py4dgeo 0.6.0__cp312-cp312-win_amd64.whl

py4dgeo/segmentation.py

@@ -0,0 +1,1202 @@
+from py4dgeo.epoch import Epoch, as_epoch
+from py4dgeo.logger import logger_context
+from py4dgeo.util import Py4DGeoError, find_file
+from py4dgeo.UpdateableZipFile import UpdateableZipFile
+
+import datetime
+import json
+import logging
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+import os
+import pickle
+import seaborn
+import tempfile
+import zipfile
+
+import _py4dgeo
+
+
+# Get the py4dgeo logger instance
+logger = logging.getLogger("py4dgeo")
+
+
+# This integer controls the versioning of the _segmentation file format. Whenever the
+# format is changed, this version should be increased, so that py4dgeo can warn
+# about incompatibilities of py4dgeo with loaded data. This version is intentionally
+# different from py4dgeo's version, because not all releases of py4dgeo necessarily
+# change the _segmentation file format and we want to be as compatible as possible.
+PY4DGEO_SEGMENTATION_FILE_FORMAT_VERSION = 0
+
+
+class SpatiotemporalAnalysis:
+    def __init__(self, filename, compress=True, allow_pickle=True, force=False):
+        """Construct a spatiotemporal _segmentation object
+
+        This is the basic data structure for the 4D objects by change algorithm
+        and its derived variants. It manages storage of M3C2 distances and other
+        intermediate results for a time series of epochs. The original point clouds
+        themselves are not needed after initial distance calculation and additional
+        epochs can be added to an existing analysis. The class uses a disk backend
+        to store information and allows lazy loading of additional data like e.g.
+        M3C2 uncertainty values for postprocessing.
+
+        :param filename:
+            The filename used for this analysis. If it does not exist on the file
+            system, a new analysis is created. Otherwise, the data is loaded from the existent file.
+        :type filename: str
+        :param compress:
+            Whether to compress the stored data. This is a tradeoff decision between
+            disk space and runtime. Especially appending new epochs to an existing
+            analysis is an operation whose runtime can easily be dominated by
+            decompression/compression of data.
+        :type compress: bool
+        :param allow_pickle:
+            Whether py4dgeo is allowed to use the pickle module to store some data
+            in the file representation of the analysis. If set to false, some data
+            may not be stored and needs to be recomputed instead.
+        :type allow_pickle: bool
+        :param force:
+            Force creation of a new analysis object, even if a file of this name
+            already exists.
+        """
+
+        # Store the given parameters
+        self.filename = find_file(filename, fatal=False)
+        self.compress = compress
+        self.allow_pickle = allow_pickle
+
+        # Instantiate some properties used later on
+        self._m3c2 = None
+
+        # This is the cache for lazily loaded data
+        self._corepoints = None
+        self._distances = None
+        self._smoothed_distances = None
+        self._uncertainties = None
+        self._reference_epoch = None
+
+        # If the filename does not already exist, we create a new archive
+        if force or not os.path.exists(self.filename):
+            logger.info(f"Creating analysis file {self.filename}")
+            with zipfile.ZipFile(self.filename, mode="w") as zf:
+                # Write the _segmentation file format version number
+                zf.writestr(
+                    "SEGMENTATION_FILE_FORMAT",
+                    str(PY4DGEO_SEGMENTATION_FILE_FORMAT_VERSION),
+                )
+
+                # Write the compression algorithm used for all suboperations
+                zf.writestr("USE_COMPRESSION", str(self.compress))
+
+        # Assert that the _segmentation file format is still valid
+        with zipfile.ZipFile(self.filename, mode="r") as zf:
+            # Read the _segmentation file version number and compare to current
+            version = int(zf.read("SEGMENTATION_FILE_FORMAT").decode())
+            if version != PY4DGEO_SEGMENTATION_FILE_FORMAT_VERSION:
+                raise Py4DGeoError("_segmentation file format is out of date!")
