🧠 Overview
Our project focuses on detecting non-existent or misplaced Points of Interest (POIs) by leveraging satellite imagery and geospatial data. Using a CNN trained on manually labeled satellite images, we identified inconsistencies in the placement of POIs relative to digitised routes.
🚀 What it does
- 🔄 Data Integration: Merged diverse datasets from different formats using
GeoPandas. - 📍 POI Segmentation: Grouped POIs based on geolocation, orientation, and proximity to route networks.
- 🛰️ Satellite Image Extraction: Retrieved area-specific images via the HERE API.
- 🧠 Deep Learning-based Prediction: Trained a custom CNN to detect misplaced or non-existent POIs.
- 🌐 Multi-scenario Handling: Designed logic to classify and address different misplacement cases.
- 🖥️ Interactive UI: Built a visual tool to inspect results and predictions effectively.
🔧 How we built it
- Geospatial preprocessing and cleaning using
GeoPandas - Geometrical computations for defining areas of interest
- Manual annotation of satellite imagery for training the CNN
- CNN built from scratch using
PyTorch - Developed a user-friendly interface with filtering and insights
🧩 Challenges faced
- Limited feature availability from open datasets
- Complexity of geospatial reasoning
- High volume of data and API rate limits
- Handling inconsistencies in POI sources
🏆 Key accomplishments
- Devised a novel geometric approach to infer POI-route relationships
- Built a convolutional neural network from the ground up
- End-to-end pipeline: from data acquisition to UI deployment
- Tackled real-world GIS problems with scalable solutions
What we learned
- Datasets are valuable
- Geographical routing problems are hard
- Usefulness of algorithms and mathematics for resource intensive problems
Log in or sign up for Devpost to join the conversation.