This is a capstone project for the IBM AI Engineering Specialization, focusing on scalable deep learning for classifying agricultural versus non-agricultural land in gigapixel-scale satellite imagery (>100 GB datasets). It implements memory-efficient pipelines in Keras and PyTorch, progressing from CNN baselines to hybrid CNN-Vision Transformer (ViT) models that achieve high accuracy without excessive RAM or GPU demands.
Satellite imagery analysis enables precision agriculture and environmental monitoring, but large-scale datasets pose memory challenges. This project addresses them by developing streaming data loaders, on-the-fly augmentations, and hybrid architectures. Using ESA-derived satellite tiles, models classify land types with >94% validation accuracy and 0.984 AUC, optimizing for real-world deployment on limited hardware. The workflow covers data ingestion, training, evaluation (accuracy, precision, recall, F1, ROC-AUC), and ViT integration, demonstrating a complete end-to-end pipeline.
- Stream and augment >100 GB datasets using generators to reduce RAM usage by 5x.
- Train CNNs (ResNet-50 in Keras, EfficientNet-B3 in PyTorch) as baselines.
- Build CNN-ViT hybrids: Extract local features via CNN, capture global dependencies via ViT.
- Evaluate on 50k test tiles with rigorous metrics and visualizations (ROC curves, confusion matrices).
- Optimize for efficiency: Achieve ≥95% GPU utilization via caching, prefetching, and mixed precision.
| Technique | Keras Implementation | PyTorch Implementation | Impact |
|---|---|---|---|
| Generator Loading | tf.keras.utils.Sequence |
IterableDataset |
RAM ↓5x |
| Caching | .cache(filename) |
persistent_workers=True |
Epoch time ↓30% |
| Prefetching | tf.data.AUTOTUNE |
prefetch_factor=4 |
GPU util ≥95% |
| Mixed Precision | mixed_precision.set_global_policy('mixed_float16') |
torch.cuda.amp.autocast |
VRAM ↓50% |
| Augmentation | tf.image (flip, rotate, CutMix) |
albumentations |
No extra disk usage |
Validated on binary classification (agri/non-agri):
| Model | Framework | Parameters | Val Accuracy | AUC |
|---|---|---|---|---|
| ResNet-50 | Keras | 23M | 92.3% | 0.967 |
| EfficientNet-B3 | PyTorch | 47M | 93.1% | 0.972 |
| CNN-ViT Hybrid | PyTorch | 78M | 94.7% | 0.984 |
Hybrids excel in recall (0.999), minimizing false negatives for agricultural detection.
This project is licensed under the MIT License - see the LICENSE file for details.
Built for the IBM AI Engineering Professional Certificate. Dataset from ESA Sentinel-2. Code inspired by official Keras/PyTorch docs.
MIT © 2025 wusinyee – see LICENSE