tommonopoli's picture
update readme
7829054
import numpy as np
import torch
import rasterio
import xarray as xr
import rioxarray as rxr
import cv2
from transformers import SegformerForSemanticSegmentation
from tqdm import tqdm
from scipy.ndimage import grey_dilation
import matplotlib as mpl
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import make_axes_locatable
from .viz_utils import alpha_composite
from loguru import logger
def resize(img, shape=None, scaling_factor=1., order='CHW'):
"""Resize an image by a given scaling factor"""
assert order in ['HWC', 'CHW'], f"Got unknown order '{order}', expected one of ['HWC','CHW']"
assert shape is None or scaling_factor == 1., "Got both shape and scaling_factor. Please provide only one of them"
# resize image
if order == 'CHW':
img = np.moveaxis(img, 0, -1) # CHW -> HWC
if shape is not None:
img = cv2.resize(img, shape[::-1], interpolation=cv2.INTER_LINEAR)
else:
img = cv2.resize(img, None, fx=scaling_factor, fy=scaling_factor, interpolation=cv2.INTER_LINEAR)
# NB: cv2.resize returns a HW image if the input image is HW1: restore the C dimension
if len(img.shape) == 2:
img = img[..., None]
if order == 'CHW':
img = np.moveaxis(img, -1, 0) # HWC -> CHW
return img
def minimum_needed_padding(img_size, patch_size: int, stride: int):
"""
Compute the minimum padding needed to make an image divisible by a patch size with a given stride.
Args:
image_shape (tuple): the shape (H,W) of the image tensor
patch_size (int): the size of the patches to extract
stride (int): the stride to use when extracting patches
Returns:
tuple: the padding needed to make the image tensor divisible by the patch size with the given stride
"""
img_size = np.array(img_size)
pad = np.where(
img_size <= patch_size,
(patch_size - img_size) % patch_size, # the % patch_size is to handle the case img_size = (0,0)
(stride - (img_size - patch_size)) % stride
)
pad_t, pad_l = pad // 2
pad_b, pad_r = pad[0] - pad_t, pad[1] - pad_l
return pad_t, pad_b, pad_l, pad_r
def pad(img, pad, order='CHW'):
"""Pad an image by the given pad values, in the format (pad_t, pad_b, pad_l, pad_r)"""
assert order in ['HWC', 'CHW'], f"Got unknown order '{order}', expected one of ['HWC','CHW']"
pad_t, pad_b, pad_l, pad_r = pad
# pad image
if order == 'HWC':
padded_img = np.pad(img, ((pad_t,pad_b), (pad_l,pad_r), (0,0)), mode='constant', constant_values=0) # can also try mode='reflect'
else:
padded_img = np.pad(img, ((0,0), (pad_t,pad_b), (pad_l,pad_r)), mode='constant', constant_values=0) # can also try mode='reflect'
if isinstance(img, torch.Tensor):
padded_img = torch.tensor(padded_img)
return padded_img
def extract_patches(img, patch_size=512, stride=256, order='CHW', only_return_idx=True, include_last=True):
"""Extract patches from an image, in the format (h_start, h_end, w_start, w_end)"""
assert order in ['HWC', 'CHW'], f"Got unknown order '{order}', expected one of ['HWC','CHW']"
assert len(img.shape) == 3, f"Got image with {len(img.shape)} dimensions, expected 3 dimensions (C,H,W) or (H,W,C)"
assert img.shape[0] >= patch_size, f"Got image with height {img.shape[0]}, expected at least {patch_size}. Maybe apply padding first?"
assert img.shape[1] >= patch_size, f"Got image with width {img.shape[1]}, expected at least {patch_size}. Maybe apply padding first?"
