to run <em>m.neural_network.preparetraining</em> repeatedly with the same <em>output</em> directory, as the training/validation split up happens during runtime.
+  Hence, <em>m.neural_network.preparetraining</em> expects that the <em>output</em> directory does not exist.
+</p>
+<p>
+  With the <em>class_values</em> and the <em>no_class_value</em> parameters, the user defines the allowed range of values in the <em>class_column</em> of the labelled data. In
+  case an unexpected value is found, an error is thrown which indicates the affected tile.
+</p>
+<p>
+ If a tile is not completely covered either by <em>class_values</em> or <em>no_class_value</em>, the not allocated areas will be filled with <em>no_class_value</em> in the rasterized version.
+</p>
+
+<h2>EXAMPLES</h2>
+
+<div class="code"><pre>
+  m.neural_network.preparetraining input_traindir=nn_data_with_labels/train input_applydir=nn_data_with_labels/apply nprocs=6 class_column=class_number class_values=2 no_class_value=1 output=nn_data_structured
+</pre></div>
+
+
+<h2>SEE ALSO</h2>
+
+<em>
+<a href="https://grass.osgeo.org/grass-stable/manuals/v.import.html">v.import</a>,
+<a href="https://grass.osgeo.org/grass-stable/manuals/g.region.html">g.region</a>
+<a href="https://grass.osgeo.org/grass-stable/manuals/r.mapcalc.html">r.mapcalc</a>,
+<a href="https://grass.osgeo.org/grass-stable/manuals/v.to.rast.html">v.to.rast</a>,
+</em>
+
+<h2>REQUIREMENTS</h2>
+ <ul>
+  <li>GDAL and OGR Python bindings</li>
+   <li><a href="https://pypi.org/project/grass-gis-helpers/">grass-gis-helpers</a> Python library >= 2.2.0</li>
+ </ul>
+
+<h2>AUTHORS</h2>
+Guido Riembauer, <a href="https://www.mundialis.de/">mundialis GmbH &amp; Co. KG</a><br>