Vishalpainjane's picture
Update app.py
22c89e3 verified
import gradio as gr
import torch
import torch.nn as nn
from PIL import Image
import numpy as np
import pandas as pd
import matplotlib
import matplotlib.pyplot as plt
import segmentation_models_pytorch as smp
import albumentations as A
from albumentations.pytorch import ToTensorV2
import os
import random
from datetime importdatetime
# --- Best Practice: Set Matplotlib backend for server environments ---
matplotlib.use('Agg')
# --- CONFIGURATION (UPDATED FOR DEPLOYMENT) ---
class CFG:
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
# CRITICAL: Use relative paths for deployment.
# Place your model file in the root of your Hugging Face Space repository.
MODEL_PATH = "best_model_optimized_83.98.pth"
# The app will scan this local folder for example images.
EXAMPLES_DIR = "examples"
MODEL_NAME = "CustomDeepLabV3+"
ENCODER_NAME = "timm-efficientnet-b2"
NUM_CLASSES = 8
IMG_SIZE = 256
# Constants for area calculation
ORIGINAL_PATCH_DIM = 64
RESOLUTION_M_PER_PIXEL = 10
SQ_METERS_PER_HECTARE = 10000
TOTAL_PATCH_AREA_HECTARES = (ORIGINAL_PATCH_DIM**2 * RESOLUTION_M_PER_PIXEL**2) / SQ_METERS_PER_HECTARE
# --- DATA & CLASS INFO ---
CLASS_INFO = {
0: {"name": "Unclassified", "hex": "#969696"}, 1: {"name": "Water Bodies", "hex": "#0000FF"},
2: {"name": "Dense Forest", "hex": "#006400"}, 3: {"name": "Built up", "hex": "#800080"},
4: {"name": "Agriculture land", "hex": "#00FF00"}, 5: {"name": "Barren land", "hex": "#FFFF00"},
6: {"name": "Fallow land", "hex": "#D2B48C"}, 7: {"name": "Sparse Forest", "hex": "#3CB371"},
}
# --- MODEL DEFINITION (REFORMATTED FOR READABILITY) ---
class SELayer(nn.Module):
def __init__(self, channel, reduction=16):
super(SELayer, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
nn.Linear(channel, channel // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channel // reduction, channel, bias=False),
nn.Sigmoid()
)
def forward(self, x):
b, c, _, _ = x.size()
y = self.avg_pool(x).view(b, c)
y = self.fc(y).view(b, c, 1, 1)
return x * y.expand_as(x)
class CustomDeepLabV3Plus(nn.Module):
def __init__(self, encoder_name, in_channels, classes):
super().__init__()
self.smp_model = smp.DeepLabV3Plus(
encoder_name=encoder_name,
encoder_weights="imagenet",
in_channels=in_channels,
classes=classes
)
decoder_channels = self.smp_model.segmentation_head[0].in_channels
self.se_layer = SELayer(decoder_channels)
self.segmentation_head = self.smp_model.segmentation_head
self.smp_model.segmentation_head = nn.Identity()
def forward(self, x):
decoder_features = self.smp_model(x)
attended_features = self.se_layer(decoder_features)
output = self.segmentation_head(attended_features)
return output
# --- MODEL LOADING & TRANSFORMS ---
def load_model():
print(f"Loading model from {CFG.MODEL_PATH} on device {CFG.DEVICE}...")
model = CustomDeepLabV3Plus(encoder_name=CFG.ENCODER_NAME, in_channels=3, classes=CFG.NUM_CLASSES)
if not os.path.exists(CFG.MODEL_PATH):
raise FileNotFoundError(f"CRITICAL: Model file not found at '{CFG.MODEL_PATH}'. Please ensure the model file is in the root directory of your Space.")
# Using weights_only=True is safer
model.load_state_dict(torch.load(CFG.MODEL_PATH, map_location=torch.device(CFG.DEVICE), weights_only=True))
model.to(CFG.DEVICE)
model.eval()
print("Model loaded successfully!")
return model
model = load_model()
transform = A.Compose([
A.Resize(height=CFG.IMG_SIZE, width=CFG.IMG_SIZE),
A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
# --- HELPER & ANALYSIS FUNCTIONS ---
def create_color_map():
color_map = np.zeros((256, 3), dtype=np.uint8)
for class_id, info in CLASS_INFO.items():
color_map[class_id] = tuple(int(info['hex'].lstrip('#')[i:i+2], 16) for i in (0, 2, 4))
return color_map
COLOR_MAP_NUMPY = create_color_map()
def create_colored_mask(mask_np):
return Image.fromarray(COLOR_MAP_NUMPY[mask_np])
def analyze_one_image(image_filepath: str):
if image_filepath is None: return None, {}
image = Image.open(image_filepath)
image_np = np.array(image.convert("RGB"))
transformed = transform(image=image_np)
input_tensor = transformed['image'].unsqueeze(0).to(CFG.DEVICE)
with torch.no_grad():
prediction = model(input_tensor)
pred_mask = torch.argmax(prediction.squeeze(), dim=0).cpu().numpy()
area_results = {}
class_indices, pixel_counts = np.unique(pred_mask, return_counts=True)
total_pixels_in_mask = pred_mask.size
for class_id, count in zip(class_indices, pixel_counts):
if class_id in CLASS_INFO:
pixel_proportion = count / total_pixels_in_mask
area_hectares = pixel_proportion * CFG.TOTAL_PATCH_AREA_HECTARES
area_results[CLASS_INFO[class_id]["name"]] = area_hectares
return pred_mask, area_results
def single_image_analysis(image_filepath: str):
if image_filepath is None: raise gr.Error("Please upload an image to analyze.")
