Object-oriented classification example based on Python Ⅱ

Continue the introduction from the previous part~

1. Convert land type vector to raster

This step is to make the ground class value and the object of the image fall in the same area, thereby assimilating the segmented objects in the image into the actual ground object class.

train_fn = r".\train_data1.shp"
train_ds = ogr.Open(train_fn)
lyr = train_ds.GetLayer()
driver = gdal.GetDriverByName('MEM')
target_ds = driver.Create('', im_width, im_height, 1, gdal.GDT_UInt16)
target_ds.SetGeoTransform(im_geotrans)
target_ds.SetProjection(im_proj)
options = ['ATTRIBUTE=tyPE']
gdal.RasterizeLayer(target_ds, [1], lyr, options=options)
data = target_ds.GetRasterBand(1).ReadAsArray()
ground_truth = target_ds.GetRasterBand(1).ReadAsArray()
ground_truth = ground_truth.transpose((1, 0))
classes = np.unique(ground_truth)[1:]

Finally, raster point data with feature category data is obtained.

2. Feature matching

After matching the obtained grid point real surface object data with the image object through iteration, the corresponding characteristics of the object are found through the iterator.

segments_per_class = {}
for klass in classes:
    segments_of_class = segments[ground_truth == klass]
    segments_per_class[klass] = set(segments_of_class)
 
intersection = set()
accum = set()
for class_segments in segments_per_class.values():
    intersection |= accum.intersection(class_segments)
    accum |= class_segments
assert len(intersection) == 0, "Segment(s) represent multiple classes"
print("adjust complete")
end3 = time.time()
print('Running time2: %s Seconds'%(end3-start3))




print("start train randomforest classification")
start4 = time.time()
train_img = np.copy(segments)
threshold = train_img.max() + 1 

for klass in classes:
    class_label = threshold + klass
    for segment_id in segments_per_class[klass]:
        train_img[train_img == segment_id] = class_label
 
train_img[train_img <= threshold] = 0
train_img[train_img > threshold] -= threshold

training_objects = []
training_labels = []
for klass in classes:
    class_train_object = [v for i, v in enumerate(objects) if segment_ids[i] in segments_per_class[klass]]
    training_labels += [klass] * len(class_train_object)
    training_objects += class_train_object

In the actual construction process of image samples, some ground object samples may have a small distance from each other, causing two or more samples to fall in the same segmentation area. This will cause the feature matching iteration to continue indefinitely, so we have to start from the two Choose one of one or more samples.

3. Classification

Finally, scikit-learn's random forest classifier is used to predict the sample segmentation blocks and other undefined segmentation blocks, and finally the results are output to the raster.

def PolygonizeTheRaster(inputfile,outputfile):
    dataset = gdal.Open(inputfile, gdal.GA_ReadOnly)
    srcband=dataset.GetRasterBand(1)
    im_proj = dataset.GetProjection()
    prj = osr.SpatialReference() 
    prj.ImportFromWkt(im_proj)
    drv = ogr.GetDriverByName('ESRI Shapefile')
    dst_ds = drv.CreateDataSource(outputfile)
    dst_layername = 'out'
    dst_layer = dst_ds.CreateLayer(dst_layername, srs=prj)
    dst_fieldname = 'DN'
    fd = ogr.FieldDefn(dst_fieldname, ogr.OFTInteger)
    dst_layer.CreateField(fd)
    dst_field = 0
    gdal.Polygonize(srcband, None, dst_layer, dst_field) 

classifier = RandomForestClassifier(n_jobs=-1)  
classifier.fit(training_objects, training_labels) 
predicted = classifier.predict(objects)  

clf = segments.copy()
for segment_id, klass in zip(segment_ids, predicted):
    clf[clf == segment_id] = klass
# temp = temp.transpose((2, 1, 0))
mask = np.sum(temp, axis=2)  
mask[mask > 0.0] = 1.0   
mask[mask == 0.0] = -1.0

clf = np.multiply(clf, mask)
clf[clf < 0] = -9999.0
clf = clf.transpose((1, 0))
clfds = driverTiff.Create(r"D:\Data\testdata\classification\result.tif", im_width, im_height,
                          1, gdal.GDT_Float32) 
clfds.SetGeoTransform(im_geotrans)
clfds.SetProjection(im_proj)
clfds.GetRasterBand(1).SetNoDataValue(-9999.0)
clfds.GetRasterBand(1).WriteArray(clf)
clfds = None

end4 = time.time()
outputfile = r".\result2.shp"
print('Running time2: %s Seconds'%(end4-start4))
PolygonizeTheRaster(r".\result.tif",outputfile)

The tag ID of a class in the image is 1, 2, 3 and so on, so it is difficult to view it with the naked eye using general raster viewing software. In order to facilitate readers' viewing and mapping, I also converted raster to vector, so that the classification situation can be clearly viewed in arcmap.

The final classification result:

Figure 1 Classification result diagram

Figure 2 Woodland classification effect