+
+            # Read the compression algorithm
+            self.compress = eval(zf.read("USE_COMPRESSION").decode())
+
+    @property
+    def reference_epoch(self):
+        """Access the reference epoch of this analysis"""
+
+        if self._reference_epoch is None:
+            with zipfile.ZipFile(self.filename, mode="r") as zf:
+                # Double check that the reference has already been set
+                if "reference_epoch.zip" not in zf.namelist():
+                    raise Py4DGeoError("Reference epoch for analysis not yet set")
+
+                # Extract it from the archive
+                with tempfile.TemporaryDirectory() as tmp_dir:
+                    ref_epochfile = zf.extract("reference_epoch.zip", path=tmp_dir)
+                    self._reference_epoch = Epoch.load(ref_epochfile)
+
+        return self._reference_epoch
+
+    @reference_epoch.setter
+    def reference_epoch(self, epoch):
+        """Set the reference epoch of this analysis (only possible once)"""
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            # If we already have a reference epoch, the user should start a
+            # new analysis instead
+            if "reference_epoch.zip" in zf.namelist():
+                raise Py4DGeoError(
+                    "Reference epoch cannot be changed - please start a new analysis"
+                )
+
+            # Ensure that we do have a timestamp on the epoch
+            epoch = check_epoch_timestamp(epoch)
+
+            # Ensure that the KDTree is built - no-op if triggered by the user
+            epoch.build_kdtree()
+
+            # Write the reference epoch into the archive
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                epochfilename = os.path.join(tmp_dir, "reference_epoch.zip")
+                epoch.save(epochfilename)
+                zf.write(epochfilename, arcname="reference_epoch.zip")
+
+        # Also cache it directly
+        self._reference_epoch = epoch
+
+    @reference_epoch.deleter
+    def reference_epoch(self):
+        self._reference_epoch = None
+
+    @property
+    def corepoints(self):
+        """Access the corepoints of this analysis"""
+        if self._corepoints is None:
+            with zipfile.ZipFile(self.filename, mode="r") as zf:
+                # Double check that the reference has already been set
+                if "corepoints.zip" not in zf.namelist():
+                    raise Py4DGeoError("Corepoints for analysis not yet set")
+
+                # Extract it from the archive
+                with tempfile.TemporaryDirectory() as tmp_dir:
+                    cpfile = zf.extract("corepoints.zip", path=tmp_dir)
+                    self._corepoints = Epoch.load(cpfile)
+
+        return self._corepoints
+
+    @corepoints.setter
+    def corepoints(self, _corepoints):
+        """Set the corepoints for this analysis (only possible once)"""
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            # If we already have corepoints in the archive, the user should start a
+            # new analysis instead
+            if "corepoints.zip" in zf.namelist():
+                raise Py4DGeoError(
+                    "Corepoints cannot be changed - please start a new analysis"
+                )
+
+            # Ensure that the corepoints are stored as an epoch and build its KDTree
+            self._corepoints = as_epoch(_corepoints)
+            self._corepoints.build_kdtree()
+
+            # Write the corepoints into the archive
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                cpfilename = os.path.join(tmp_dir, "corepoints.zip")
+                self._corepoints.save(cpfilename)
+                zf.write(cpfilename, arcname="corepoints.zip")
+
+    @corepoints.deleter
+    def corepoints(self):
+        self._corepoints = None
+
+    @property
+    def m3c2(self):
+        """Access the M3C2 algorithm of this analysis"""
+        # If M3C2 has not been set, we use a default constructed one
+        return self._m3c2
+
+    @m3c2.setter
+    def m3c2(self, _m3c2):
+        """Set the M3C2 algorithm of this analysis"""
+        self._m3c2 = _m3c2
+
+    @property
+    def timedeltas(self):
+        """Access the sequence of time stamp deltas for the time series"""
+        with zipfile.ZipFile(self.filename, mode="r") as zf:
+            if "timestamps.json" not in zf.namelist():
+                return []
+
+            # Read timedeltas
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                timestampsfile = zf.extract("timestamps.json", path=tmp_dir)
+                with open(timestampsfile) as f:
+                    timedeltas = json.load(f)
+
+                # Convert the serialized deltas to datetime.timedelta
+                return [datetime.timedelta(**data) for data in timedeltas]
+
+    @timedeltas.setter
+    def timedeltas(self, _timedeltas):
+        """Set the timedeltas manually
+
+        This is only possible exactly once and mutually exclusive with adding
+        epochs via the :ref:`add_epochs` method.
+        """
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            # If we already have timestamps in the archive, this is not possible
+            if "timestamps.json" in zf.namelist():
+                raise Py4DGeoError(
+                    "Timestamps can only be set on freshly created analysis instances"
+                )
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                timestampsfile = os.path.join(tmp_dir, "timestamps.json")
+                with open(timestampsfile, "w") as f:
+                    json.dump(
+                        [
+                            {
+                                "days": td.days,
+                                "seconds": td.seconds,
+                                "microseconds": td.microseconds,
+                            }
+                            for td in _timedeltas
+                        ],
+                        f,
+                    )
+                zf.write(timestampsfile, arcname="timestamps.json")
+
+    @property
+    def distances(self):
+        """Access the M3C2 distances of this analysis"""
+
+        if self._distances is None:
+            with zipfile.ZipFile(self.filename, mode="r") as zf:
+                filename = self._numpy_filename("distances")
+                if filename not in zf.namelist():
+                    self.distances = np.empty(
+                        (self.corepoints.cloud.shape[0], 0), dtype=np.float64
+                    )
+                    return self._distances
+
+                with tempfile.TemporaryDirectory() as tmp_dir:
+                    distancefile = zf.extract(filename, path=tmp_dir)
+                    read_func = (
+                        (lambda f: np.load(f)["arr_0"]) if self.compress else np.load
+                    )
+                    self._distances = read_func(distancefile)
+
+        return self._distances
+
+    @distances.setter
+    def distances(self, _distances):
+        """Set the distances manually
+
+        This is only possible exactly once and mutually exclusive with adding
+        epochs via the :ref:`add_epochs` method.
+        """
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            filename = self._numpy_filename("distances")
+            write_func = np.savez_compressed if self.compress else np.save
+
+            # If we already have distacces in the archive, this is not possible
+            if filename in zf.namelist():
+                raise Py4DGeoError(
+                    "Distances can only be set on freshly created analysis instances, use add_epochs instead."
