Image Recognition with EuroSAT Dataset¶

This notebook demonstrates how to train an image recognition (classification) model using the geoai.recognize module. The module provides a high-level, single-function API for training classifiers on ImageFolder-style datasets — directories of images organized by class.

Key Features¶

• Single-function training API: Pass a directory, get a trained model back
• 1000+ architectures: ResNet, EfficientNet, Vision Transformers, ConvNeXt, and more via timm
• Multi-format support: Works with JPEG, PNG, and multi-band GeoTIFF images
• Built-in evaluation: Classification reports, confusion matrices, and prediction visualization
• Transfer learning: Pretrained ImageNet weights with optional backbone freezing

Install packages¶

In [ ]:

# %pip install geoai-py

# %pip install geoai-py

Import libraries¶

In [ ]:

import os
from geoai.utils import download_file
from geoai.recognize import (
    load_image_dataset,
    train_image_classifier,
    predict_images,
    evaluate_classifier,
    plot_training_history,
    plot_confusion_matrix,
    plot_predictions,
)

import os from geoai.utils import download_file from geoai.recognize import ( load_image_dataset, train_image_classifier, predict_images, evaluate_classifier, plot_training_history, plot_confusion_matrix, plot_predictions, )

Download the EuroSAT RGB Dataset¶

The EuroSAT dataset contains 27,000 Sentinel-2 satellite image patches (64x64 pixels, RGB) in 10 land use/land cover classes:

• AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial
• Pasture, PermanentCrop, Residential, River, SeaLake

In [ ]:

url = "https://data.source.coop/opengeos/geoai/EuroSAT_RGB.zip"
download_dir = download_file(url)

url = "https://data.source.coop/opengeos/geoai/EuroSAT_RGB.zip" download_dir = download_file(url)

In [ ]:

data_dir = os.path.join(download_dir, "EuroSAT_RGB")
print(f"Dataset directory: {data_dir}")
print(f"Classes: {sorted(os.listdir(data_dir))}")

data_dir = os.path.join(download_dir, "EuroSAT_RGB") print(f"Dataset directory: {data_dir}") print(f"Classes: {sorted(os.listdir(data_dir))}")

Explore the Dataset¶

Use load_image_dataset to scan the directory and display the class distribution.

In [ ]:

dataset_info = load_image_dataset(data_dir)

dataset_info = load_image_dataset(data_dir)

In [ ]:

import matplotlib.pyplot as plt
from PIL import Image

class_names = dataset_info["class_names"]
image_paths = dataset_info["image_paths"]
labels = dataset_info["labels"]
class_to_idx = dataset_info["class_to_idx"]

fig, axes = plt.subplots(2, 5, figsize=(20, 8))

for idx, class_name in enumerate(class_names):
    ax = axes[idx // 5, idx % 5]
    # Find first image of this class
    img_idx = labels.index(class_to_idx[class_name])
    img = Image.open(image_paths[img_idx])
    ax.imshow(img)
    ax.set_title(class_name, fontsize=12)
    ax.axis("off")

plt.suptitle("Sample Image from Each Class", fontsize=14)
plt.tight_layout()
plt.show()

import matplotlib.pyplot as plt from PIL import Image class_names = dataset_info["class_names"] image_paths = dataset_info["image_paths"] labels = dataset_info["labels"] class_to_idx = dataset_info["class_to_idx"] fig, axes = plt.subplots(2, 5, figsize=(20, 8)) for idx, class_name in enumerate(class_names): ax = axes[idx // 5, idx % 5] # Find first image of this class img_idx = labels.index(class_to_idx[class_name]) img = Image.open(image_paths[img_idx]) ax.imshow(img) ax.set_title(class_name, fontsize=12) ax.axis("off") plt.suptitle("Sample Image from Each Class", fontsize=14) plt.tight_layout() plt.show()

Train a ResNet50 Classifier¶

The train_image_classifier function handles everything: scanning the directory, splitting into train/val/test, creating datasets, training, and evaluating.

In [ ]:

result = train_image_classifier(
    data_dir=data_dir,
    model_name="resnet50",
    num_epochs=5,
    batch_size=32,
    learning_rate=1e-3,
    image_size=64,
    in_channels=3,
    pretrained=True,
    output_dir="image_recognition_output/resnet50",
    num_workers=4,
    seed=42,
)

result = train_image_classifier( data_dir=data_dir, model_name="resnet50", num_epochs=5, batch_size=32, learning_rate=1e-3, image_size=64, in_channels=3, pretrained=True, output_dir="image_recognition_output/resnet50", num_workers=4, seed=42, )

Plot Training History¶

Visualize the training and validation loss and accuracy curves.

In [ ]:

fig = plot_training_history("image_recognition_output/resnet50/models")
plt.show()

fig = plot_training_history("image_recognition_output/resnet50/models") plt.show()

Evaluate on Test Set¶

Generate a classification report with precision, recall, and F1-score for each class.

