rishii100/galaxeye-tile-classifier

🛰️ GalaxEye Satellite Tile Land-Use Classifier

A fine-tuned ResNet-18 model for classifying 64×64 RGB satellite tiles into 7 land-use categories, designed for offline, edge-deployable inference on CPU.

Built as part of the GalaxEye Backend Engineer (ML Systems) take-home assignment.

---

📊 Model Summary

Property	Value
Architecture	ResNet-18 (ImageNet pretrained, fine-tuned)
Parameters	11.19M (FP32)
Input	64×64 RGB satellite tile (PNG)
Output	7-class probability distribution
Format	Hugging Face safetensors (zero-copy, mmap-safe)
Framework	PyTorch 2.2+
License	MIT

---

🏷️ Supported Classes

Label ID	Class	Description
0	AnnualCrop	Seasonal agricultural fields
1	Forest	Dense tree cover / woodland
2	Highway	Road infrastructure and corridors
3	Industrial	Factories, warehouses, industrial zones
4	Residential	Urban housing and neighborhoods
5	River	Waterways and river channels
6	SeaLake	Large water bodies (seas, lakes)

---

📈 Performance

Evaluated on a 210-tile holdout set (30 tiles per class), unseen during training:

Class	Accuracy
AnnualCrop	29/30 (96.7%)
Forest	28/30 (93.3%)
Highway	28/30 (93.3%)
Industrial	29/30 (96.7%)
Residential	30/30 (100.0%)
River	29/30 (96.7%)
SeaLake	28/30 (93.3%)
Overall	201/210 (95.71%)

---

🚀 Quick Start

Using PyTorch + Safetensors (Recommended)

import torch
import torch.nn as nn
import json
from torchvision import models, transforms
from safetensors.torch import load_file
from PIL import Image

# Load config and model
with open("config.json") as f:
    config = json.load(f)

classes = [config["id2label"][str(i)] for i in range(config["num_labels"])]

model = models.resnet18()
model.fc = nn.Linear(model.fc.in_features, len(classes))
model.load_state_dict(load_file("model.safetensors", device="cpu"))
model.eval()

# Preprocess and classify
transform = transforms.Compose([
    transforms.Resize((64, 64)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

image = Image.open("tile.png").convert("RGB")
tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    probs = torch.softmax(model(tensor), dim=1)[0]
    label = classes[probs.argmax().item()]
    confidence = probs.max().item()

print(f"Prediction: {label} ({confidence:.2%})")

Download from Hugging Face Hub

from huggingface_hub import hf_hub_download

# Download model artifacts
safetensors_path = hf_hub_download("rishii100/galaxeye-tile-classifier", "model.safetensors")
config_path = hf_hub_download("rishii100/galaxeye-tile-classifier", "config.json")

---

🏗️ Training Details

Property	Value
Base Model	torchvision.models.resnet18(pretrained=True)
Dataset	EuroSAT 7-class subset (1,050 tiles: 150 per class)
Train / Eval Split	840 train / 210 holdout (80-20 stratified)
Optimizer	Adam (lr=1e-4)
Loss	CrossEntropyLoss
Epochs	10
Augmentation	RandomHorizontalFlip, RandomRotation(15), ColorJitter
Hardware	NVIDIA T4 GPU (Google Colab)
Training Time	~2 minutes

---

🔒 Why Safetensors?

This model uses the safetensors format instead of Python pickle (.pth) for critical safety and performance benefits:

• 🛡️ No Arbitrary Code Execution: Unlike .pth files which use pickle (vulnerable to code injection), safetensors is a pure tensor serialization format with zero deserialization attack surface.
• ⚡ Zero-Copy Memory Mapping: Supports mmap for instant model loading without full deserialization — ideal for resource-constrained edge hardware.
• 📦 Framework Agnostic: Compatible across PyTorch, TensorFlow, JAX, and ONNX ecosystems.

---

🌍 Intended Use

This model is designed for offline satellite tile classification in isolated, internet-free environments such as:

• Edge base stations and field laboratories
• Satellite ground segment compute modules
• Disaster response and environmental monitoring units

It is part of a complete FastAPI-based inference service — see the GitHub repository for the full backend implementation.

---

⚠️ Limitations

• Resolution: Trained exclusively on 64×64 pixel tiles. Performance on other resolutions is not guaranteed.
• Class Coverage: Limited to 7 EuroSAT land-use categories. Novel terrain types (e.g., desert, glacier, wetland) will be misclassified.
• Geographic Bias: EuroSAT tiles are sourced from European Sentinel-2 imagery. Accuracy may degrade on tiles from other geographic regions with different spectral characteristics.
• Small Training Set: Only 840 training tiles (120 per class). Production deployments should fine-tune on larger, domain-specific datasets.

---

📎 Links

• 🔗 GitHub Repository: rishii100/galaxeye-tile-classifier
• 🏢 GalaxEye Space: galaxeye.space
• 📄 EuroSAT Dataset: Helber et al., 2019