Spaces:
Running
Running
initial commit
Browse files- app.py +302 -0
- requirements.txt +7 -0
app.py
ADDED
@@ -0,0 +1,302 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import numpy as np
|
3 |
+
import os
|
4 |
+
import cv2
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
from huggingface_hub import snapshot_download
|
7 |
+
import rasterio
|
8 |
+
from rasterio.enums import Resampling
|
9 |
+
from rasterio.plot import reshape_as_image
|
10 |
+
import sys
|
11 |
+
|
12 |
+
# Download the entire repository to a subdirectory
|
13 |
+
repo_id = "truthdotphd/cloud-detection"
|
14 |
+
repo_subdir = "."
|
15 |
+
repo_dir = snapshot_download(repo_id=repo_id, local_dir=repo_subdir)
|
16 |
+
|
17 |
+
# Add the repository directory to the Python path
|
18 |
+
sys.path.append(repo_dir)
|
19 |
+
|
20 |
+
# Import the necessary functions from the downloaded modules
|
21 |
+
try:
|
22 |
+
from omnicloudmask import predict_from_array
|
23 |
+
except ImportError:
|
24 |
+
omnicloudmask_dir = os.path.join(repo_dir, "omnicloudmask")
|
25 |
+
if os.path.exists(omnicloudmask_dir):
|
26 |
+
sys.path.append(omnicloudmask_dir)
|
27 |
+
from omnicloudmask import predict_from_array
|
28 |
+
else:
|
29 |
+
raise ImportError("Could not find the omnicloudmask module in the downloaded repository")
|
30 |
+
|
31 |
+
def visualize_rgb(red_file, green_file, blue_file, nir_file):
|
32 |
+
"""
|
33 |
+
Create and display an RGB visualization immediately after images are uploaded.
|
34 |
+
"""
|
35 |
+
if not all([red_file, green_file, blue_file, nir_file]):
|
36 |
+
return None
|
37 |
+
|
38 |
+
try:
|
39 |
+
# Get dimensions from red band to use for resampling
|
40 |
+
with rasterio.open(red_file) as src:
|
41 |
+
target_height = src.height
|
42 |
+
target_width = src.width
|
43 |
+
|
44 |
+
# Load bands
|
45 |
+
blue_data = load_band(blue_file)
|
46 |
+
green_data = load_band(green_file)
|
47 |
+
red_data = load_band(red_file)
|
48 |
+
|
49 |
+
# Compute max values for each channel for dynamic normalization
|
50 |
+
red_max = np.max(red_data)
|
51 |
+
green_max = np.max(green_data)
|
52 |
+
blue_max = np.max(blue_data)
|
53 |
+
|
54 |
+
# Create RGB image for visualization with dynamic normalization
|
55 |
+
rgb_image = np.zeros((red_data.shape[0], red_data.shape[1], 3), dtype=np.float32)
|
56 |
+
|
57 |
+
# Normalize each channel individually
|
58 |
+
epsilon = 1e-10
|
59 |
+
rgb_image[:, :, 0] = red_data / (red_max + epsilon)
|
60 |
+
rgb_image[:, :, 1] = green_data / (green_max + epsilon)
|
61 |
+
rgb_image[:, :, 2] = blue_data / (blue_max + epsilon)
|
62 |
+
|
63 |
+
# Clip values to 0-1 range
|
64 |
+
rgb_image = np.clip(rgb_image, 0, 1)
|
65 |
+
|
66 |
+
# Apply contrast enhancement for better visualization
|
67 |
+
p2 = np.percentile(rgb_image, 2)
|
68 |
+
p98 = np.percentile(rgb_image, 98)
|
69 |
+
rgb_image_enhanced = np.clip((rgb_image - p2) / (p98 - p2), 0, 1)
|
70 |
+
|
71 |
+
# Convert to uint8 for display
|
72 |
+
rgb_display = (rgb_image_enhanced * 255).astype(np.uint8)
|
73 |
+
|
74 |
+
return rgb_display
|
75 |
+
except Exception as e:
|
76 |
+
print(f"Error generating RGB preview: {e}")
|
77 |
+
return None
|
78 |
+
|
79 |
+
|
80 |
+
def visualize_jp2(file_path):
|
81 |
+
"""
|
82 |
+
Visualize a single JP2 file.
