| --- |
| license: mit |
| tags: |
| - remote-sensing |
| - satellite-imagery |
| - multi-label-classification |
| - pytorch |
| - efficientnetv2 |
| - mobilenetv2 |
| datasets: |
| - planet-understanding-the-amazon-from-space |
| metrics: |
| - fbeta_score |
| --- |
| |
| # 🛰️ Planet Amazon Satellite Scene Multi-Label Classifier |
|
|
| Multi-label classification of remote sensing satellite images over the Amazon rainforest using **MobileNetV2** (Baseline) and **EfficientNetV2-S** (Proposed Novel Architecture). |
|
|
| ## 📊 Performance Benchmarks |
| - **Dataset:** Planet Amazon (40,479 images, 17 classes) |
| - **Macro Recall:** 98.72% (EfficientNetV2-S) |
| - **Official Kaggle Public Score:** 0.60475 |
| - **Official Kaggle Private Score:** 0.59929 (Zero Overfitting, Δ = 0.0054) |
|
|
| ## 🏗️ Architecture Innovations |
| - Backbone: `EfficientNetV2-S` with Fused-MBConv layers |
| - Loss: Label Smoothing Binary Cross-Entropy (ε = 0.05) |
| - Schedule: Cosine Annealing Learning Rate |
| - Activation: GELU non-linearities in 2-stage classification head |
|
|