+                )
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                distancesfile = os.path.join(tmp_dir, filename)
+                write_func(distancesfile, _distances)
+                zf.write(distancesfile, arcname=filename)
+
+        self._distances = _distances
+
+    @distances.deleter
+    def distances(self):
+        self._distances = None
+
+    @property
+    def smoothed_distances(self):
+        if self._smoothed_distances is None:
+            with zipfile.ZipFile(self.filename, mode="r") as zf:
+                filename = self._numpy_filename("smoothed_distances")
+                if filename in zf.namelist():
+                    with tempfile.TemporaryDirectory() as tmp_dir:
+                        smoothedfile = zf.extract(filename, path=tmp_dir)
+                        read_func = (
+                            (lambda f: np.load(f)["arr_0"])
+                            if self.compress
+                            else np.load
+                        )
+                        self._smoothed_distances = read_func(smoothedfile)
+
+        return self._smoothed_distances
+
+    @smoothed_distances.setter
+    def smoothed_distances(self, _smoothed_distances):
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            filename = self._numpy_filename("smoothed_distances")
+            write_func = np.savez_compressed if self.compress else np.save
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                smoothedfile = os.path.join(tmp_dir, filename)
+                write_func(smoothedfile, _smoothed_distances)
+                zf.write(smoothedfile, arcname=filename)
+
+        self._smoothed_distances = _smoothed_distances
+
+    @smoothed_distances.deleter
+    def smoothed_distances(self):
+        self._smoothed_distances = None
+
+    @property
+    def uncertainties(self):
+        """Access the M3C2 uncertainties of this analysis"""
+
+        if self._uncertainties is None:
+            with zipfile.ZipFile(self.filename, mode="r") as zf:
+                filename = self._numpy_filename("uncertainties")
+                if filename not in zf.namelist():
+                    self.uncertainties = np.empty(
+                        (self.corepoints.cloud.shape[0], 0),
+                        dtype=np.dtype(
+                            [
+                                ("lodetection", "<f8"),
+                                ("spread1", "<f8"),
+                                ("num_samples1", "<i8"),
+                                ("spread2", "<f8"),
+                                ("num_samples2", "<i8"),
+                            ]
+                        ),
+                    )
+                    return self._uncertainties
+
+                with tempfile.TemporaryDirectory() as tmp_dir:
+                    uncertaintyfile = zf.extract(filename, path=tmp_dir)
+                    read_func = (
+                        (lambda f: np.load(f)["arr_0"]) if self.compress else np.load
+                    )
+                    self._uncertainties = read_func(uncertaintyfile)
+
+        return self._uncertainties
+
+    @uncertainties.setter
+    def uncertainties(self, _uncertainties):
+        """Set the uncertainties manually
+
+        This is only possible exactly once and mutually exclusive with adding
+        epochs via the :ref:`add_epochs` method.
+        """
+        with zipfile.ZipFile(self.filename, mode="a") as zf:
+            filename = self._numpy_filename("uncertainties")
+            write_func = np.savez_compressed if self.compress else np.save
+
+            # If we already have distacces in the archive, this is not possible
+            if filename in zf.namelist():
+                raise Py4DGeoError(
+                    "Uncertainties can only be set on freshly created analysis instances, use add_epochs instead."
+                )
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                uncertaintiesfile = os.path.join(tmp_dir, filename)
+                write_func(uncertaintiesfile, _uncertainties)
+                zf.write(uncertaintiesfile, arcname=filename)
+
+        self._uncertainties = _uncertainties
+
+    @uncertainties.deleter
+    def uncertainties(self):
+        self._uncertainties = None
+
+    def add_epochs(self, *epochs):
+        """Add a numbers of epochs to the existing analysis"""
+
+        # Remove intermediate results from the archive
+        self.invalidate_results()
+
+        # Assert that all epochs have a timestamp
+        for epoch in epochs:
+            check_epoch_timestamp(epoch)
+
+        # Lazily fetch required data
+        reference_epoch = self.reference_epoch
+        timedeltas = self.timedeltas
+
+        # Collect the calculated results to only add them once to the archive
+        new_distances = []
+        new_uncertainties = []
+
+        # Iterate over the given epochs
+        for i, epoch in enumerate(sorted(epochs, key=lambda e: e.timestamp)):
+            with logger_context(f"Adding epoch {i+1}/{len(epochs)} to analysis object"):
+                # Prepare the M3C2 instance
+                self.m3c2.corepoints = self.corepoints.cloud
+                self.m3c2.epochs = (reference_epoch, epoch)
+
+                # Calculate the M3C2 distances
+                d, u = self.m3c2.calculate_distances(reference_epoch, epoch)
+                new_distances.append(d)
+                new_uncertainties.append(u)
+                timedeltas.append(epoch.timestamp - reference_epoch.timestamp)
+
+        # We do not need the reference_epoch at this point
+        del self.reference_epoch
+
+        # Prepare all archive data in a temporary directory
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            # Write a new timestamps file
+            timestampsfile = os.path.join(tmp_dir, "timestamps.json")
+            with open(timestampsfile, "w") as f:
+                json.dump(
+                    [
+                        {
+                            "days": td.days,
+                            "seconds": td.seconds,
+                            "microseconds": td.microseconds,