pred_mask_np, areas_dict = analyze_one_image(image_filepath)
pred_mask_pil = create_colored_mask(pred_mask_np)
area_data = sorted(areas_dict.items(), key=lambda item: item[1], reverse=True)
area_df = pd.DataFrame(area_data, columns=["Land Cover Class", "Area (Hectares)"])
area_df["Area (Hectares)"] = area_df["Area (Hectares)"].map('{:.4f}'.format)
analysis_results = {"areas": areas_dict, "area_df": area_df, "image_path": image_filepath}
return pred_mask_pil, area_df, analysis_results
def compare_land_cover(filepath1: str, filepath2: str):
if filepath1 is None or filepath2 is None:
raise gr.Error("Please upload both a 'Before' and 'After' image for comparison.")
_, areas1_dict = analyze_one_image(filepath1)
_, areas2_dict = analyze_one_image(filepath2)
mask1_pil = create_colored_mask(analyze_one_image(filepath1)[0])
mask2_pil = create_colored_mask(analyze_one_image(filepath2)[0])
all_class_names = sorted(list(set(areas1_dict.keys()) | set(areas2_dict.keys())))
data_for_df = [[name, areas1_dict.get(name, 0), areas2_dict.get(name, 0)] for name in all_class_names]
df = pd.DataFrame(data_for_df, columns=["Class", "Area 1 (ha)", "Area 2 (ha)"])
df['Change (ha)'] = df['Area 2 (ha)'] - df['Area 1 (ha)']
df['% Change'] = df.apply(lambda row: (row['Change (ha)'] / row['Area 1 (ha)'] * 100) if row['Area 1 (ha)'] > 0 else float('inf'), axis=1)
df_display = df.copy()
for col in ["Area 1 (ha)", "Area 2 (ha)"]: df_display[col] = df_display[col].map('{:.2f}'.format)
df_display["Change (ha)"] = df_display["Change (ha)"].map('{:+.2f}'.format)
df_display["% Change"] = df_display["% Change"].apply(lambda x: f"{x:+.2f}%" if x != float('inf') else "New")
plt.style.use('seaborn-v0_8-whitegrid')
fig, ax = plt.subplots(figsize=(10, 6))
index = np.arange(len(df))
bar_width = 0.35
ax.bar(index - bar_width/2, df['Area 1 (ha)'], bar_width, label='Area 1 (Before)', color='cornflowerblue')
ax.bar(index + bar_width/2, df['Area 2 (ha)'], bar_width, label='Area 2 (After)', color='salmon')
ax.set_xlabel('Land Cover Class', fontweight='bold')
ax.set_ylabel('Area (Hectares)', fontweight='bold')
ax.set_title('Land Cover Change Analysis', fontsize=16, fontweight='bold')
ax.set_xticks(index)
ax.set_xticklabels(df['Class'], rotation=45, ha="right")
ax.legend()
fig.tight_layout()
analysis_results = {"df": df_display, "path1": filepath1, "path2": filepath2, "raw_df": df}
return mask1_pil, mask2_pil, df_display, fig, analysis_results
# --- REPORTING FUNCTIONS ---
def generate_report(analysis_results, report_type):
if not analysis_results:
raise gr.Error("Please run an analysis first before generating a report.")