# Get image height and width
if order == 'HWC':
H, W = img.shape[:2]
else:
H, W = img.shape[1:]
# Compute the number of "proper" patches in each dimension
n_patches_H = (H - patch_size) // stride + 1
n_patches_W = (W - patch_size) // stride + 1
# Extract patches indices
patches_idx = []
for i in range(n_patches_H): # iterate over height
for j in range(n_patches_W): # iterate over width
# Get the current patch indices
patches_idx.append((i*stride, i*stride+patch_size, j*stride, j*stride+patch_size)) # (top, bottom, left, right)
# Include leftmost and lowermost patch if needed
if include_last:
if j == n_patches_W-1 and j*stride+patch_size < W:
patches_idx.append((i*stride, i*stride+patch_size, W-patch_size, W))
if i == n_patches_H-1 and i*stride+patch_size < H:
patches_idx.append((H-patch_size, H, j*stride, j*stride+patch_size))
if i == n_patches_H-1 and j == n_patches_W-1 and i*stride+patch_size < H and j*stride+patch_size < W:
patches_idx.append((H-patch_size, H, W-patch_size, W))
if only_return_idx:
return patches_idx
else:
# Extract patches
patches = []
for t,b,l,r in patches_idx:
if order == 'HWC':
patch = img[t:b, l:r, :]
else:
patch = img[:, t:b, l:r]
patches.append(patch)
return patches, patches_idx
def segment_batch(batch, model):
# perform prediction
with torch.no_grad():
out = model(batch) # (n_patches, 1, H, W) logits
if isinstance(model, SegformerForSemanticSegmentation):
out = upsample(out.logits, size=batch.shape[-2:])
# apply sigmoid
out = torch.sigmoid(out) # logits -> confidence scores
return out
def upsample(x, size):
"""Upsample a 3D/4D/5D tensor"""
return torch.nn.functional.interpolate(x, size=size, mode='bilinear', align_corners=False)
def merge_patches(patches, patches_idx, rotate=False, canvas_shape=None, order='CHW'): # TODO
"""Merge patches into a single image"""
assert order in ['HWC', 'CHW'], f"Got unknown order '{order}', expected one of ['HWC','CHW']"
if rotate:
axes_to_rotate = (0,1) if order == 'HWC' else (1,2)
patches = [np.rot90(p, -i, axes=axes_to_rotate) for i,p in enumerate(patches)]
else:
assert len(patches) == len(patches_idx), f"Got {len(patches)} patches and {len(patches_idx)} indexes"
# if canvas_shape is None, infer it from patches_idx
if canvas_shape is None:
patches_idx_zipped = list(zip(*patches_idx))
canvas_H = max(patches_idx_zipped[1])
canvas_W = max(patches_idx_zipped[3])
else:
canvas_H, canvas_W = canvas_shape
# initialize canvas
dtype = patches[0].dtype
if order == 'HWC':
canvas_C = patches[0].shape[-1]
canvas = np.zeros((canvas_H, canvas_W, canvas_C), dtype=dtype) # HWC
n_overlapping_patches = np.zeros((canvas_H, canvas_W, 1))
else:
canvas_C = patches[0].shape[0]
canvas = np.zeros((canvas_C, canvas_H, canvas_W, ), dtype=dtype) # CHW
n_overlapping_patches = np.zeros((1, canvas_H, canvas_W))
# merge patches
for p, (t,b,l,r) in zip(patches, patches_idx):
if order == 'HWC':
canvas[t:b, l:r, :] += p
n_overlapping_patches[t:b, l:r, 0] += 1
else:
canvas[:, t:b, l:r] += p
n_overlapping_patches[0, t:b, l:r] += 1
# compute average
canvas = np.divide(canvas, n_overlapping_patches, where=(n_overlapping_patches != 0))
return canvas
def segment(img, model, patch_size=512, stride=256, scaling_factor=1., rotate=False, device=None, batch_size=16, verbose=False):
"""Segment an RGB image by using a segmentation model. Returns a probability
map (and performance metrics, if requested)"""
# some checks
assert isinstance(img, np.ndarray), f"Input must be a numpy array. Got {type(img)}"
assert img.shape[0] in [3,4], f"Input image must be formatted as CHW, with C = 3,4. Got a shape of {img.shape}"
assert img.dtype == np.uint8, f"Input image must be a numpy array with dtype np.uint8. Got {img.dtype}"
# prepare model for evaluation
model = model.to(device)
model.eval()
# prepare alpha channel
original_shape = img.shape
if img.shape[0] == 3:
# create dummy alpha channel
alpha = np.full(original_shape[1:], 255, dtype=np.uint8)
else:
# extract alpha channel
img, alpha = img[:3], img[3]
# resize image
img = resize(img, scaling_factor=scaling_factor)
# pad image
pad_t, pad_b, pad_l, pad_r = minimum_needed_padding(img.shape[1:], patch_size, stride)
padded_img = pad(img, pad=(pad_t, pad_b, pad_l, pad_r))
padded_shape = padded_img.shape
# extract patches indexes
patches_idx = extract_patches(padded_img, patch_size=patch_size, stride=stride)
### segment
masks = []
masks_idx = []
batch = []
for i, p_idx in enumerate(tqdm(patches_idx, disable=not verbose, desc="Predicting...", total=len(patches_idx))):
t, b, l, r = p_idx
# extract patch
patch = padded_img[:, t:b, l:r]
# consider patch only if it is valid (i.e. not all black or all white)
if np.any(patch != 0) and np.any(patch != 255):
# convert patch to torch.tensor with float32 values in [0,1] (as required by torch)
patch = torch.tensor(patch).float() / 255.
