OBIABase.py

import os.path
import sys
import rasterio as rio
from rasterio.windows import Window
from rasterio.features import shapes
import numpy as np
import geopandas as gpd
import pandas as pd
from enum import Enum
from Common.otb_numpy_proc import to_otb_pipeline
from skimage.measure import regionprops, label
import otbApplication as otb
from tqdm import tqdm
from sklearn.impute import SimpleImputer

class OStats(Enum):
    MEAN = 'mean'
    STD = 'std'
    MIN = 'min'
    MAX = 'max'

def intensity_std(region, intensities):
    # note the ddof arg to get the sample var if you so desire!
    return np.std(intensities[region])

class OBIABase:

    def __init__(self, object_layer,
                 ref_data=None, ref_id_field='id', ref_class_field='class',
                 nominal_tile_size=5000, background_label=0):
        self.object_layer = object_layer
        with rio.open(self.object_layer, 'r') as src:
            self.crs = src.crs
            self.transform = src.transform
        self.background_label = background_label
        if type(nominal_tile_size) is tuple:
            self.nominal_tile_size = nominal_tile_size
        elif type(nominal_tile_size) is int:
            self.nominal_tile_size = (nominal_tile_size, nominal_tile_size)
        else:
            sys.exit('Invalid nominal tile size')
        self.tiles, self.obj_to_tile = self.generate_tile_info(self.nominal_tile_size, self.background_label)
        # Init database
        self.ref_db = None
        self.ref_obj_layer_pipe = None
        if ref_data is not None:
            assert(ref_id_field is not None and ref_class_field is not None)
            self.init_ref_db(ref_data, ref_id_field, ref_class_field)

        # Computation rasters
        self.raster_files = []
        self.raster_stats = []
        self.raster_var_names = []
        self.raster_groups = []
        self.n_vars = 0

        # Output raster
        self.output_maps = []

    def init_ref_db(self, vector_file, id_field, class_field):
        if isinstance(class_field, str):
            class_field = [class_field]
        ras_id = otb.Registry.CreateApplication('Rasterization')
        ras_id.SetParameterString('in', vector_file)
        ras_id.SetParameterString('im', self.object_layer)
        ras_id.SetParameterString('mode', 'attribute')
        ras_id.SetParameterString('mode.attribute.field', id_field)
        ras_id.Execute()
        #ids = ras_id.GetImageAsNumpyArray('out')

        ras_cl = []
        for cf in class_field:
            ras_cl.append(otb.Registry.CreateApplication('Rasterization'))
            ras_cl[-1].SetParameterString('in', vector_file)
            ras_cl[-1].SetParameterString('im', self.object_layer)
            ras_cl[-1].SetParameterString('mode', 'attribute')
            ras_cl[-1].SetParameterString('mode.attribute.field', cf)
            ras_cl[-1].Execute()

        in_seg = to_otb_pipeline(self.object_layer)

        intensity_img = otb.Registry.CreateApplication('ConcatenateImages')
        intensity_img.AddImageToParameterInputImageList('il', in_seg.GetParameterOutputImage('out'))
        intensity_img.AddImageToParameterInputImageList('il', ras_id.GetParameterOutputImage('out'))
        [intensity_img.AddImageToParameterInputImageList('il', rcl.GetParameterOutputImage('out')) for rcl in ras_cl]
        intensity_img.Execute()

        ref_ol = otb.Registry.CreateApplication('BandMath')
        ref_ol.AddImageToParameterInputImageList('il', in_seg.GetParameterOutputImage('out'))
        ref_ol.AddImageToParameterInputImageList('il', ras_id.GetParameterOutputImage('out'))
        ref_ol.SetParameterString('exp', 'im1b1 * (im2b1 > 0)')
        ref_ol.Execute()
        self.ref_obj_layer_pipe = [in_seg, ras_id, ref_ol]