+                        }
+                        for td in timedeltas
+                    ],
+                    f,
+                )
+
+            # Depending on whether we compress, we use different numpy functionality
+            write_func = np.savez_compressed if self.compress else np.save
+            distance_filename = self._numpy_filename("distances")
+            uncertainty_filename = self._numpy_filename("uncertainties")
+
+            with logger_context("Rearranging space-time array in memory"):
+                # Load the distance array and append new data
+                distance_file = os.path.join(tmp_dir, distance_filename)
+                write_func(
+                    distance_file,
+                    np.concatenate(
+                        (self.distances, np.column_stack(tuple(new_distances))), axis=1
+                    ),
+                )
+
+                # Load the uncertainty array and append new data
+                uncertainty_file = os.path.join(tmp_dir, uncertainty_filename)
+                write_func(
+                    uncertainty_file,
+                    np.concatenate(
+                        (self.uncertainties, np.column_stack(tuple(new_uncertainties))),
+                        axis=1,
+                    ),
+                )
+
+            # Invalidate potential caches for distances/uncertainties
+            self._distances = None
+            self._uncertainties = None
+
+            # Dump the updated files into the archive
+            with logger_context("Updating disk-based analysis archive with new epochs"):
+                with UpdateableZipFile(self.filename, mode="a") as zf:
+                    if "timestamps.json" in zf.namelist():
+                        zf.remove("timestamps.json")
+                    zf.write(timestampsfile, arcname="timestamps.json")
+                    if distance_filename in zf.namelist():
+                        zf.remove(distance_filename)
+                    zf.write(distance_file, arcname=distance_filename)
+                    if uncertainty_filename in zf.namelist():
+                        zf.remove(uncertainty_filename)
+                    zf.write(uncertainty_file, arcname=uncertainty_filename)
+
+        # (Potentially) remove caches
+        del self.distances
+        del self.uncertainties
+
+    @property
+    def seeds(self):
+        """The list of seed candidates for this analysis"""
+
+        with zipfile.ZipFile(self.filename, mode="r") as zf:
+            if "seeds.pickle" not in zf.namelist():
+                return None
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                zf.extract("seeds.pickle", path=tmp_dir)
+                with open(os.path.join(tmp_dir, "seeds.pickle"), "rb") as f:
+                    return pickle.load(f)
+
+    @seeds.setter
+    def seeds(self, _seeds):
+        # Assert that we received the correct type
+        for seed in _seeds:
+            if not isinstance(seed, RegionGrowingSeed):
+                raise Py4DGeoError(
+                    "Seeds are expected to inherit from RegionGrowingSeed"
+                )
+
+        if not self.allow_pickle:
+            return
+
+        with UpdateableZipFile(self.filename, mode="a") as zf:
+            if "seeds.pickle" in zf.namelist():
+                zf.remove("seeds.pickle")
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                seedsfile = os.path.join(tmp_dir, "seeds.pickle")
+                with open(seedsfile, "wb") as f:
+                    pickle.dump(_seeds, f)
+
+                zf.write(seedsfile, arcname="seeds.pickle")
+
+    @property
+    def objects(self):
+        """The list of objects by change for this analysis"""
+
+        with zipfile.ZipFile(self.filename, mode="r") as zf:
+            if "objects.pickle" not in zf.namelist():
+                return None
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                zf.extract("objects.pickle", path=tmp_dir)
+                with open(os.path.join(tmp_dir, "objects.pickle"), "rb") as f:
+                    return pickle.load(f)
+
+    @objects.setter
+    def objects(self, _objects):
+        # Assert that we received the correct type
+        for seed in _objects:
+            if not isinstance(seed, ObjectByChange):
+                raise Py4DGeoError(
+                    "Objects are expected to inherit from ObjectByChange"
+                )
+
+        if not self.allow_pickle:
+            return
+
+        with UpdateableZipFile(self.filename, mode="a") as zf:
+            if "objects.pickle" in zf.namelist():
+                zf.remove("objects.pickle")
+
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                objectsfile = os.path.join(tmp_dir, "objects.pickle")
+                with open(objectsfile, "wb") as f:
+                    pickle.dump(_objects, f)
+
+                zf.write(objectsfile, arcname="objects.pickle")
+
+    def invalidate_results(self, seeds=True, objects=True, smoothed_distances=True):
+        """Invalidate (and remove) calculated results
+
+        This is automatically called when new epochs are added or when
+        an algorithm sets the :code:`force` option.
+        """
+
+        logger.info(
+            f"Removing intermediate results from the analysis file {self.filename}"
+        )
+        with UpdateableZipFile(self.filename, mode="a") as zf:
+            if seeds and "seeds.pickle" in zf.namelist():
+                zf.remove("seeds.pickle")
+
+            if objects and "objects.pickle" in zf.namelist():
+                zf.remove("objects.pickle")
+
+            smoothed_file = self._numpy_filename("smoothed_distances")
+            if smoothed_distances and smoothed_file in zf.namelist():
+                zf.remove(smoothed_file)
+
+    def _numpy_filename(self, name):
+        extension = "npz" if self.compress else "npy"
+        return f"{name}.{extension}"
+
+    @property
+    def distances_for_compute(self):
+        """Retrieve the distance array used for computation
+
+        This might be the raw data or smoothed data, based on whether
+        a smoothing was provided by the user.