# normalize patch with ImageNet mean and std
patch = (patch - torch.tensor([0.485, 0.456, 0.406]).view(3,1,1)) / torch.tensor([0.229, 0.224, 0.225]).view(3,1,1)
# add patch to batch
batch.append(patch)
masks_idx.append(p_idx)
# (optional) for each patch extracted, consider also its rotated versions
if rotate:
for rot in range(1,4):
patch = torch.rot90(patch, rot, dims=[1,2])
batch.append(patch)
masks_idx.append(p_idx)
# if the batch is full, perform prediction
if len(batch) >= batch_size or i == len(patches_idx)-1:
# move batch to GPU
batch = torch.stack(batch).to(device)
# perform prediction
out = segment_batch(batch, model)
# append predictions to masks
masks.append(out.cpu().numpy())
# reset batch
batch = []
# concatenate predictions
masks = np.concatenate(masks) # (n_patches, 1, H, W)
# merge patches
mask = merge_patches(masks, masks_idx, rotate=rotate, canvas_shape=padded_shape[1:]) # (1, H, W)
# undo padding
mask = mask[:, pad_t:padded_shape[1]-pad_b, pad_l:padded_shape[2]-pad_r]
# resize mask to original shape
mask = resize(mask, shape=original_shape[1:])
# apply alpha channel, i.e. set to -1 the pixels where alpha is 0
mask = np.where(alpha == 0, -1, mask)
return mask.squeeze()
def sliding_window_avg_pooling(img, window, granularity, alpha=None, min_nonblank_pixels=0., order="HWC", normalize=False, return_min_max=False, verbose=False):
assert isinstance(img, np.ndarray), f'Input image must be a numpy array. Got {type(img)}'
if order == "HWC":
assert img.shape[2] == 1, f'Input image must be formatted as HWC, with C = 1. Got a shape of {img.shape}'
elif order == "CHW":
assert img.shape[0] == 1, f'Input image must be formatted as CHW, with C = 1. Got a shape of {img.shape}'
# check if alpha channel was given, and cast it to np.float32 with values in [0,1]
if alpha is not None:
assert img.shape == alpha.shape, f'The shape of input image {img.shape} and alpha channel {alpha.shape} do not match'
if alpha.dtype == np.uint8:
alpha = (alpha / 255).astype(np.float32)
elif alpha.dtype == bool:
alpha = alpha.astype(np.float32)
else:
alpha = np.ones_like(img, dtype=np.float32)
# compute threshold
thresh = min_nonblank_pixels * window**2
# extract patches idxs
patches_idx = extract_patches(img, patch_size=window, stride=granularity, order=order, only_return_idx=True)
# initialize canvas
canvas = np.zeros_like(img, dtype=np.float32)
n_overlapping_patches = np.zeros_like(img, dtype=np.float32)
# cycle through patches idxs
for t,b,l,r in tqdm(patches_idx, disable=not verbose):
p_a = alpha[t:b,l:r]
n_valid_pixels = p_a.sum()
# keep only if it has more than min_nonblank_pixels
if n_valid_pixels <= thresh:
continue
# compute average patch value (i.e. density inside the patch)
p = img[t:b,l:r]
p_density = (p * p_a).sum() / n_valid_pixels
# add to canvas
canvas[t:b,l:r] += p_density
n_overlapping_patches[t:b,l:r] += 1
# compute average density
density_map = np.divide(canvas, n_overlapping_patches, where=(n_overlapping_patches != 0))
# apply alpha
density_map = density_map * alpha
if normalize:
# [0,1]-normalize
density_map_min = density_map.min()
density_map_max = density_map.max()
density_map = (density_map - density_map_min) / (density_map_max - density_map_min)
if return_min_max:
return density_map, density_map_min, density_map_max
return density_map
def compute_vndvi(
raster: np.ndarray,
mask: np.ndarray,
dilate_rows=True,
