MA/01

Deep Learning / Case study

Satellite Land Classification

CNN and Vision Transformer workflows for classifying satellite terrain in precision-agriculture scenarios.

View repository

Problem

Classify crops, forests, and water bodies from satellite imagery to support agricultural land-use decisions.

Solution

Built parallel Keras and PyTorch workflows covering image loading, augmentation, CNN development, Vision Transformer fine-tuning, and comparative evaluation.

Architecture

Geospatial images → framework-specific preprocessing and augmentation → CNN / ViT training paths → metric comparison.

Technical implementation

IBM AI Engineering capstone organized into four modules. Models are evaluated using accuracy, precision, recall, F1-score, and AU-ROC.

Key challenges

Compare convolutional and transformer approaches consistently across two deep-learning frameworks.

Lessons learned

The repository documents data-loading trade-offs, framework differences, transfer learning, and comparative model evaluation.