if report_type == "single":
filename = os.path.basename(analysis_results['image_path'])
report = f"# LULC Analysis Report: {filename}\n"
report += f"**Date:** {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n\n"
report += "## Area Distribution (Hectares)\n"
report += analysis_results['area_df'].to_markdown(index=False)
elif report_type == "change":
file1 = os.path.basename(analysis_results['path1'])
file2 = os.path.basename(analysis_results['path2'])
df = analysis_results['raw_df']
summary = ""
df_sorted = df.reindex(df['Change (ha)'].abs().sort_values(ascending=False).index)
for _, row in df_sorted.head(3).iterrows():
if abs(row['Change (ha)']) > 0.01:
direction = "increased" if row['Change (ha)'] > 0 else "decreased"
summary += f"- **{row['Class']}** has {direction} by **{abs(row['Change (ha)']):.2f} hectares**.\n"
report = f"# LULC Change Detection Report\n"
report += f"**Comparison:** `{file1}` (Before) vs. `{file2}` (After)\n"
report += f"**Date:** {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n\n"
report += "## Key Summary of Changes\n"
report += summary + "\n"
report += "## Detailed Comparison Table\n"
report += analysis_results['df'].to_markdown(index=False)
# Switch to the report tab and populate it
return {
report_editor: gr.update(value=report),
download_btn: gr.update(visible=True),
tabs: gr.update(selected=2)
}
def save_report_to_file(report_content):
filepath = "LULC_Report.md"
with open(filepath, "w", encoding="utf-8") as f:
f.write(report_content)
return filepath
# --- EXAMPLE FINDER ---
def find_examples():
single_examples = []
change_examples = []
if os.path.isdir(CFG.EXAMPLES_DIR):
files = sorted([os.path.join(CFG.EXAMPLES_DIR, f) for f in os.listdir(CFG.EXAMPLES_DIR) if f.lower().endswith(('.png', '.jpg', '.jpeg', '.tif'))])
single_examples = files[:10] # Take up to 10 for single analysis
# Create pairs for change detection
if len(files) >= 2:
for i in range(0, min(len(files) - 1, 10), 2): # Take up to 5 pairs
change_examples.append([files[i], files[i+1]])
return single_examples, change_examples
single_examples, change_examples = find_examples()
# --- GRADIO UI LAYOUT ---
with gr.Blocks(theme=gr.themes.Soft(), title="LULC Analysis Platform") as demo:
gr.Markdown("# Land Use & Land Cover (LULC) Analysis Platform")
gr.Markdown("An AI-powered tool to analyze satellite imagery for environmental monitoring and planning.")
# Hidden state objects to store analysis results robustly
single_analysis_results = gr.State()
change_analysis_results = gr.State()
with gr.Tabs() as tabs:
with gr.TabItem("Single Image Analysis", id=0):
with gr.Row(variant="panel"):
with gr.Column(scale=1):
single_img_input = gr.Image(type="filepath", label="Upload Satellite Image")
single_analyze_btn = gr.Button("Analyze Image", variant="primary")
with gr.Column(scale=1):
single_mask_output = gr.Image(type="pil", label="Predicted Mask")
with gr.Row():
area_df_output = gr.DataFrame(label="Predicted Area Distribution", wrap=True)
send_single_report_btn = gr.Button("➡ Create Report from this Analysis")
gr.Examples(examples=single_examples, inputs=single_img_input, label="Click an Example to Start")
with gr.TabItem("Change Detection Tool", id=1):
with gr.Row(variant="panel"):
compare_img1 = gr.Image(type="filepath", label="Image 1 (e.g., Before / 2020)")
compare_img2 = gr.Image(type="filepath", label="Image 2 (e.g., After / 2024)")
compare_analyze_btn = gr.Button("Analyze Changes", variant="primary")
with gr.Row():
compare_mask1 = gr.Image(type="pil", label="Mask for Image 1")
compare_mask2 = gr.Image(type="pil", label="Mask for Image 2")
with gr.Tabs():
with gr.TabItem("📊 Change Chart"): compare_plot = gr.Plot()
with gr.TabItem("📑 Comparison Table"): compare_df = gr.DataFrame(interactive=False)
send_change_report_btn = gr.Button("➡ Create Report from this Analysis")
if change_examples:
gr.Examples(examples=change_examples, inputs=[compare_img1, compare_img2], label="Click an Example Pair to Start")
with gr.TabItem("Report Builder", id=2):
gr.Markdown("### Create and Download Your Analysis Report")
gr.Markdown("1. Run an analysis on one of the other tabs.\n"
"2. Click the **'➡ Create Report'** button.\n"
"3. Your report will appear below. You can edit it before downloading.\n")
with gr.Column():
report_editor = gr.Textbox(label="Your Report (Editable)", lines=20, interactive=True)
download_btn = gr.DownloadButton(label="Download Report (.md)", visible=False)
# --- BUTTON CLICK EVENTS & DATA FLOW ---
# Single Image Analysis Flow
single_analyze_btn.click(
fn=single_image_analysis,
inputs=single_img_input,
outputs=[single_mask_output, area_df_output, single_analysis_results]
).then(
lambda: gr.update(interactive=False, value="Analyzing..."), None, single_analyze_btn
).then(
lambda: gr.update(interactive=True, value="Analyze Image"), None, single_analyze_btn
)
send_single_report_btn.click(
fn=lambda res: generate_report(res, "single"),
inputs=single_analysis_results,
outputs=[report_editor, download_btn, tabs]
)
# Change Detection Flow
compare_analyze_btn.click(
fn=compare_land_cover,
inputs=[compare_img1, compare_img2],
outputs=[compare_mask1, compare_mask2, compare_df, compare_plot, change_analysis_results]
).then(
lambda: gr.update(interactive=False, value="Analyzing..."), None, compare_analyze_btn
).then(
lambda: gr.update(interactive=True, value="Analyze Changes"), None, compare_analyze_btn
)
send_change_report_btn.click(
fn=lambda res: generate_report(res, "change"),
inputs=change_analysis_results,
outputs=[report_editor, download_btn, tabs]
)
# Report Download Flow
download_btn.click(fn=save_report_to_file, inputs=report_editor, outputs=download_btn)
if __name__ == "__main__":
demo.launch(debug=True)