        self.ref_db = pd.DataFrame(data=[],
                                   columns=['area', 'orig_label', 'polygon_id'] + class_field,
                                   index=[])
        r = otb.itkRegion()
        for tn, t in tqdm(self.tiles.items(), desc='Init. Ref. DB', total=len(self.tiles)):
            r['index'][0], r['index'][1] = t[0], t[1]
            r['size'][0], r['size'][1] = t[2], t[3]
            ref_ol.PropagateRequestedRegion('out', r)
            intensity_img.PropagateRequestedRegion('out', r)
            tile_ref_ol = ref_ol.GetImageAsNumpyArray('out').astype(np.int32)
            tile_int_img = intensity_img.GetVectorImageAsNumpyArray('out').astype(int)
            rp = regionprops(tile_ref_ol, intensity_image=tile_int_img)
            self.ref_db = pd.concat([
                self.ref_db,
                pd.DataFrame(
                    data=[np.insert(o.intensity_min, 0, o.area) for o in rp if self.obj_to_tile[o.label] == tn],
                    columns=self.ref_db.columns,
                    index=[o.label for o in rp if self.obj_to_tile[o.label] == tn]
                )]
            )

        return

    def generate_tile_info(self, nominal_tile_size, background_label):
        in_seg = to_otb_pipeline(self.object_layer)
        self.W, self.H = in_seg.GetImageSize('out')
        r = otb.itkRegion()
        obj_bbox = {}
        rw, rh = range(0, self.W, nominal_tile_size[0]), range(0, self.H, nominal_tile_size[1])
        with tqdm(total=len(rw)*len(rh), desc='Gen. tile info') as pb:
            for i in rw:
                for j in rh:
                    r['index'][0], r['index'][1] = i, j
                    r['size'][0], r['size'][1] = min(nominal_tile_size[0], self.W-i), min(nominal_tile_size[1], self.H-j)
                    in_seg.PropagateRequestedRegion('out', r)
                    tile = in_seg.GetImageAsNumpyArray('out')
                    rp = regionprops(tile.astype(np.uint32))
                    for o in rp:
                        bbox = np.array([o.bbox[0]+j, o.bbox[1]+i, o.bbox[2]+j, o.bbox[3]+i])
                        if o.label in obj_bbox.keys():
                            obj_bbox[o.label].append(bbox)
                        else:
                            obj_bbox[o.label] = [bbox]
                    pb.update(1)

        for o in obj_bbox.keys():
            obj_bbox[o] = np.array(obj_bbox[o])
            mn, mx = np.min(obj_bbox[o], axis=0), np.max(obj_bbox[o], axis=0)
            obj_bbox[o] = [mn[0], mn[1], mx[2], mx[3]]

        tx, ty = int(self.W / nominal_tile_size[0]) + 1, int(self.H / nominal_tile_size[1]) + 1

        obj_to_tile = {}
        tiles = dict.fromkeys(range(tx * ty))
        for i in range(tx * ty):
            tiles[i] = [np.inf, np.inf, 0, 0]

        for o,bb in obj_bbox.items():
            if o != background_label:
                ix, iy = int(bb[1] / nominal_tile_size[0]), int(bb[0] / nominal_tile_size[1])
                idx = ix * ty + iy
                obj_to_tile[o] = idx
                tiles[idx][0] = int(min(bb[1], tiles[idx][0]))
                tiles[idx][1] = int(min(bb[0], tiles[idx][1]))
                tiles[idx][2] = int(max(bb[3], tiles[idx][2]))
                tiles[idx][3] = int(max(bb[2], tiles[idx][3]))

        tiles = {k: [v[0],v[1],v[2]-v[0],v[3]-v[1]] for k,v in tiles.items() if not np.isinf(v[0])}