In [ ]:

eval_result = evaluate_classifier(
    model=result["model"],
    dataset=result["test_dataset"],
    class_names=result["class_names"],
)

eval_result = evaluate_classifier( model=result["model"], dataset=result["test_dataset"], class_names=result["class_names"], )

Plot Confusion Matrix¶

In [ ]:

fig = plot_confusion_matrix(
    eval_result["confusion_matrix"],
    result["class_names"],
)
plt.show()

fig = plot_confusion_matrix( eval_result["confusion_matrix"], result["class_names"], ) plt.show()

In [ ]:

fig = plot_confusion_matrix(
    eval_result["confusion_matrix"],
    result["class_names"],
    normalize=True,
)
plt.show()

fig = plot_confusion_matrix( eval_result["confusion_matrix"], result["class_names"], normalize=True, ) plt.show()

Visualize Predictions¶

Show model predictions on test images with confidence scores. Green titles indicate correct predictions; red titles indicate misclassifications.

In [ ]:

test_dataset = result["test_dataset"]
test_paths = test_dataset.image_paths
test_labels = test_dataset.labels

pred_result = predict_images(
    model=result["model"],
    image_paths=test_paths[:20],
    class_names=result["class_names"],
    image_size=64,
    in_channels=3,
)

fig = plot_predictions(
    image_paths=test_paths[:20],
    predictions=pred_result["predictions"],
    true_labels=test_labels[:20],
    class_names=result["class_names"],
    probabilities=pred_result["probabilities"],
)
plt.show()

test_dataset = result["test_dataset"] test_paths = test_dataset.image_paths test_labels = test_dataset.labels pred_result = predict_images( model=result["model"], image_paths=test_paths[:20], class_names=result["class_names"], image_size=64, in_channels=3, ) fig = plot_predictions( image_paths=test_paths[:20], predictions=pred_result["predictions"], true_labels=test_labels[:20], class_names=result["class_names"], probabilities=pred_result["probabilities"], ) plt.show()

Train an EfficientNet-B0 Classifier¶

EfficientNet models provide an excellent balance between accuracy and efficiency. Let's compare it with ResNet50.

In [ ]:

result_effnet = train_image_classifier(
    data_dir=data_dir,
    model_name="efficientnet_b0",
    num_epochs=5,
    batch_size=32,
    learning_rate=1e-3,
    image_size=64,
    in_channels=3,
    pretrained=True,
    output_dir="image_recognition_output/efficientnet_b0",
    num_workers=4,
    seed=42,
)

result_effnet = train_image_classifier( data_dir=data_dir, model_name="efficientnet_b0", num_epochs=5, batch_size=32, learning_rate=1e-3, image_size=64, in_channels=3, pretrained=True, output_dir="image_recognition_output/efficientnet_b0", num_workers=4, seed=42, )

In [ ]:

eval_effnet = evaluate_classifier(
    model=result_effnet["model"],
    dataset=result_effnet["test_dataset"],
    class_names=result_effnet["class_names"],
)

eval_effnet = evaluate_classifier( model=result_effnet["model"], dataset=result_effnet["test_dataset"], class_names=result_effnet["class_names"], )

In [ ]:

fig = plot_confusion_matrix(
    eval_effnet["confusion_matrix"],
    result_effnet["class_names"],
    normalize=True,
)
plt.show()

fig = plot_confusion_matrix( eval_effnet["confusion_matrix"], result_effnet["class_names"], normalize=True, ) plt.show()

Compare Results¶

Compare the overall accuracy of both models.

In [ ]:

print(f"ResNet50 accuracy:       {eval_result['accuracy']:.4f}")
print(f"EfficientNet-B0 accuracy: {eval_effnet['accuracy']:.4f}")

print(f"ResNet50 accuracy: {eval_result['accuracy']:.4f}") print(f"EfficientNet-B0 accuracy: {eval_effnet['accuracy']:.4f}")

Summary¶

This notebook demonstrated:

1. Dataset download: Using download_file to fetch the EuroSAT RGB dataset
2. Single-function training: train_image_classifier handles splitting, dataset creation, training, and evaluation
3. Evaluation: Classification reports and confusion matrices with evaluate_classifier
4. Visualization: Training curves, confusion matrices, and prediction grids
5. Architecture comparison: Comparing ResNet50 and EfficientNet-B0

Key Parameters¶

• model_name: Choose from 1000+ timm models
• image_size: Resize images to this size (default 64 for EuroSAT)
• in_channels: 3 for RGB, 4+ for multispectral GeoTIFF
• pretrained: Use ImageNet pretrained weights for transfer learning
• freeze_backbone: Freeze backbone for faster fine-tuning
• num_epochs, batch_size, learning_rate: Standard training hyperparameters

Next Steps¶

• Try more architectures (ConvNeXt, Swin Transformer, ViT)
• Experiment with data augmentation transforms
• Use freeze_backbone=True for faster fine-tuning
• Apply to multi-band GeoTIFF datasets by setting in_channels > 3