pokkiri's picture
Update app.py
eb80d64 verified
"""
Biomass Prediction Gradio App with Exact 99 Features
Author: najahpokkiri
Date: 2025-05-19
Updated with side-by-side RGB comparison, fixed sample image loading, and corrected biomass calculation.
"""
import os
import sys
import torch
import numpy as np
import gradio as gr
import joblib
import tempfile
import matplotlib.pyplot as plt
import matplotlib.colors as colors
from PIL import Image
import io
import logging
from huggingface_hub import hf_hub_download
# Configure logger
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
# Import model architecture
from model import StableResNet
# Import feature engineering
from feature_engineering import extract_all_features
# Import config - this must happen before loading model_package.pkl
try:
from config import BiomassPipelineConfig
logger.info("Successfully imported config.BiomassPipelineConfig")
except ImportError as e:
logger.error(f"Failed to import config.BiomassPipelineConfig: {e}")
logger.error("This will likely cause errors when loading the model package")
class BiomassPredictorApp:
"""Gradio app for biomass prediction from satellite imagery"""
def __init__(self, model_repo="pokkiri/biomass-model"):
"""Initialize the app with model repository information"""
self.model = None
self.package = None
self.feature_names = []
self.model_repo = model_repo
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Cache for storing temporary files
self.temp_files = []
# Load the model
self.load_model()
def load_model(self):
"""Load the model and preprocessing pipeline from HuggingFace Hub"""
try:
logger.info(f"Loading model from {self.model_repo}")
# Download model files from HuggingFace
model_path = hf_hub_download(repo_id=self.model_repo, filename="model.pt")
package_path = hf_hub_download(repo_id=self.model_repo, filename="model_package.pkl")
try:
# Try to load package with metadata
logger.info(f"Loading package from {package_path}")
self.package = joblib.load(package_path)
logger.info("Successfully loaded model package")
# Extract information from package
n_features = self.package['n_features']
self.feature_names = self.package.get('feature_names', [f"feature_{i}" for i in range(n_features)])
logger.info(f"Package keys: {list(self.package.keys())}")
logger.info(f"Model expects {n_features} features")
# Verify feature count is 99
if n_features != 99:
logger.warning(f"Warning: Model expects {n_features} features, not the expected 99. This may cause issues.")
except Exception as e:
logger.error(f"Error loading package file: {e}")
# Fallback to default values
n_features = 99 # We know there are 99 features
self.feature_names = [f"feature_{i}" for i in range(n_features)]
# Create a minimal package with essential components
self.package = {
'n_features': n_features,
'use_log_transform': True,
'epsilon': 1.0,
'scaler': None # Will handle the None case in prediction
}
# Initialize model
self.model = StableResNet(n_features=n_features)
self.model.load_state_dict(torch.load(model_path, map_location=self.device))
self.model.to(self.device)
self.model.eval()
logger.info(f"Model loaded successfully from {self.model_repo}")
logger.info(f"Number of features: {n_features}")
logger.info(f"Using device: {self.device}")
logger.info(f"Log transform: {self.package.get('use_log_transform', True)}")
logger.info(f"Epsilon: {self.package.get('epsilon', 1.0)}")
return True
except Exception as e:
logger.error(f"Error loading model: {e}")
import traceback
logger.error(traceback.format_exc())
return False
def cleanup(self):
"""Clean up temporary files"""
for tmp_path in self.temp_files:
try:
if os.path.exists(tmp_path):
os.unlink(tmp_path)
except Exception as e:
logger.warning(f"Failed to remove temporary file {tmp_path}: {e}")
self.temp_files = []
def load_sample_image(self):
"""Load the sample image and return a file-like object"""
try:
sample_path = "input_chip_1.tif"
if os.path.exists(sample_path):
logger.info(f"Loading sample image from {sample_path}")
return sample_path
else:
logger.warning(f"Sample image not found at {sample_path}")
return None
except Exception as e:
logger.error(f"Error loading sample image: {e}")
return None
def predict_biomass(self, image_file, display_type="heatmap"):
"""Predict biomass from a satellite image"""
if self.model is None:
return None, "Error: Model not loaded. Please check logs for details."
if image_file is None:
return None, "Error: No file uploaded. Please upload a GeoTIFF file or use the sample image."
