How vasanthv07/geospatial-object-detection works
Project type and purpose
• This is a machine‐learning research project for semantic segmentation (“Geospatial Object Detection using Aerial Imagery”) built primarily as a Jupyter notebook pipeline.
• It ingests an annotated aerial imagery dataset, applies preprocessing (Min-Max scaling, patchify), trains a custom Multi U-Net model in TensorFlow/Keras using Focal+D ice losses, evaluates with the Jaccard coefficient, and produces segmentation outputs for downstream analysis.
File structure
• Top‐level directories:
• Certificates, Paper, Presentation, Report → documentation/artifacts
• Dataset → raw imagery tiles (images + masks) + classes.json
• Single Notebook “Object_Detection_using_Aerial_Imagery.ipynb” houses the entire data‐prep → training → evaluation logic.
• No separate backend/frontend or microservices—this is a linear ML pipeline.
Key insights from the README
• Technical stack: Python, Jupyter, NumPy, Pandas, Matplotlib, Keras, TensorFlow
• Data: 72 images in 8 tiles with 6 semantic classes
• Processing steps and metrics are explicitly described (Patchify, Min-Max scaling, Focal + Dice loss, Jaccard index)
a. Main components
• Data Storage (Dataset folder with raw images, masks, classes.json)
• Preprocessing Module (patchify, scaling, any augmentation)
• Model Definition (Multi U-Net layers, custom loss functions)
• Training Orchestrator (Notebook controlling epochs/batches)
• Evaluation Module (computing Jaccard, generating prediction images)
• Artifacts Store (exported model files on Google Drive)
• External Services (Colab runtime, GitHub repo)
b. Relationships and data flows
c. Architectural patterns and design principles
• Pipeline pattern: sequential data‐prep → train → evaluate → store
• Modular stages: isolate data ingestion, preprocessing, model logic, evaluation
• Reproducibility: versioned dataset, notebook code, model artifacts
d. Key technologies/frameworks
• Python + Jupyter Notebook (control flow)
• TensorFlow/Keras (model)
• NumPy/Pandas (data manipulation)
• Matplotlib (visualization)
• Google Colab & Drive (compute, artifact storage)