Revolutionizing Land Classification with AI
deepTerra is an innovative AI application designed to simplify land classification from remote sensing data, including drone footage and high spatial resolution satellite imagery. It streamlines the machine learning process into four key aspects:
Data Collection and Organization
deepTerra can automatically source, download, and organize satellite image patches by simply specifying the latitude/longitude coordinates of a 1km land square.
Image Augmentation
The application employs geometric image augmentation techniques to enhance datasets, thereby improving the accuracy and generalization of the machine learning process.
Model Training
deepTerra supports a variety of open-source deep learning models, including ResNet50, VGG16, DenseNet, InceptionV3, InceptionResNet, and a proprietary convolutional network model tailored for specific dataset categories. It offers a wide range of tunable hyperparameters for optimizing training speed, accuracy, and generalizability. A new version has been released that includes sophisticated automatic hypertuning features that automatically search and optimize hyperparameter values without manual intervention
Image Prediction and Data Export
Once trained, models can be applied to new image sets for classification, with rich support for prediction visualization and filtering through interactive heatmaps. The latest release introduces ensemble learning, combining multiple models and statistical techniques to generate more accurate and reliable predictions.
Results can be visualized directly on Google Maps or exported in a variety of formats, including Google Earth KML files, JSON, CSV, and templated documents such as LaTeX or PDF.
Key Points
deepTerra is designed to be user-friendly, requiring no coding or scripting. It comes with several pre-trained datasets and can be applied to general image classification tasks beyond land classification. For instance, it has been trained on a dataset of 25,000 cat and dog images and can distinguish between cat and dog images with over 98% accuracy.
Future Developments
Development of deepTerra is ongoing, with plans to support multi-class applications, use historical data for land change mapping, target relevant parts of images using hotspots, extract heat characteristics using multispectral data, and apply advanced models for incremental learning as new data becomes available.
Screenshots
The following show several screenshots that give a flavour of deepTerra and its main functionalities.
The start screen contains buttons that link to the four main aspects of image based machine learning: image collection, augmentation, supervised training, and prediction
Data collection consists of downloading, preprocessing and organizing images so that they are suitable for training and/or test and prediction. deepTerra supports downloading images either from local folders or tarred-zip files or by downloading tarred-zip files from the internet.
Additionally, and perhaps most powerfully, the user can specify the top-left latitude/longitude coordinate and the tool will use the google earth API to download and preprocess the high resolution images for a 1km square area - resulting in 1296 images which can be used for training, testing and prediction.
A partial set of images downloaded from google Earth.
If only a small set of images are available for training (which is often the case) then various transformations can be applied to each image in order to make variations on each one. This process, known as "image augmentation", can greatly increase the number of images available and improve the accuracy and generalization of the resulting models.
The core of the machine learning process is refered to as "training". We use supervised training where examples of each category of interest are pre-labelled. The machine learning algorithms can then use this labelled dataset to learn the various characeristic for each category.
The ultimate goal of all these processes is to be able to take new images not seen before, and from the learnt models to make predictions of how each image should be categorized. This screen shows the raw results of such a prediction run.
it's possible to see details of a prediction run where the categorization for each image is shown, together with an associated level of confidence. If possible, a google maps link is provided that highlights the approximate location of each image.
The "positive" results for a prediction run can be viewed as an automatically generated google map, with the areas of interest overlaid with semi-transparent red rectangles.
The "positive" results for a larger scale prediction, automatically collected from a 1km square area, comprising 1296 prediction patches with positive results overlaid in red.
Technical Overview
A paper describing the capabilities and architecture of deepTerra has been uploaded to both ResearchGate and arXiv.