try:
# Check if we're using the sample image (string path) or an uploaded file
if isinstance(image_file, str):
logger.info(f"Using sample image: {image_file}")
tmp_path = image_file # Use the sample path directly
cleanup_tmp = False # Don't delete the sample file
else:
# Create a temporary file to save the uploaded file
with tempfile.NamedTemporaryFile(suffix='.tif', delete=False) as tmp_file:
tmp_path = tmp_file.name
with open(image_file.name, 'rb') as f:
tmp_file.write(f.read())
# Add to list for cleanup later
self.temp_files.append(tmp_path)
cleanup_tmp = True
# Ensure rasterio is available
try:
import rasterio
except ImportError:
return None, "Error: rasterio is required but not installed. Please install with: pip install rasterio"
# Open the image file
with rasterio.open(tmp_path) as src:
image = src.read()
height, width = image.shape[1], image.shape[2]
transform = src.transform
crs = src.crs
# Check if we need to limit to 59 bands
if image.shape[0] > 59:
logger.info(f"Image has {image.shape[0]} bands, selecting first 59 for model compatibility")
image = image[:59, :, :]
logger.info(f"Processing image: {height}x{width} pixels, {image.shape[0]} bands")
# Validate minimum band count
if image.shape[0] < 1:
return None, f"Error: Image has no bands. Please use multi-band satellite imagery."
# Generate all features using feature engineering
logger.info("Generating all 99 features from bands...")
feature_matrix, valid_mask, generated_features = extract_all_features(image)
# Print basic feature statistics for debugging
logger.info(f"Feature statistics - Min: {np.min(feature_matrix, axis=0)[:5]}, " +
f"Max: {np.max(feature_matrix, axis=0)[:5]}, " +
f"Mean: {np.mean(feature_matrix, axis=0)[:5]}")
# Verify we have exactly 99 features
if feature_matrix.shape[1] != 99:
logger.error(f"Error: Generated {feature_matrix.shape[1]} features, but model expects 99.")
return None, f"Error: Generated {feature_matrix.shape[1]} features, but model expects 99."
# Apply feature scaling if available
try:
if 'scaler' in self.package and self.package['scaler'] is not None:
logger.info("Applying feature scaling...")
feature_matrix = self.package['scaler'].transform(feature_matrix)
logger.info("Scaling complete")
logger.info(f"After scaling - Min: {np.min(feature_matrix, axis=0)[:5]}, " +
f"Max: {np.max(feature_matrix, axis=0)[:5]}")
except Exception as e:
logger.warning(f"Error applying scaler: {e}. Using original features.")
# Initialize predictions array
predictions = np.zeros((height, width), dtype=np.float32)
# Get valid pixel coordinates
valid_y, valid_x = np.where(valid_mask)
# Make predictions
logger.info(f"Running model inference on {len(valid_y)} valid pixels...")
with torch.no_grad():
# Process in batches to avoid memory issues
batch_size = 10000
for i in range(0, len(valid_y), batch_size):
end_idx = min(i + batch_size, len(valid_y))
batch = feature_matrix[i:end_idx]
# Convert to tensor
batch_tensor = torch.tensor(batch, dtype=torch.float32).to(self.device)
# Get predictions
batch_predictions = self.model(batch_tensor).cpu().numpy()
# Handle scalar case for single-item batches
if batch_predictions.ndim == 0:
batch_predictions = np.array([batch_predictions])
# Log raw predictions
if i == 0:
logger.info(f"Raw prediction sample: {batch_predictions[:5]}")
# Fix: Correct log transform reversal
if self.package.get('use_log_transform', True):
# Get epsilon value, default to 1.0
epsilon = self.package.get('epsilon', 1.0)
# Log transform should be exp(x) - epsilon
batch_predictions = np.exp(batch_predictions)
# Only subtract epsilon if it's not zero or close to zero
if abs(epsilon) > 1e-10:
batch_predictions = batch_predictions - epsilon
# Ensure non-negative
batch_predictions = np.maximum(batch_predictions, 0)
# Log transformed predictions
if i == 0:
logger.info(f"Transformed prediction sample: {batch_predictions[:5]}")
logger.info(f"Using log transform: {self.package.get('use_log_transform', True)}, " +
f"epsilon: {self.package.get('epsilon', 1.0)}")
# Map predictions back to image
for j, pred in enumerate(batch_predictions):
y_idx = valid_y[i + j]
x_idx = valid_x[i + j]
predictions[y_idx, x_idx] = pred
# Log progress
if (i // batch_size) % 5 == 0 or end_idx == len(valid_y):
logger.info(f"Processed {end_idx}/{len(valid_y)} pixels")
# Calculate and log prediction statistics
valid_predictions = predictions[valid_mask]
logger.info(f"Prediction statistics - Min: {np.min(valid_predictions):.2f}, " +
f"Max: {np.max(valid_predictions):.2f}, " +
f"Mean: {np.mean(valid_predictions):.2f}, " +
f"Median: {np.median(valid_predictions):.2f}")
# Create visualization
logger.info("Creating visualization...")