        return tiles, obj_to_tile

    def compute_stats_on_tile(self, tile_num, input_image, stats=[OStats.MEAN], on_ref=False):
        r = otb.itkRegion()
        r['index'][0], r['index'][1] = self.tiles[tile_num][0], self.tiles[tile_num][1]
        r['size'][0], r['size'][1] = self.tiles[tile_num][2], self.tiles[tile_num][3]

        if not on_ref:
            obj = to_otb_pipeline(self.object_layer)
            obj.PropagateRequestedRegion('out', r)
            clip_obj = obj.ExportImage('out')
            tile_obj = np.squeeze(clip_obj['array']).astype(np.uint32)
        else:
            assert (self.ref_obj_layer_pipe is not None)
            tmp_er = otb.Registry.CreateApplication('ExtractROI')
            tmp_er.SetParameterInputImage('in', self.ref_obj_layer_pipe[-1].GetParameterOutputImage('out'))
            tmp_er.SetParameterInt('startx', r['index'][0])
            tmp_er.SetParameterInt('starty', r['index'][1])
            tmp_er.SetParameterInt('sizex', r['size'][0])
            tmp_er.SetParameterInt('sizey', r['size'][1])
            tmp_er.Execute()
            tile_obj = tmp_er.GetImageAsNumpyArray('out').astype(np.uint32)

        si = otb.Registry.CreateApplication('Superimpose')
        si.SetParameterString('inm', input_image)
        si.SetParameterString('inr', self.object_layer)
        si.SetParameterString('interpolator', 'nn')
        si.Execute()
        si.PropagateRequestedRegion('out', r)
        clip_img = si.ExportImage('out')
        tile_img = np.squeeze(clip_img['array'])

        if OStats.STD in stats:
            op = regionprops(tile_obj, intensity_image=tile_img, extra_properties=[intensity_std])
        else:
            op = regionprops(tile_obj, intensity_image=tile_img)

        out_stats = {}
        for o in op:
            if (not on_ref and self.obj_to_tile[o.label] == tile_num) or \
                    (on_ref and self.obj_to_tile[self.ref_db["orig_label"][o.label]] == tile_num):
                out_stats[o.label] = []
                if OStats.MEAN in stats:
                    out_stats[o.label].append(o.intensity_mean)
                if OStats.STD in stats:
                    out_stats[o.label].append(o.intensity_std)
                if OStats.MIN in stats:
                    out_stats[o.label].append(o.intensity_min)
                if OStats.MAX in stats:
                    out_stats[o.label].append(o.intensity_max)
                out_stats[o.label] = np.array(out_stats[o.label]).flatten(order='F')

        if len(out_stats) > 0:
            return out_stats.keys(), np.vstack(list(out_stats.values()))
        else:
            return None, None

    def add_raster_for_stats(self, raster_file, raster_name=None, stats=[OStats.MEAN], as_group=True):
        self.raster_files.append(raster_file)
        self.raster_stats.append(stats)
        if raster_name is None:
            raster_name = os.path.splitext(os.path.basename(raster_file))[0]
        with rio.open(raster_file) as ds:
            n_bands = ds.count
        ds = None
        to_append = [raster_name + '_{}_{}'.format(i+1,j.value) for i in range(n_bands) for j in stats]
        if as_group:
            self.raster_groups.extend(range(self.n_vars, self.n_vars+len(to_append)))
        self.raster_var_names.append(to_append)
        self.n_vars += len(to_append)
        return

    def add_raster_time_series_for_stats(self, raster_list, ts_name=None, stats=[OStats.MEAN]):
        #Check
        n_bands = []
        for r in raster_list:
            with rio.open(r) as ds:
                n_bands.append(ds.count)
        assert all(x == n_bands[0] for x in n_bands)
        n_bands = n_bands[0]

        pos = self.n_vars
        for i,r in enumerate(raster_list):
            if ts_name is not None:
                raster_name = '{}_D{}'.format(ts_name,i)
            self.add_raster_for_stats(r, stats=stats, raster_name=ts_name, as_group=False)
        for i in range(n_bands):
            for s in range(len(stats)):
                self.raster_groups.append(list(range(len(stats)*i+s+pos,
                                                     len(stats)*i+s+pos+len(stats)*n_bands*len(raster_list),
                                                     len(stats)*n_bands)))
        return