+        """
+        distances = self.smoothed_distances
+        if distances is None:
+            distances = self.distances
+        return distances
+
+
+class RegionGrowingAlgorithmBase:
+    def __init__(
+        self,
+        neighborhood_radius=1.0,
+        thresholds=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],
+        min_segments=20,
+        max_segments=None,
+    ):
+        """Construct a spatiotemporal _segmentation algorithm.
+
+        This class can be derived from to customize the algorithm behaviour.
+
+        :param neighborhood_radius:
+            The size of the neighborhood of a core point. All core points within
+            this radius are considered adjacent and are therefore considered as
+            candidates for inclusion in the region growing algorithm.
+        :type neighborhood_radius: float
+        :param thresholds:
+            A list of thresholds to use as candidates in 4D-OBC's adaptive
+            thresholding procedure.
+        :type thresholds: list
+        :param min_segments:
+            The minimum number of core points in an object-by-change. Defaults to
+            20.
+        :type min_segments: int
+        :param max_segments:
+            The maximum number of core points in an object-by-change. This is mainly
+            used to bound the runtime of expensive region growing. By default, no
+            maximum is applied.
+        :type max_segments: int
+        """
+
+        self.neighborhood_radius = neighborhood_radius
+        self.thresholds = thresholds
+        self.min_segments = min_segments
+        self.max_segments = max_segments
+
+        self._analysis = None
+
+    def distance_measure(self):
+        """Distance measure between two time series
+
+        Expected to return a function that accepts two time series and returns
+        the distance.
+        """
+
+        return _py4dgeo.normalized_dtw_distance
+
+    def find_seedpoints(self):
+        """Calculate seedpoints for the region growing algorithm"""
+
+        raise NotImplementedError
+
+    def seed_sorting_scorefunction(self):
+        """A function that computes a score for a seed candidate
+
+        This function is used to prioritize seed candidates.
+        """
+
+        # The base class does not perform sorting.
+        return lambda seed: 0.0
+
+    def filter_objects(self, obj):
+        """A filter for objects produced by the region growing algorithm
+
+        Objects are discarded if this method returns False.
+        """
+
+        # The base class does not perform filtering
+        return True
+
+    @property
+    def analysis(self):
+        """Access the analysis object that the algorithm operates on
+
+        This is only available after :ref:`run` has been called.
+        """
+        if self._analysis is None:
+            raise Py4DGeoError(
+                "Analysis object is only available when the algorithm is run"
+            )
+        return self._analysis
+
+    def run(self, analysis, force=False):
+        """Calculate the _segmentation
+
+        :param analysis:
+            The analysis object we are working with.
+        :type analysis: py4dgeo.segmentation.SpatiotemporalAnalysis
+        :param force:
+            Force recalculation of results. If false, some intermediate results will be
+            restored from the analysis object instead of being recalculated.
+        """
+
+        # Make the analysis object known to all members
+        self._analysis = analysis
+
+        # Enforce the removal of intermediate results
+        if force:
+            analysis.invalidate_results()
+
+        # Return pre-calculated objects if they are available
+        precalculated = analysis.objects
+        if precalculated is not None:
+            logger.info("Reusing objects by change stored in analysis object")
+            return precalculated
+
+        # Get corepoints from M3C2 class and build a KDTree on them
+        corepoints = as_epoch(analysis.corepoints)
+        corepoints.build_kdtree()
+
+        # Calculate the list of seed points and sort them
+        seeds = analysis.seeds
+        if seeds is None:
+            with logger_context("Find seed candidates in time series"):
+                seeds = self.find_seedpoints()
+
+            # Sort the seed points
+            with logger_context("Sort seed candidates by priority"):
+                seeds = list(sorted(seeds, key=self.seed_sorting_scorefunction()))
+
+            # Store the seeds
+            analysis.seeds = seeds
+        else:
+            logger.info("Reusing seed candidates stored in analysis object")
+
+        objects = []
+
+        # Iterate over the seeds to maybe turn them into objects
+        for i, seed in enumerate(seeds):
+            # Check all already calculated objects whether they overlap with this seed.
+            found = False
+            for obj in objects:
+                if seed.index in obj.indices and (
+                    obj.end_epoch > seed.start_epoch
+                    and seed.end_epoch > obj.start_epoch
+                ):
+                    found = True
+                    break
+
+            # If we found an overlap, we skip this seed
+            if found:
+                continue
+
+            # Apply a numeric default to the max_segments parameter
+            max_segments = self.max_segments
+            if max_segments is None:
+                max_segments = corepoints.cloud.shape[0] + 1
+
+            data = _py4dgeo.RegionGrowingAlgorithmData(
+                analysis.distances_for_compute,
+                corepoints,
+                self.neighborhood_radius,
+                seed._seed,
+                self.thresholds,
+                self.min_segments,
+                max_segments,
+            )
+
+            # Perform the region growing
+            with logger_context(
+                f"Performing region growing on seed candidate {i+1}/{len(seeds)}"
+            ):
+                objdata = _py4dgeo.region_growing(data, self.distance_measure())
+
+                # If the returned object has 0 indices, the min_segments threshold was violated
+                if objdata.indices_distances:
+                    obj = ObjectByChange(objdata, seed, analysis)
+                    if self.filter_objects(obj):
+                        objects.append(obj)
+
+                # If the returned object is larger than max_segments we issue a warning
+                if len(objdata.indices_distances) >= max_segments:
+                    logger.warning(
+                        f"An object by change exceeded the given maximum size of {max_segments}"
+                    )
+
+        # Store the results in the analysis object
+        analysis.objects = objects
+
+        # Potentially remove objects from memory # TODO Why do we remove these?
