aaronr43210/duk-feature-extraction

🛰️ DIGITAL UNIVERSITY KERALA FEATURE EXTRACTION MODEL V1 Feature Extraction: Unified AI Pipeline

Developed by Students of Digital University Kerala (DUK)

This repository contains a state-of-the-art AI solution developed for Problem Statement 1: Feature Extraction from Drone Images. Our pipeline is specifically designed to automate the extraction of high-precision geospatial features from DIGITAL UNIVERSITY KERALA FEATURE EXTRACTION MODEL V1 drone orthophotos, achieving a target accuracy of 95%.

🏆 Hackathon Solution Overview

Problem Statement: Feature Extraction from Drone Images (DIGITAL UNIVERSITY KERALA FEATURE EXTRACTION MODEL V1 Scheme) Objective: Develop an AI model capable of identifying key features in high-resolution orthophotos with high precision, optimized for efficient processing and deployment.

🎯 Key Features Extracted

• —Building footprint extraction: High-precision polygonal footprints.
• —Roof-top Classification: Automated classification into RCC, Tiled, Tin, and Others.
• —Road features: Continuous road network extraction (Polygons & Centerlines).
• —Waterbodies: Accurate delineation of ponds and rivers.
• —Point Feature Identification: Automated localization of Distribution Transformers and Wells.

🧠 Technical Architecture

1. Foundation Model Backbone

We utilize Segformer-B4 (Mix Transformer) pre-trained on high-resolution ImageNet and Cityscapes datasets as our core feature extractor. Unlike traditional CNNs, this Vision Transformer natively outputs 4 scale-agnostic feature maps, making it the industry standard for complex aerial semantics like "RCC Roof" vs "Tin Roof".

For a detailed breakdown of the model design, see architecture.md.

2. Multi-Head UPerFPN Decoder

Our custom Unified Perceptual Pyramid Network (UPerFPN) fuses the 4 multi-scale features from the Segformer encoder using Bilinear Interpolation and CBAM (Convolutional Block Attention Module) to handle objects of drastically varying sizes.

• —Segmentation Heads: Binary masks for Buildings, Roads, and Waterbodies.
• —Classification Head: Multi-class branch for Roof Categorization.
• —Connectivity Heads: Dilated convolution branches for linear features (roads/pipelines).
• —Point Detection: Specialized heads for localized utility features.

3. Advanced Post-Processing Pipeline

Our export module applies research-backed geometric refinement to every extracted feature layer:

• —Buildings: PolyMapper-style dominant angle Orthogonalization using SamGeo Feature Edge Reconstruction (FER) for mathematically perfect square/L-shaped structures.
• —Roads: Morphological closing (7px kernel) to bridge tree-canopy gaps; hole filling for continuous surfaces.
• —All Lines (Centerlines, Utilities): skan-based skeleton pruning → Chaikin corner-cutting smoothing → dead-end snapping for connected networks.
• —Waterbodies: Large morphological closing (9px) for smooth natural shorelines; convex hull for tiny ponds.
• —Point Features: YOLOv8 centroid extraction (no geometric post-processing needed).

5. Security & Production Hardening

• —Per-class adaptive confidence thresholds (instead of global 0.5)
• —Input file size validation (10 GB limit)
• —File extension whitelist validation
• —Output filename sanitization (path traversal prevention)
• —Pinned dependency versions for reproducible builds

🚀 Efficient Processing & Deployment

⚡ Optimization for Large-Scale Data

• —MMSegmentation Patterns: Loss functions utilizing Lovász Hinge and Online Hard Example Mining (Focal Loss) directly optimize IoU over massive orthophotos.
• —Intelligent Tiling: 512x512 tiling with 192px overlap to ensure NoData handling and seamless edge reconstruction.
• —Negative Sampling Quota: Automatically skips 99% empty tiles (farmland/forest) to focus GPU cycles on feature-rich areas.

💻 Deployment

A built-in Streamlit Dashboard (app.py) provides:

• —One-click processing for GeoTIFF orthophotos.
• —Real-time visualization of all 11 feature layers.
• —GIS Export: Direct download of vectorized .gpkg (GeoPackage) datasets ready for QGIS/ArcGIS.

🛠️ Usage Guidelines

Installation

git clone https://github.com/aaron43210/FEATURE.git
cd FEATURE
pip install -r requirements.txt

Training (Unified Pipeline)

The train.py script unifies the three-stage training process into a single command, automating dataset preparation and multi-model training.

# Train everything (Segmentation + YOLO)
# For 10 villages on DGX (8 GPUs), this takes ~8-12 hours
python train.py --train_dirs ./data/villages/ --epochs 150 --batch_size 8

# EMERGENCY RESUME (If DGX server crashes/restarts)
python train.py --resume --train_dirs ./data/villages/

Workflow:

1. 1.YOLO Prep: Converts shapefile points to YOLO bounding boxes and image tiles.
2. 2.Segmentation: Trains the Segformer+UPerFPN ensemble for buildings, roads, water, and roof types.
3. 3.YOLO Train: Trains the YOLOv8 point detector for high-precision utility localization.

For DGX environments, the script automatically leverages all available GPUs via DataParallel/DDP, routing data directly through the HuggingFace Segformer backbone.

Inference

streamlit run app.py

Developed with ❤️ by Digital University Kerala Students Committed to advancing the DIGITAL UNIVERSITY KERALA FEATURE EXTRACTION MODEL V1 scheme through innovative AI/ML techniques.