ONNX export of cm93/resnet50-eurosat, a ResNet-50 model fine-tuned on the EuroSAT satellite imagery dataset for land-use and land-cover (LULC) classification.

Model Details

Property	Value
Architecture	ResNet-50 (ImageNet-1k pretrained)
Task	Image Classification
Classes	10
Format	ONNX
Source model	cm93/resnet50-eurosat
Source library	timm
License	MIT

Classes

Index	Label
0	AnnualCrop
1	Forest
2	HerbaceousVegetation
3	Highway
4	Industrial
5	Pasture
6	PermanentCrop
7	Residential
8	River
9	SeaLake

Note: Index order reflects id2label from the source model config. Always load it via AutoConfig.from_pretrained("cm93/resnet50-eurosat") rather than assuming alphabetical order.

Evaluation Results

Evaluated on the cm93/eurosat test split. Preprocessing applied using the source model's TimmWrapperImageProcessor via AutoImageProcessor.from_pretrained("cm93/resnet50-eurosat").

Metric	Score
Accuracy	0.9670
Precision (macro)	0.9662
Recall (macro)	0.9669
F1 (macro)	0.9662

Source model reported metrics (from cm93/resnet50-eurosat):

Split	Accuracy	F1
Train	0.9756	—
Validation	0.9830	—
Test	0.9807	0.9807

Usage

Installation

pip install onnxruntime huggingface_hub transformers

Inference

python

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from transformers import AutoImageProcessor, AutoConfig
from PIL import Image

# Load ONNX session
model_path = hf_hub_download("krylox/resnet50-eurosat-onnx", filename="model.onnx")
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])

# Load processor and label map from the source model
processor = AutoImageProcessor.from_pretrained("cm93/resnet50-eurosat")
id2label  = AutoConfig.from_pretrained("cm93/resnet50-eurosat").id2label

# Preprocess and infer
image = Image.open("your_image.jpg").convert("RGB")
pixel_values = processor(images=image, return_tensors="pt")["pixel_values"].numpy().astype(np.float32)

input_name  = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name
logits = session.run([output_name], {input_name: pixel_values})[0]

predicted_label = id2label[int(np.argmax(logits, axis=1)[0])]
print(f"Predicted class: {predicted_label}")

Important: Use AutoImageProcessor.from_pretrained("cm93/resnet50-eurosat") for preprocessing — the source model uses TimmWrapperImageProcessor, which may differ from standard ImageNet normalization. Using hardcoded stats will degrade accuracy significantly.

Training Details (Source Model)

Parameter	Value
Base weights	ResNet-50 pretrained on ImageNet-1k
Optimizer	AdamW (β₁=0.9, β₂=0.999)
Learning rate	3e-5
Weight decay	0.01
Batch size	64
Scheduler	PolynomialLR
Loss	CrossEntropyLoss
Epochs	20
Augmentation	RandomHorizontalFlip, RandomVerticalFlip
Preprocessing	Resize → CenterCrop → Normalize

Training Data

The source model was trained on EuroSAT, which contains 27,000 labeled, geo-referenced Sentinel-2 satellite image patches (JPEG, 64×64 px) across 10 LULC classes.

Intended Uses

Land-use / land-cover classification from Sentinel-2 imagery
Remote sensing research and benchmarking
Deployment in non-PyTorch or resource-constrained environments via ONNX Runtime

Limitations & Biases

Trained exclusively on European Sentinel-2 imagery; performance may degrade on other sensors or non-European regions.
The 10-class taxonomy is coarse and may not capture fine-grained land-use distinctions.
Inherits any biases present in the source checkpoint cm93/resnet50-eurosat.

License

MIT