+        del analysis.smoothed_distances
+        del analysis.distances
+
+        return objects
+
+
+class RegionGrowingAlgorithm(RegionGrowingAlgorithmBase):
+    def __init__(
+        self,
+        seed_subsampling=1,
+        seed_candidates=None,
+        window_width=24,
+        window_min_size=12,
+        window_jump=1,
+        window_penalty=1.0,
+        minperiod=24,
+        height_threshold=0.0,
+        **kwargs,
+    ):
+        """Construct the 4D-OBC algorithm.
+
+        :param seed_subsampling:
+            A subsampling factor for the set of core points for the generation
+            of _segmentation seed candidates. This can be used to speed up
+            the generation of seeds. The default of 1 does not perform any
+            subsampling, a value of, e.g., 10 would only consider every 10th
+            corepoint for adding seeds.
+        :type seed_subsampling: int
+        :param seed_candidates:
+            A set of indices specifying which core points should be used for seed detection. This can be used to perform _segmentation for selected locations. The default of None does not perform any selection and uses all corepoints. The subsampling parameter is applied additionally.
+        :type seed_candidates: list
+        :param window_width:
+            The width of the sliding temporal window for change point detection. The sliding window
+            moves along the signal and determines the discrepancy between the first and the second
+            half of the window (i.e. subsequent time series segments within the window width). The
+            default value is 24, corresponding to one day in case of hourly data.
+        :type window_width: int
+        :param window_min_size:
+            The minimum temporal distance needed between two seed candidates, for the second one to be considered.
+            The default value is 1, such that all detected seeds candidates are considered.
+        :type window_min_size: int
+        :param window_jump:
+            The interval on which the sliding temporal window moves and checks for seed candidates.
+            The default value is 1, corresponding to a check for every epoch in the time series.
+        :type window_jump: int
+        :param window_penalty:
+            A complexity penalty that determines how strict the change point detection is.
+            A higher penalty results in stricter change point detection (i.e, fewer points are detected), while a low
+            value results in a large amount of detected change points. The default value is 1.0.
+        :type window_penalty: float
+        :param minperiod:
+            The minimum period of a detected change to be considered as seed candidate for subsequent
+            _segmentation. The default is 24, corresponding to one day for hourly data.
+        :type minperiod: int
+        :param height_threshold:
+            The height threshold represents the required magnitude of a detected change to be considered
+            as seed candidate for subsequent _segmentation. The magnitude of a detected change is derived
+            as unsigned difference between magnitude (i.e. distance) at start epoch and peak magnitude.
+            The default is 0.0, in which case all detected changes are used as seed candidates.
+        :type height_threshold: float
+
+        """
+
+        # Initialize base class
+        super().__init__(**kwargs)
+
+        # Store the given parameters
+        self.seed_subsampling = seed_subsampling
+        self.seed_candidates = seed_candidates
+        self.window_width = window_width
+        self.window_min_size = window_min_size
+        self.window_jump = window_jump
+        self.window_penalty = window_penalty
+        self.minperiod = minperiod
+        self.height_threshold = height_threshold
+
+    def find_seedpoints(self):
+        """Calculate seedpoints for the region growing algorithm"""
+
+        # These are some arguments used below that we might consider
+        # exposing to the user in the future. For now, they are considered
+        # internal, but they are still defined here for readability.
+        window_costmodel = "l1"
+        # window_min_size = 12
+        # window_jump = 1
+        # window_penalty = 1.0
+
+        # The list of generated seeds
+        seeds = []
+
+        # The list of core point indices to check as seeds
+        if self.seed_candidates is None:
+            if self.seed_subsampling == 0:
+                raise Py4DGeoError(
+                    "Subsampling factor cannot be 0, use 1 or any integer larger than 1"
+                )
+            # Use all corepoints if no selection specified, considering subsampling
+            seed_candidates_curr = range(
+                0, self.analysis.distances_for_compute.shape[0], self.seed_subsampling
+            )
+        else:
+            # Use the specified corepoint indices, but consider subsampling
+            seed_candidates_curr = self.seed_candidates  # [::self.seed_subsampling]
+
+        # Iterate over all time series to analyse their change points
+        for i in seed_candidates_curr:
+            # Extract the time series and interpolate its nan values
+            timeseries = self.analysis.distances_for_compute[i, :]
+            bad_indices = np.isnan(timeseries)
+            num_nans = np.count_nonzero(bad_indices)
+
+            # If we too many nans, this timeseries does not make sense
+            if num_nans > timeseries.shape[0] - 3:
+                continue
+
+            # If there are nan values, we try fixing things by interpolation
+            if num_nans > 0:
+                good_indices = np.logical_not(bad_indices)
+                timeseries[bad_indices] = np.interp(
+                    bad_indices.nonzero()[0],
+                    good_indices.nonzero()[0],
+                    timeseries[good_indices],
+                )
+
+            # Run detection of change points
+            cpdata = _py4dgeo.ChangePointDetectionData(
+                ts=timeseries,
+                window_size=self.window_width,
+                min_size=self.window_min_size,
