Friday 07 March 2025
The quest for a comprehensive platform to classify land surface features using machine learning and satellite imagery has finally taken shape. deepTerra, a new tool, offers an integrated suite of modules that streamline the entire workflow for image classification tasks. This system is designed to simplify the process of collecting, organizing, augmenting, training, testing, and predicting images.
The platform’s core strength lies in its data collection module, which allows users to import and organize image patches from various sources. These can be local folders, compressed archives, or even remote download links. The system also provides a convenient way to label datasets, which is crucial for model development. This feature is particularly useful when dealing with large datasets that require manual annotation.
Another key aspect of deepTerra is its image augmentation module. Data augmentation is a technique used to expand the size of a dataset by applying various transformations, such as rotation, shifting, and zooming, to existing images. This helps to introduce variations in the training data, which can improve model robustness and performance. The platform supports several standard geometric augmentation techniques, making it easy for users to customize their datasets.
The training module is where deepTerra really shines. Users can select from a range of popular convolutional neural network (CNN) architectures, including ResNet50, InceptionV3, and DenseNet. The system also provides sensible default settings for hyperparameters, allowing users to fine-tune them according to their specific needs.
Once the model is trained, it’s time to test its performance on an independently collected dataset. deepTerra’s testing module provides detailed metrics, such as accuracy and F1-score, giving users a clear understanding of how well their model generalizes to new data.
The platform also includes a prediction module, which allows users to apply their trained models to novel, unlabeled datasets. The results can be filtered by confidence level or significance, enabling more targeted analysis. Users can even manipulate the predictions to focus on specific criteria, such as filtering out low-confidence predictions.
To demonstrate deepTerra’s capabilities, researchers have applied it to various real-world scenarios, including garbage dump detection, private swimming pool identification, and beehive localization. In one notable example, they used deepTerra to identify garbage dumps in Cyprus with impressive accuracy. The system was able to detect garbage patches in 94,288 images, covering nearly a quarter of the surveyed area.
The success of deepTerra is not limited to satellite imagery.
Cite this article: “DeepTerra: A Comprehensive Platform for Machine Learning-Based Image Classification”, The Science Archive, 2025.
Machine Learning, Satellite Imagery, Image Classification, Deep Learning, Convolutional Neural Networks, Data Augmentation, Computer Vision, Geospatial Analysis, Remote Sensing, Land Surface Features.
Reference: Andrew Keith Wilkinson, “deepTerra — AI Land Classification Made Easy” (2025).