|
83 |
+
"""
|
84 |
+
with rasterio.open(file_path) as src:
|
85 |
+
# Read the data
|
86 |
+
data = src.read(1)
|
87 |
+
|
88 |
+
# Normalize the data for visualization
|
89 |
+
data = (data - np.min(data)) / (np.max(data) - np.min(data))
|
90 |
+
|
91 |
+
# Apply a colormap for better visualization
|
92 |
+
cmap = plt.get_cmap('viridis')
|
93 |
+
colored_image = cmap(data)
|
94 |
+
|
95 |
+
# Convert to 8-bit for display
|
96 |
+
return (colored_image[:, :, :3] * 255).astype(np.uint8)
|
97 |
+
|
98 |
+
def load_band(file_path, resample=False, target_height=None, target_width=None):
|
99 |
+
"""
|
100 |
+
Load a single band from a raster file with optional resampling.
|
101 |
+
"""
|
102 |
+
with rasterio.open(file_path) as src:
|
103 |
+
if resample and target_height is not None and target_width is not None:
|
104 |
+
band_data = src.read(
|
105 |
+
out_shape=(src.count, target_height, target_width),
|
106 |
+
resampling=Resampling.bilinear
|
107 |
+
)[0].astype(np.float32)
|
108 |
+
else:
|
109 |
+
band_data = src.read()[0].astype(np.float32)
|
110 |
+
|
111 |
+
return band_data
|
112 |
+
|
113 |
+
def prepare_input_array(red_file, green_file, blue_file, nir_file):
|
114 |
+
"""
|
115 |
+
Prepare a stacked array of satellite bands for cloud mask prediction.
|
116 |
+
"""
|
117 |
+
# Get dimensions from red band to use for resampling
|
118 |
+
with rasterio.open(red_file) as src:
|
119 |
+
target_height = src.height
|
120 |
+
target_width = src.width
|
121 |
+
|
122 |
+
# Load bands (resample NIR band to match 10m resolution)
|
123 |
+
blue_data = load_band(blue_file)
|
124 |
+
green_data = load_band(green_file)
|
125 |
+
red_data = load_band(red_file)
|
126 |
+
nir_data = load_band(
|
127 |
+
nir_file,
|
128 |
+
resample=True,
|
129 |
+
target_height=target_height,
|
130 |
+
target_width=target_width
|
131 |
+
)
|
132 |
+
|
133 |
+
# Print band shapes for debugging
|
134 |
+
print(f"Band shapes - Blue: {blue_data.shape}, Green: {green_data.shape}, Red: {red_data.shape}, NIR: {nir_data.shape}")
|
135 |
+
|
136 |
+
# Compute max values for each channel for dynamic normalization
|
137 |
+
red_max = np.max(red_data)
|
138 |
+
green_max = np.max(green_data)
|
139 |
+
blue_max = np.max(blue_data)
|
140 |
+
|
141 |
+
print(f"Max values - Red: {red_max}, Green: {green_max}, Blue: {blue_max}")
|
142 |
+
|
143 |
+
# Create RGB image for visualization with dynamic normalization
|
144 |
+
rgb_image = np.zeros((red_data.shape[0], red_data.shape[1], 3), dtype=np.float32)
|
145 |
+
|
146 |
+
# Normalize each channel individually
|
147 |
+
# Add a small epsilon to avoid division by zero
|
148 |
+
epsilon = 1e-10
|
149 |
+
rgb_image[:, :, 0] = red_data / (red_max + epsilon)
|
150 |
+
rgb_image[:, :, 1] = green_data / (green_max + epsilon)
|
151 |
+
rgb_image[:, :, 2] = blue_data / (blue_max + epsilon)
|
152 |
+
|
153 |
+
# Clip values to 0-1 range
|
154 |
+
rgb_image = np.clip(rgb_image, 0, 1)
|
155 |
+
|
156 |
+
# Optional: Apply contrast enhancement for better visualization
|
157 |
+
p2 = np.percentile(rgb_image, 2)
|
158 |
+
p98 = np.percentile(rgb_image, 98)
|
159 |
+
rgb_image_enhanced = np.clip((rgb_image - p2) / (p98 - p2), 0, 1)
|
160 |
+
|
161 |
+
# Stack bands in CHW format for cloud mask prediction (red, green, nir)
|
162 |
+
prediction_array = np.stack([red_data, green_data, nir_data], axis=0)
|
163 |
+
|
164 |
+
return prediction_array, rgb_image_enhanced
|
165 |
+
|
166 |
+
|
167 |
+
def visualize_cloud_mask(rgb_image, pred_mask):
|
168 |
+
"""
|
169 |
+
Create a visualization of the cloud mask overlaid on the RGB image.