+                jump=self.window_jump,
+                penalty=self.window_penalty,
+            )
+            changepoints = _py4dgeo.change_point_detection(cpdata)[:-1]
+
+            # Shift the time series to positive values
+            timeseries = timeseries + abs(np.nanmin(timeseries) + 0.1)
+            # create a flipped version for negative change volumes
+            timeseries_flipped = timeseries * -1.0 + abs(np.nanmax(timeseries)) + 0.1
+
+            # Create seeds for this timeseries
+            corepoint_seeds = []
+            for start_idx in changepoints:
+                # Skip this changepoint if it was included into a previous seed
+                if corepoint_seeds and start_idx <= corepoint_seeds[-1].end_epoch:
+                    continue
+
+                # Skip this changepoint if this to close to the end
+                if start_idx >= timeseries.shape[0] - self.minperiod:
+                    break
+
+                # Decide whether we need use the flipped timeseries
+                used_timeseries = timeseries
+                if timeseries[start_idx] >= timeseries[start_idx + self.minperiod]:
+                    used_timeseries = timeseries_flipped
+
+                previous_volume = -999.9
+                for target_idx in range(start_idx + 1, timeseries.shape[0]):
+                    # Calculate the change volume
+                    height = used_timeseries[start_idx]
+                    volume = np.nansum(
+                        used_timeseries[start_idx : target_idx + 1] - height
+                    )
+
+                    # Check whether the volume started decreasing
+                    if previous_volume > volume:
+                        # Only add seed if larger than the minimum period
+                        if target_idx - start_idx >= self.minperiod:
+                            corepoint_seeds.append(
+                                RegionGrowingSeed(i, start_idx, target_idx)
+                            )
+                        break
+                    else:
+                        previous_volume = volume
+
+                    # This causes a seed to always be detected if the volume doesn't decrease before present
+                    #  Useful when used in an online setting, can be filtered before region growing
+                    # Only if the last epoch is reached we use the segment as seed
+                    if target_idx == timeseries.shape[0] - 1:
+                        # We reached the present and add a seed based on it
+                        corepoint_seeds.append(
+                            RegionGrowingSeed(i, start_idx, timeseries.shape[0] - 1)
+                        )
+
+            # Add all the seeds found for this corepoint to the full list
+            seeds.extend(corepoint_seeds)
+
+        return seeds
+
+    def seed_sorting_scorefunction(self):
+        """Neighborhood similarity sorting function"""
+
+        # The 4D-OBC algorithm sorts by similarity in the neighborhood
+        # of the seed.
+        def neighborhood_similarity(seed):
+            neighbors = self.analysis.corepoints.kdtree.radius_search(
+                self.analysis.corepoints.cloud[seed.index, :], self.neighborhood_radius
+            )
+            similarities = []
+            for n in neighbors:
+                data = _py4dgeo.TimeseriesDistanceFunctionData(
+                    self.analysis.distances_for_compute[
+                        seed.index, seed.start_epoch : seed.end_epoch + 1
+                    ],
+                    self.analysis.distances_for_compute[
+                        n, seed.start_epoch : seed.end_epoch + 1
+                    ],
+                )
+                similarities.append(self.distance_measure()(data))
+
+            return sum(similarities, 0.0) / (len(neighbors) - 1)
+
+        return neighborhood_similarity
+
+    def filter_objects(self, obj):
+        """A filter for objects produced by the region growing algorithm"""
+
+        # Filter based on coefficient of variation
+        distarray = np.fromiter(obj._data.indices_distances.values(), np.float64)
+
+        # Check if mean is 0.0, if so, set to very small value to avoid division by 0
+        mean_distarray = np.mean(distarray)
+        if mean_distarray == 0.0:
+            mean_distarray = 10**-10
+
+        # Calculate coefficient of variation
+        cv = np.std(distarray) / mean_distarray
+
+        # TODO: Make this threshold configurable?
+        return cv <= 0.8
+
+
+class RegionGrowingSeed:
+    def __init__(self, index, start_epoch, end_epoch):
+        self._seed = _py4dgeo.RegionGrowingSeed(index, start_epoch, end_epoch)
+
+    @property
+    def index(self):
+        return self._seed.index
+
+    @property
+    def start_epoch(self):
+        return self._seed.start_epoch
+
+    @property
+    def end_epoch(self):
+        return self._seed.end_epoch
+
+
+class ObjectByChange:
+    """Representation a change object in the spatiotemporal domain"""
+
+    def __init__(self, data, seed, analysis=None):
+        self._data = data
+        self._analysis = analysis
+        self.seed = seed
+
+    @property
+    def indices(self):
+        """The set of corepoint indices that compose the object by change"""
+        return list(self._data.indices_distances.keys())
+
+    def distance(self, index):
+        return self._data.indices_distances[index]
+
+    @property
+    def start_epoch(self):
+        """The index of the start epoch of the change object"""
+        return self._data.start_epoch
+
+    @property
+    def end_epoch(self):
+        """The index of the end epoch of the change object"""
+        return self._data.end_epoch
+
+    @property
+    def threshold(self):
+        """The distance threshold that produced this object"""
+        return self._data.threshold
+
+    def plot(self, filename=None):
+        """Create an informative visualization of the Object By Change
+
+        :param filename:
+            The filename to use to store the plot. Can be omitted to only show
+            plot in a Jupyter notebook session.