if display_type == "heatmap":
# Create heatmap visualization
fig, ax = plt.subplots(figsize=(10, 8))
# Use masked array for better visualization
masked_predictions = np.ma.masked_where(~valid_mask, predictions)
# Set min/max values based on percentiles for better contrast
vmin = np.percentile(predictions[valid_mask], 1)
vmax = np.percentile(predictions[valid_mask], 99)
im = ax.imshow(masked_predictions, cmap='viridis', vmin=vmin, vmax=vmax)
fig.colorbar(im, ax=ax, label='Biomass (Mg/ha)')
ax.set_title('Predicted Above-Ground Biomass')
ax.axis('off') # Hide axes for cleaner visualization
elif display_type == "rgb_overlay":
# Create side-by-side comparison (RGB and Biomass)
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 8))
# Prepare RGB image using bands 4,3,2 (0-indexed: 3,2,1)
rgb_bands = [3, 2, 1] # Using 4,3,2 for RGB (0-indexed)
if image.shape[0] >= 5: # Ensure we have enough bands (need at least 5 for 0-indexed band 4)
# Create RGB image
rgb = np.zeros((height, width, 3), dtype=np.float32)
for i, band_idx in enumerate(rgb_bands):
if band_idx < image.shape[0]:
rgb[:, :, i] = image[band_idx]
# Handle potential NaN values
rgb = np.nan_to_num(rgb)
# Enhance contrast with percentile-based normalization
for i in range(3):
p2 = np.percentile(rgb[:,:,i], 2)
p98 = np.percentile(rgb[:,:,i], 98)
if p98 > p2:
rgb[:,:,i] = np.clip((rgb[:,:,i] - p2) / (p98 - p2), 0, 1)
# Display RGB image
ax1.imshow(rgb)
ax1.set_title('RGB Image (Bands 4,3,2)')
ax1.axis('off')
# Display biomass prediction
masked_predictions = np.ma.masked_where(~valid_mask, predictions)
vmin = np.percentile(predictions[valid_mask], 1)
vmax = np.percentile(predictions[valid_mask], 99)
im = ax2.imshow(masked_predictions, cmap='viridis', vmin=vmin, vmax=vmax)
fig.colorbar(im, ax=ax2, label='Biomass (Mg/ha)')
ax2.set_title('Predicted Biomass')
ax2.axis('off')
# Add super title
plt.suptitle('RGB Image and Biomass Prediction', fontsize=16)
plt.tight_layout()
else:
# Fallback to heatmap if not enough bands
logger.warning(f"Not enough bands for RGB display (need 5, got {image.shape[0]}). Showing biomass only.")