|
170 |
+
"""
|
171 |
+
# Ensure pred_mask has the right dimensions
|
172 |
+
if pred_mask.ndim > 2:
|
173 |
+
pred_mask = np.squeeze(pred_mask)
|
174 |
+
|
175 |
+
print(f"RGB image shape: {rgb_image.shape}, Pred mask shape: {pred_mask.shape}")
|
176 |
+
|
177 |
+
# Ensure mask has the same spatial dimensions as the image
|
178 |
+
if pred_mask.shape != rgb_image.shape[:2]:
|
179 |
+
pred_mask = cv2.resize(
|
180 |
+
pred_mask.astype(np.float32),
|
181 |
+
(rgb_image.shape[1], rgb_image.shape[0]),
|
182 |
+
interpolation=cv2.INTER_NEAREST
|
183 |
+
).astype(np.uint8)
|
184 |
+
print(f"Resized mask shape: {pred_mask.shape}")
|
185 |
+
|
186 |
+
# Define colors for each class
|
187 |
+
colors = {
|
188 |
+
0: [0, 255, 0], # Clear - Green
|
189 |
+
1: [255, 255, 255], # Thick Cloud - White
|
190 |
+
2: [200, 200, 200], # Thin Cloud - Light Gray
|
191 |
+
3: [100, 100, 100] # Cloud Shadow - Dark Gray
|
192 |
+
}
|
193 |
+
|
194 |
+
# Create a color-coded mask
|
195 |
+
mask_vis = np.zeros((pred_mask.shape[0], pred_mask.shape[1], 3), dtype=np.uint8)
|
196 |
+
for class_idx, color in colors.items():
|
197 |
+
mask_vis[pred_mask == class_idx] = color
|
198 |
+
|
199 |
+
# Create a blended visualization
|
200 |
+
alpha = 0.5
|
201 |
+
blended = cv2.addWeighted((rgb_image * 255).astype(np.uint8), 1-alpha, mask_vis, alpha, 0)
|
202 |
+
|
203 |
+
# Get the width of the blended image for the legend
|
204 |
+
image_width = blended.shape[1]
|
205 |
+
|
206 |
+
# Create a legend with the same width as the image
|
207 |
+
legend = np.ones((100, image_width, 3), dtype=np.uint8) * 255
|
208 |
+
legend_text = ["Clear", "Thick Cloud", "Thin Cloud", "Cloud Shadow"]
|
209 |
+
legend_colors = [colors[i] for i in range(4)]
|
210 |
+
|
211 |
+
for i, (text, color) in enumerate(zip(legend_text, legend_colors)):
|
212 |
+
cv2.rectangle(legend, (10, 10 + i*20), (30, 30 + i*20), color, -1)
|
213 |
+
cv2.putText(legend, text, (40, 25 + i*20), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
|
214 |
+
|
215 |
+
# Combine image and legend
|
216 |
+
final_output = np.vstack([blended, legend])
|
217 |
+
|
218 |
+
return final_output
|
219 |
+
|
220 |
+
def process_satellite_images(red_file, green_file, blue_file, nir_file, batch_size, patch_size, patch_overlap):
|
221 |
+
"""
|
222 |
+
Process the satellite images and detect clouds.
