Satellite_Image-Visual_Search

Training a ML model that performs General Visual Search and Retrieval from Satellite Images given sample image chip of the target.

Kaggle Notebook -> click here

Model Training Pipeline

Siamese Network Model for Satellite Object Detection

Overview

Few-shot learning approach using Siamese networks to detect objects in satellite imagery with minimal training examples (3-5 per class). Achieved 3156 detections through metric learning rather than traditional object detection.

Architecture

Model Structure

• Input: 64×64 pixel chips with 4 spectral bands (Blue, Green, Red, Near-Infrared)
• Backbone: ResNet50 pre-trained on ImageNet
• Embedding Dimension: 128-dimensional normalized vectors
• Output: L2-normalized embeddings for similarity comparison

Key Components

1. 4→3 Channel Projection: 1×1 convolution converts satellite bands to RGB-compatible format
2. ResNet50 Backbone: Transfer learning from ImageNet for feature extraction
3. Embedding Head: Dense layers (2048→512→128) with unit normalization

Training Process

Data Preparation

• Chip Extraction: 64×64 patches with 15% padding around labeled objects
• Background Class: Random sampling from non-object regions for negative learning
• Normalization: Per-band mean/std normalization with auto-detected scale factor
• Augmentation: Random horizontal/vertical flips and 90° rotations

P×K Batch Sampling

• P_CLASSES = 8: 8 different classes per batch
• K_SAMPLES = 3: 3 examples per class
• Batch Size = 24: Balanced positive/negative pairs for triplet learning

Training Strategy

1. Phase 1 - Warmup (5 epochs): Freeze ResNet50, train embedding head only
2. Phase 2 - Fine-tuning (145 epochs): Unfreeze conv5 block + embedding head
3. Optimizer: AdamW with learning rate 1e-4 and weight decay 1e-4
4. Early Stopping: Patience of 10 epochs on validation loss

Loss Function: Batch-Hard Triplet Loss

• Concept: For each anchor, find hardest positive (furthest same-class) and hardest negative (closest different-class)
• Margin: 0.5 safety buffer between positive and negative distances
• Formula: loss = softplus(d(A,P) - d(A,N) + margin)
• Mining: Automatic hard pair selection within each batch

Performance Optimization

Mixed Precision Training

• Policy: float16 for most operations, float32 for critical computations
• Benefit: 2x faster training without accuracy loss

Calibration

• Purpose: Convert cosine similarities to detection probabilities
• Method: Logistic regression on validation pairs (same-class vs different-class)
• Threshold: 0.3 probability for detection confidence

Inference Pipeline

Prototype Generation

• Average embeddings from 3-5 exemplar chips per class
• Creates representative "fingerprint" for each object type

Sliding Window Search

• Stride: 32 pixels (50% overlap for comprehensive coverage)
• Batch Processing: 64 patches simultaneously for efficiency
• Similarity Computation: Dot product between patch embeddings and class prototype
• Non-Maximum Suppression: Remove overlapping detections with IoU threshold

Key Advantages

• Data Efficient: Works with 3-5 examples per class vs 1000+ for traditional methods
• Flexible: Add new classes without retraining the entire model
• Generalizable: Learns similarity rather than class-specific features
• Fast Training: 2-3 hours vs 3-5 days for conventional object detection

Technical Specifications

• Framework: TensorFlow with Keras API
• Hardware: P100 GPU in Kaggle
• Output Format: Bounding boxes with confidence scores for submission in a csv format