geotree

🌳 GeoTree — Deep Learning Tree Crown Detector for Satellite Imagery

GeoTree is a production-grade Convolutional Neural Network with Residual Blocks optimized for automated tree crown detection, land cover analysis, and carbon/biomass estimation from high-resolution satellite imagery (Sentinel-2, PlanetScope, drone orthomosaics) across Bangladesh.

---

🖼️ Sample Model Prediction & Detection Output

---

📈 Training Progress & Metrics History (40 Epochs)

<details open> <summary><b>📋 View Full 40-Epoch Training Metrics Progression</b></summary> </details>

---

📊 Final Performance Benchmarks & Confusion Matrix

<div align="center">

</div>

---

🗺️ Geographic Training & Validation Data Distribution

<details> <summary><b>📐 Bounding Box Coordinate Accuracy (MAE & RMSE)</b></summary> </details>

---

🚀 Model Specifications

• Model Name: geotree
• Architecture: Residual ConvNet (TreeDetectorModel) with Batch Normalization & SiLU
• Loss Function: Complete IoU (CIoU) Loss (10.0× weight) + BCE Logits Loss
• Input Dimensions: 640x640 RGB / Multispectral tiles

---

💻 Quickstart Inference Code

import torch
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download
from model import TreeDetectorModel

# 1. Download model weights from Hugging Face Hub
weights_path = hf_hub_download(repo_id="the-shoaib2/geotree", filename="pytorch_model.bin")

# 2. Instantiate and load model
model = TreeDetectorModel()
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

# 3. Load & preprocess image
img = Image.open("sample_tile.png").convert("RGB").resize((640, 640))
img_tensor = torch.tensor(np.array(img).transpose(2, 0, 1).astype(np.float32) / 255.0).unsqueeze(0)

# 4. Predict
with torch.no_grad():
    output = model(img_tensor).squeeze(0)
    conf = torch.sigmoid(output[0]).item()
    bbox = output[1:].tolist()

print(f"Tree Detected: {conf > 0.3} | Confidence: {conf:.2f} | BBox: {bbox}")

🌍 Applications

GeoTree can be used in a wide range of environmental, forestry, agricultural, and urban monitoring applications:

• Tree detection and counting
• Tree crown detection
• Forest cover mapping
• Tree density estimation
• Deforestation monitoring
• Reforestation tracking
• Land cover classification
• Vegetation health analysis (NDVI, NDWI, NDMI)
• Carbon and biomass estimation
• Plantation and orchard monitoring
• Urban green space analysis
• Water body detection
• Grassland and bare soil detection
• Disaster impact assessment (fire, flood, cyclone)
• Change detection using multi-temporal satellite imagery
• Protected forest and wildlife habitat monitoring
• Smart city environmental planning
• GIS-based environmental reporting and analytics