|
223 |
+
"""
|
224 |
+
if not all([red_file, green_file, blue_file, nir_file]):
|
225 |
+
return None, None, "Please upload all four channel files (Red, Green, Blue, NIR)"
|
226 |
+
|
227 |
+
# Prepare input array and RGB image for visualization
|
228 |
+
input_array, rgb_image = prepare_input_array(red_file, green_file, blue_file, nir_file)
|
229 |
+
|
230 |
+
# Convert RGB image to format suitable for display
|
231 |
+
rgb_display = (rgb_image * 255).astype(np.uint8)
|
232 |
+
|
233 |
+
# Predict cloud mask using omnicloudmask
|
234 |
+
pred_mask = predict_from_array(
|
235 |
+
input_array,
|
236 |
+
batch_size=batch_size,
|
237 |
+
patch_size=patch_size,
|
238 |
+
patch_overlap=patch_overlap
|
239 |
+
)
|
240 |
+
|
241 |
+
# Calculate class distribution
|
242 |
+
if pred_mask.ndim > 2:
|
243 |
+
flat_mask = np.squeeze(pred_mask)
|
244 |
+
else:
|
245 |
+
flat_mask = pred_mask
|
246 |
+
|
247 |
+
clear_pixels = np.sum(flat_mask == 0)
|
248 |
+
thick_cloud_pixels = np.sum(flat_mask == 1)
|
249 |
+
thin_cloud_pixels = np.sum(flat_mask == 2)
|
250 |
+
cloud_shadow_pixels = np.sum(flat_mask == 3)
|
251 |
+
total_pixels = flat_mask.size
|
252 |
+
|
253 |
+
stats = f"""
|
254 |
+
Cloud Mask Statistics:
|
255 |
+
- Clear: {clear_pixels} pixels ({clear_pixels/total_pixels*100:.2f}%)
|
256 |
+
- Thick Cloud: {thick_cloud_pixels} pixels ({thick_cloud_pixels/total_pixels*100:.2f}%)
|
257 |
+
- Thin Cloud: {thin_cloud_pixels} pixels ({thin_cloud_pixels/total_pixels*100:.2f}%)
|
258 |
+
- Cloud Shadow: {cloud_shadow_pixels} pixels ({cloud_shadow_pixels/total_pixels*100:.2f}%)
|
259 |
+
- Total Cloud Cover: {(thick_cloud_pixels + thin_cloud_pixels)/total_pixels*100:.2f}%
|
260 |
+
"""
|
261 |
+
|
262 |
+
# Visualize the cloud mask on the original image
|
263 |
+
visualization = visualize_cloud_mask(rgb_image, flat_mask)
|
264 |
+
|
265 |
+
return rgb_display, visualization, stats
|
266 |
+
|
267 |
+
|
268 |
+
# Create Gradio interface
|
269 |
+
demo = gr.Interface(
|
270 |
+
fn=process_satellite_images,
|
271 |
+
inputs=[
|
272 |
+
gr.Image(type="filepath", label="Red Channel (JP2)"),
|
273 |
+
gr.Image(type="filepath", label="Green Channel (JP2)"),
|
274 |
+
gr.Image(type="filepath", label="Blue Channel (JP2)"),
|
275 |
+
gr.Image(type="filepath", label="NIR Channel (JP2)"),
|
276 |
+
gr.Slider(minimum=1, maximum=32, value=1, step=1, label="Batch Size", info="Higher values use more memory but process faster"),
|
277 |
+
gr.Slider(minimum=500, maximum=2000, value=1000, step=100, label="Patch Size", info="Size of image patches for processing"),
|
278 |
+
gr.Slider(minimum=100, maximum=500, value=300, step=50, label="Patch Overlap", info="Overlap between patches to avoid edge artifacts")
|
279 |
+
],
|
280 |
+
outputs=[
|
281 |
+
gr.Image(label="Original RGB Image"),
|
282 |
+
gr.Image(label="Cloud Detection Visualization"),
|
283 |
+
gr.Textbox(label="Statistics")
|
284 |
+
],
|
285 |
+
title="Satellite Cloud Detection",
|
286 |
+
description="""
|
287 |
+
Upload separate JP2 files for Red, Green, Blue, and NIR channels to detect clouds in satellite imagery.
|
288 |
+
|
289 |
+
This application uses the OmniCloudMask model to classify each pixel as:
|
290 |
+
- Clear (0)
|
291 |
+
- Thick Cloud (1)
|
292 |
+
- Thin Cloud (2)
|
293 |
+
- Cloud Shadow (3)
|
294 |
+
|
295 |
+
The model works best with imagery at 10-50m resolution. For higher resolution imagery, downsampling is recommended.
|
296 |
+
""",
|
297 |
+
examples=[
|
298 |
+
["jp2s/B04.jp2", "jp2s/B03.jp2", "jp2s/B02.jp2", "jp2s/B8A.jp2", 1, 1000, 300]
|
299 |
+
]
|
300 |
+
)
|
301 |
+
# Launch the app
|
302 |
+
demo.launch(debug=True)
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
rasterio==1.3.11
|
2 |
+
matplotlib==3.7.5
|
3 |
+
fastai>=2.7
|
4 |
+
timm>=0.9
|
5 |
+
tqdm>=4.0
|
6 |
+
gdown>=5.1.0
|
7 |
+
torch>=2.2
|