masked_predictions = np.ma.masked_where(~valid_mask, predictions)
im = ax1.imshow(masked_predictions, cmap='viridis')
fig.colorbar(im, ax=ax1, label='Biomass (Mg/ha)')
ax1.set_title('Predicted Above-Ground Biomass')
ax1.axis('off')
# Save figure to bytes buffer
buf = io.BytesIO()
fig.savefig(buf, format='png', dpi=150, bbox_inches='tight')
buf.seek(0)
plt.close(fig)
# Calculate summary statistics
valid_predictions = predictions[valid_mask]
stats = {
'Mean Biomass': f"{np.mean(valid_predictions):.2f} Mg/ha",
'Median Biomass': f"{np.median(valid_predictions):.2f} Mg/ha",
'Min Biomass': f"{np.min(valid_predictions):.2f} Mg/ha",
'Max Biomass': f"{np.max(valid_predictions):.2f} Mg/ha"
}
# Add area and total biomass if transform is available
if transform is not None:
pixel_area_m2 = abs(transform[0] * transform[4]) # Assuming square pixels
total_biomass = np.sum(valid_predictions) * (pixel_area_m2 / 10000) # Convert to hectares
area_hectares = np.sum(valid_mask) * (pixel_area_m2 / 10000)
stats['Total Biomass'] = f"{total_biomass:.2f} Mg"
stats['Area'] = f"{area_hectares:.2f} hectares"
# Format statistics as markdown
stats_md = "### Biomass Statistics\n\n"
stats_md += "| Metric | Value |\n|--------|-------|\n"
for k, v in stats.items():
stats_md += f"| {k} | {v} |\n"
# Add processing info
stats_md += f"\n\n*Processed {np.sum(valid_mask):,} valid pixels with {feature_matrix.shape[1]} features*"
# Cleanup temporary files if needed
if cleanup_tmp:
self.cleanup()
# Return visualization and statistics
return Image.open(buf), stats_md
except Exception as e:
# Ensure cleanup even on error
self.cleanup()
import traceback
logger.error(f"Error predicting biomass: {e}")
logger.error(traceback.format_exc())
return None, f"Error predicting biomass: {str(e)}\n\nPlease check logs for details."
def create_interface(self):
"""Create Gradio interface"""
with gr.Blocks(title="Biomass Prediction Model") as interface:
gr.Markdown("# Above-Ground Biomass Prediction")
gr.Markdown("""
Upload a multi-band satellite image to predict above-ground biomass (AGB) across the landscape.
**Requirements:**
- Image must be a GeoTIFF with spectral bands
- For best results, use imagery with at least 59 bands or similar to training data
""")
with gr.Row():
with gr.Column(scale=1):
input_image = gr.File(
label="Upload Satellite Image (GeoTIFF)",
file_types=[".tif", ".tiff"]
)
display_type = gr.Radio(
choices=["heatmap", "rgb_overlay"],
value="heatmap",
label="Display Type"
)
with gr.Row():
submit_btn = gr.Button("Generate Biomass Prediction", variant="primary")
sample_btn = gr.Button("Use Sample Image")
with gr.Column(scale=2):
output_image = gr.Image(
label="Biomass Prediction Map",
type="pil"
)
output_stats = gr.Markdown(
label="Statistics"
)
with gr.Accordion("About", open=False):
gr.Markdown("""
## About This Model
This biomass prediction model uses the StableResNet architecture to predict above-ground biomass from satellite imagery.
### Model Details
- Architecture: StableResNet
- Input: Multi-spectral satellite imagery
- Output: Above-ground biomass (Mg/ha)
- Creator: vertify.earth for GIZ Forest Forward
- Date: 2025-05-19
### How It Works
1. The model extracts features from each pixel in the satellite image
2. These features include spectral bands, vegetation indices, texture metrics, and more
3. The model outputs a biomass prediction for each pixel
4. Results are visualized as a heatmap or RGB overlay
### Updates in This Version
- Fixed biomass value calculation issue (improved log transform handling)
- Added detailed diagnostics for troubleshooting
- Enhanced RGB visualization with band verification
""")
# Add a warning if model failed to load
if self.model is None:
gr.Warning("⚠️ Model failed to load. The app may not work correctly. Check logs for details.")
# Connect the submit button
submit_btn.click(
fn=self.predict_biomass,
inputs=[input_image, display_type],
outputs=[output_image, output_stats]
)
# Handle sample image button
def use_sample_image(display_type):
sample_path = self.load_sample_image()
if sample_path is None:
return None, "Error: Sample image not found. Please make sure 'input_chip_1.tif' exists in the app directory."
return self.predict_biomass(sample_path, display_type)
sample_btn.click(
fn=use_sample_image,
inputs=[display_type],
outputs=[output_image, output_stats]
)
return interface
def launch_app():
"""Launch the Gradio app"""
try:
# Create app instance
app = BiomassPredictorApp()
# Create interface
interface = app.create_interface()
# Launch interface - Important: no share=True in Hugging Face Spaces
interface.launch()
except Exception as e:
logger.error(f"Error launching app: {e}")
import traceback
logger.error(traceback.format_exc())
if __name__ == "__main__":
launch_app()