    def get_full_stats_on_tile(self, tilenum):
        out_db = pd.DataFrame(index=[o for o,t in self.obj_to_tile.items() if t == tilenum])
        for rf,rs,rv in zip(self.raster_files, self.raster_stats, self.raster_var_names):
            k,v = self.compute_stats_on_tile(tilenum, rf, rs)
            if k is not None:
                out_db.loc[k,rv] = v
        return out_db

    def populate_ref_db(self):
        assert(self.ref_db is not None)
        for t in tqdm(self.tiles.keys(), desc="Zonal stats on DB", total=len(self.tiles)):
            for rf, rs, rv in zip(self.raster_files, self.raster_stats, self.raster_var_names):
                k, v = self.compute_stats_on_tile(t, rf, rs, on_ref=True)
                if k is not None:
                    self.ref_db.loc[k,rv] = v
        return

    def export_reference_db(self, out_vector): #TO BE CHANGED
        with rio.open(self.object_layer) as ds:
            feats = [{'properties': {'sample_id': v}, 'geometry': s} for s, v in
                 shapes(self.ref_obj_layer.astype(np.int32), mask=self.ref_obj_layer > 0, transform=ds.transform)]
            crs = ds.crs.to_epsg()
        vds = gpd.GeoDataFrame.from_features(feats)
        vds = vds.set_crs(epsg=crs)
        vds = vds.join(self.ref_db, on="sample_id")
        vds.to_file(out_vector)
        return

    def get_vars(self):
        return [item for sublist in self.raster_var_names for item in sublist]

    def get_reference_db_as_training_base(self, class_field='class'):
        if isinstance(class_field, str):
            class_field = [class_field]
        assert(self.ref_db is not None and len(self.raster_var_names)>0)
        vars = self.get_vars()
        out = {}
        out['vars'] = vars
        out['obj_id'] = self.ref_db['orig_label'].to_numpy(dtype=int)
        imp = SimpleImputer(missing_values=np.nan, strategy='mean')
        out['X'] = imp.fit_transform(self.ref_db[vars].to_numpy())
        # compute percentiles and normalize
        out['perc2'] = np.zeros(out['X'].shape[1])
        out['perc98'] = np.zeros(out['X'].shape[1])
        for g in self.raster_groups:
            tmp = out['X'][:,g]
            m,M = np.percentile(tmp, [2, 98])
            if isinstance(g, list):
                for x in g:
                    out['perc2'][x] = m
                    out['perc98'][x] = M
            else:
                out['perc2'][g] = m
                out['perc98'][g] = M
            out['X'][:,g] = (tmp - m)/(M - m)
        for cf in class_field:
            out[cf] = self.ref_db[cf].to_numpy(dtype=int)
        out['groups'] = self.ref_db['polygon_id'].to_numpy(dtype=int)
        return out

    def tiled_data(self, normalize=None):
        vars = [item for sublist in self.raster_var_names for item in sublist]
        for tilenum in self.tiles.keys():
            tile_data = self.get_full_stats_on_tile(tilenum)
            L = tile_data.index.to_numpy(dtype=np.int32)
            # Probably sub-optimal and error-prone, uses mean value from tile only 
            imp = SimpleImputer(missing_values=np.nan, strategy='mean')
            X = imp.fit_transform(tile_data[vars].to_numpy())
            if normalize is not None:
                X = (X - normalize[0]) / (normalize[1] - normalize[0])
            yield tilenum,L,X

    def create_new_map(self):
        self.output_maps.append(np.zeros((self.H, self.W), dtype=np.int))
        return