window_size=360,
granularity=45,
):
assert isinstance(raster, np.ndarray)
assert isinstance(mask, np.ndarray)
assert len(raster.shape) == 3 # CHW
assert len(mask.shape) == 2 # HW
assert raster.shape[0] in [3,4] # RGB or RGBA
# CHW -> HWC
raster = raster.transpose(1,2,0)
# Extract channels
_raster = raster.astype(np.float32) / 255 # convert to float32 in [0,1]
R, G, B = _raster[:,:,0], _raster[:,:,1], _raster[:,:,2]
# To avoid division by 0 due to negative power, we replace 0 with 1 in R and B channels
R = np.where(R == 0, 1, R)
B = np.where(B == 0, 1, B)
# Mask has values: 0=interrows, 255=rows, 1=nodata
# Get mask for the rows and interrows
mask_rows = (mask == 255)
mask_interrows = (mask == 0)
mask_valid = mask_rows | mask_interrows
# Compute vndvi
vndvi = 0.5268 * (R**(-0.1294) * G**(0.3389) * B**(-0.3118))
# Clip values to [0,1]
vndvi = np.clip(vndvi, 0, 1)
# Compute 10th and 90th percentile on whole vineyard vndvi heatmap
vndvi_perc10, vndvi_perc90 = np.percentile(vndvi[mask_valid], [10,90])
# Clip values between 10th and 90th percentile
vndvi_clipped = np.clip(vndvi, vndvi_perc10, vndvi_perc90)
# Perform sliding window average pooling to smooth the heatmap
# NB: the window takes into account only the rows
vndvi_rows_clipped_pooled = sliding_window_avg_pooling(
np.where(mask_rows, vndvi_clipped, 0)[..., None],
window = int(window_size / 4),
granularity = granularity,
alpha = mask_rows[..., None],
min_nonblank_pixels = 0.0,
verbose=True,
)
# Same, but for interrows
vndvi_interrows_clipped_pooled = sliding_window_avg_pooling(
np.where(mask_interrows, vndvi_clipped, 0)[..., None],
window = int(window_size / 4),
granularity = granularity,
alpha = mask_interrows[..., None],
min_nonblank_pixels = 0.0,
verbose=True,
)
# Apply dilation to rows mask
dil_factor = int(window_size / 60)
mask_rows_dilated = grey_dilation(mask_rows, size=(dil_factor, dil_factor))
vndvi_rows_clipped_pooled_dilated = grey_dilation(vndvi_rows_clipped_pooled, size=(dil_factor, dil_factor, 1))
# For visualization purposes, normalize with vndvi_perc10 and
# vndvi_perc90 (because we want vndvi_perc10 to be the first color of
# the colormap and vndvi_perc90 to be the last)
vndvi_rows_clipped_pooled_normalized = (vndvi_rows_clipped_pooled - vndvi_perc10) / (vndvi_perc90 - vndvi_perc10)
vndvi_rows_clipped_pooled_dilated_normalized = (vndvi_rows_clipped_pooled_dilated - vndvi_perc10) / (vndvi_perc90 - vndvi_perc10)
vndvi_interrows_clipped_pooled_normalized = (vndvi_interrows_clipped_pooled - vndvi_perc10) / (vndvi_perc90 - vndvi_perc10)
# for visualization
vndvi_rows_img = alpha_composite(
raster,
vndvi_rows_clipped_pooled_dilated_normalized if dilate_rows else vndvi_rows_clipped_pooled_normalized,
opacity = 1.0,
colormap = 'RdYlGn',
alpha_image = np.zeros_like(raster[:,:,[0]]),
alpha_mask = mask_rows_dilated[...,None] if dilate_rows else mask_rows[...,None],
) # HW4 RGBA
vndvi_interrows_img = alpha_composite(
raster,
vndvi_interrows_clipped_pooled_normalized,
opacity = 1.0,
colormap = 'RdYlGn',
alpha_image = np.zeros_like(raster[:,:,[0]]),
alpha_mask = mask_interrows[...,None],
) # HW4 RGBA