+        :type filename: str
+        """
+
+        # Extract DTW distances from this object
+        indexarray = np.fromiter(self.indices, np.int32)
+        distarray = np.fromiter((self.distance(i) for i in indexarray), np.float64)
+
+        # Intitialize the figure and all of its subfigures
+        fig = plt.figure(figsize=plt.figaspect(0.3))
+        tsax = fig.add_subplot(1, 3, 1)
+        histax = fig.add_subplot(1, 3, 2)
+        mapax = fig.add_subplot(1, 3, 3)
+
+        # The first plot (tsax) prints all time series of chosen corepoints
+        # and colors them according to distance.
+        tsax.set_ylabel("Height change [m]")
+        tsax.set_xlabel("Time [h]")
+
+        # We pad the time series visualization with a number of data
+        # points on both sides. TODO: Expose as argument to plot?
+        timeseries_padding = 10
+        start_epoch = max(self.start_epoch - timeseries_padding, 0)
+        end_epoch = min(
+            self.end_epoch + timeseries_padding,
+            self._analysis.distances_for_compute.shape[1],
+        )
+
+        # We use the seed's timeseries to set good axis limits
+        seed_ts = self._analysis.distances_for_compute[
+            self.seed.index, start_epoch:end_epoch
+        ]
+        tsax.set_ylim(np.nanmin(seed_ts) * 0.5, np.nanmax(seed_ts) * 1.5)
+
+        # Create a colormap with distance for this object
+        cmap = matplotlib.colormaps.get_cmap("viridis")
+        maxdist = np.nanmax(distarray)
+
+        # Plot each time series individually
+        for index in self.indices:
+            tsax.plot(
+                self._analysis.distances_for_compute[index, start_epoch:end_epoch],
+                linewidth=0.7,
+                alpha=0.3,
+                color=cmap(self.distance(index) / maxdist),
+            )
+
+        # Plot the seed timeseries again, but with a thicker line
+        tsax.plot(seed_ts, linewidth=2.0, zorder=10, color="blue")
+
+        # Next, we add a histogram plot with the distance values (using seaborn)
+        seaborn.histplot(distarray, ax=histax, kde=True, color="r")
+
+        # Add labels to the histogram plot
+        histax.set_title(f"Segment size: {distarray.shape[0]}")
+        histax.set_xlabel("DTW distance")
+
+        # Create a 2D view of the segment
+        locations = self._analysis.corepoints.cloud[indexarray, 0:2]
+        mapax.scatter(locations[:, 0], locations[:, 1], c=distarray)
+
+        # Some global settings of the generated figure
+        fig.tight_layout()
+
+        # Maybe save to file
+        if filename is not None:
+            plt.savefig(filename)
+
+
+def check_epoch_timestamp(epoch):
+    """Validate an epoch to be used with SpatiotemporalSegmentation"""
+    if epoch.timestamp is None:
+        raise Py4DGeoError(
+            "Epochs need to define a timestamp to be usable in SpatiotemporalSegmentation"
+        )
+
+    return epoch
+
+
+def regular_corepoint_grid(lowerleft, upperright, num_points, zval=0.0):
+    """A helper function to create a regularly spaced grid for the analysis
+
+    :param lowerleft:
+        The lower left corner of the grid. Given as a 2D coordinate.
+    :type lowerleft: np.ndarray
+    :param upperright:
+        The upper right corner of the grid. Given as a 2D coordinate.
+    :type upperright: np.ndarray
+    :param num_points:
+        A tuple with two entries denoting the number of points to be used in
+        x and y direction
+    :type num_points: tuple
+    :param zval:
+        The value to fill for the z-direction.
+    :type zval: double
+    """
+    xspace = np.linspace(
+        lowerleft[0], upperright[0], num=num_points[0], dtype=np.float64
+    )
+    yspace = np.linspace(
+        lowerleft[1], upperright[1], num=num_points[1], dtype=np.float64
+    )
+
+    grid = np.empty(shape=(num_points[0] * num_points[1], 3), dtype=np.float64)
+    for i, x in enumerate(xspace):
+        for j, y in enumerate(yspace):
+            grid[i * num_points[0] + j, 0] = x
+            grid[i * num_points[0] + j, 1] = y
+            grid[i * num_points[0] + j, 2] = zval
+
+    return grid
+
+
+def temporal_averaging(distances, smoothing_window=24):
+    """Smoothen a space-time array of distance change using a sliding window approach
+
+    :param distances:
+        The raw data to smoothen.
+    :type distances: np.ndarray
+    :param smoothing_window:
+        The size of the sliding window used in smoothing the data. The
+        default value of 0 does not perform any smooting.
+    :type smooting_window: int
+    """
+
+    with logger_context("Smoothing temporal data"):
+        smoothed = np.empty_like(distances)
+        eps = smoothing_window // 2
+
+        for i in range(distances.shape[1]):
+            smoothed[:, i] = np.nanmedian(
+                distances[
+                    :,
+                    max(0, i - eps) : min(distances.shape[1] - 1, i + eps),
+                ],
+                axis=1,
+            )
+
+        # We use no-op smooting as the default implementation here
+        return smoothed