    def populate_map(self, tilenum, obj_id, classes, output_file=None, compress='NONE'):
        r = otb.itkRegion()
        r['index'][0], r['index'][1], r['size'][0], r['size'][1] = self.tiles[tilenum]
        obj = to_otb_pipeline(self.object_layer)
        obj.PropagateRequestedRegion('out', r)
        clip_obj = obj.ExportImage('out')
        tile_obj = np.squeeze(clip_obj['array']).astype(np.uint32)
        tmp = np.zeros(np.max(tile_obj) + 1, dtype=int)
        tmp[obj_id] = classes
        if output_file is None:
            self.output_map[-1][self.tiles[tilenum][1]:self.tiles[tilenum][1]+self.tiles[tilenum][3],
            self.tiles[tilenum][0]:self.tiles[tilenum][0]+self.tiles[tilenum][2]] += tmp[tile_obj]
        else:
            if not os.path.exists(os.path.dirname(output_file)):
                os.makedirs(os.path.dirname(output_file))
            if tilenum == 0 and os.path.exists(output_file):
                os.remove(output_file)
            mode = 'w+' if not os.path.exists(output_file) else 'r+'
            with rio.open(output_file, mode, driver='GTiff', width=self.W, height=self.H, crs=self.crs,
                          transform=self.transform, count=1, dtype=np.uint16) as dst:
                          # issue with compress option?
                          #transform=self.transform, count=1, dtype=np.uint16, compress=compress) as dst:
                win = Window(self.tiles[tilenum][0], self.tiles[tilenum][1],
                             self.tiles[tilenum][2], self.tiles[tilenum][3])
                x = dst.read(1, window=win)
                dst.write(x + tmp[tile_obj], window=win, indexes=1)
        return

    def export_raster_map(self, output_file, compress=False):
        assert (self.output_map is not None)
        if not os.path.exists(os.path.dirname(output_file)):
            os.makedirs(os.path.dirname(output_file))
        obj = to_otb_pipeline(self.object_layer)
        arr = obj.ExportImage('out')
        arr['array'] = self.output_map.astype(np.float32)
        obj.ImportImage('in', arr)
        if compress:
            output_file = output_file + "?gdal:co:compress=deflate"
        obj.SetParameterString('out', output_file)
        obj.SetParameterOutputImagePixelType('out', otb.ImagePixelType_uint16)
        obj.ExecuteAndWriteOutput()

    def populate_export_raster_map(self, obj_id, classes, output_file, compress='NONE'):
        for tn in tqdm(self.tiles.keys(), desc="Writing output map"):
            self.populate_map(tn,obj_id,classes,output_file,compress)

    def true_pred_bypixel(self, labels, predicted_classes, class_field='class'):
        pred_c = np.zeros(int(np.max(self.ref_db['orig_label']))+1)
        pred_c[labels] = predicted_classes
        support = []
        for tn, t in self.tiled_objects(on_ref=True):
            support.append(t[np.isin(t, labels)])
        support = np.concatenate(support)
        pred = pred_c[support]
        true_c = np.zeros(int(np.max(self.ref_db['orig_label']))+1)
        # ATTENTION: works if "labels" is sorted (as provided by get_reference_...)
        true_c[labels] = self.ref_db.loc[self.ref_db['orig_label'].isin(labels),class_field].to_numpy(dtype=int)
        true = true_c[support]
        return true[pred>0], pred[pred>0]

    def tiled_objects(self, on_ref=False):
        assert(self.tiles is not None)
        idx = -1 if on_ref else 0
        r = otb.itkRegion()
        for tn, t in self.tiles.items():
            r['index'][0], r['index'][1] = t[0], t[1]
            r['size'][0], r['size'][1] = t[2], t[3]
            self.ref_obj_layer_pipe[idx].PropagateRequestedRegion('out', r)
            arr = self.ref_obj_layer_pipe[idx].GetImageAsNumpyArray('out').astype(np.uint32)
            yield tn, arr