# add colorbar
# fig_rows, ax = plt.subplots(1, 1, figsize=(10, 10))
# divider = make_axes_locatable(ax)
# cax = divider.append_axes('right', size='5%', pad=0.15)
# ax.imshow(vndvi_rows_img)
# fig_rows.colorbar(
# mappable = mpl.cm.ScalarMappable(
# norm = mpl.colors.Normalize(
# vmin = vndvi_perc10,
# vmax = vndvi_perc90),
# cmap = 'RdYlGn'),
# cax = cax,
# orientation = 'vertical',
# label = 'vNDVI',
# shrink = 1)
# fig_interrows, ax = plt.subplots(1, 1, figsize=(10, 10))
# divider = make_axes_locatable(ax)
# cax = divider.append_axes('right', size='5%', pad=0.15)
# ax.imshow(vndvi_interrows_img)
# fig_interrows.colorbar(
# mappable = mpl.cm.ScalarMappable(
# norm = mpl.colors.Normalize(
# vmin = vndvi_perc10,
# vmax = vndvi_perc90),
# cmap = 'RdYlGn'),
# cax = cax,
# orientation = 'vertical',
# label = 'vNDVI',
# shrink = 1)
# return fig_rows, fig_interrows
return vndvi_rows_img, vndvi_interrows_img
def compute_vdi(
raster: np.ndarray,
mask: np.ndarray,
window_size=360,
granularity=40,
):
# CHW -> HWC
raster = raster.transpose(1,2,0)
# Mask has values: 0=interrows, 255=rows, 1=nodata
# Get mask for the rows and interrows
mask_rows = (mask == 255)
mask_interrows = (mask == 0)
mask_valid = mask_rows | mask_interrows
# compute vdi
vdi, vdi_min, vdi_max = sliding_window_avg_pooling(
mask_rows[...,None],
window=window_size,
granularity=granularity,
alpha=mask_valid[...,None],
min_nonblank_pixels=0.9,
normalize=True,
return_min_max=True,
verbose=True,
)
# for visualization
vdi_img = alpha_composite(
raster,
vdi,
opacity = 1,
colormap = 'jet_r',
alpha_image = mask_valid[...,None],
alpha_mask = mask_valid[...,None],
)
# add colorbar
# fig, ax = plt.subplots(1, 1, figsize=(10, 10))
# divider = make_axes_locatable(ax)
# cax = divider.append_axes('right', size='5%', pad=0.15)
# ax.imshow(vdi_img)
# fig.colorbar(
# mappable = mpl.cm.ScalarMappable(
# norm = mpl.colors.Normalize(
# vmin = vdi_min,
# vmax = vdi_max),
# cmap = 'jet_r'),
# cax = cax,
# orientation = 'vertical',
# label = 'VDI',
# shrink = 1)
# return fig
return vdi_img
def compute_mask(
raster: np.ndarray,
model: torch.nn.Module,
patch_size=512,
stride=256,
scaling_factor=None,
rotate=False,
batch_size=16
):
assert isinstance(raster, np.ndarray), f'Input raster must be a numpy array. Got {type(raster)}'
assert len(raster.shape) == 3, f'Input raster must have 3 dimensions (bands, rows, cols). Got shape {raster.shape}'
assert raster.shape[0] in [3,4], f'Input raster must have 3 bands (RGB) or 4 bands (RGBA). Got {raster.shape[0]} bands'
assert isinstance(model, torch.nn.Module), 'Model must be a torch.nn.Module'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Infer GSD
#gsd = abs(raster.rio.transform()[0]) # ground sampling distance (NB: valid only if image is a GeoTIFF)
# Growseg works best on orthoimages with gsd in [1, 1.7] cm/px. You may want to
# specify a scaling factor different from 1 if your image has a different gsd.
# E.g.: SCALING_FACTOR = gsd / 0.015
# logger.info(f'Image GSD: {gsd*100:.2f} cm/px')
# scaling_factor = scaling_factor or (gsd / 0.015)
scaling_factor = scaling_factor or 1
logger.info(f'Applying scaling factor: {scaling_factor:.2f}')
# segment
logger.info('Segmenting image...')
score_map = segment(
raster,
model,
patch_size=patch_size,
stride=stride,
scaling_factor=scaling_factor,
rotate=rotate,
device=device,
batch_size=batch_size,
verbose=True
) # mask is a HxW float32 array in [0, 1]
# apply threshold on confidence scores
alpha = (score_map == -1)
mask = (score_map > 0.5)
# convert to uint8
mask = (mask * 255).astype(np.uint8)
# set nodata pixels to 1
mask[alpha] = 1
return mask