SuperMap AI GIS Technology- GeoAI Workflow

We have introduced the AI GIS Technology, the basic theories and algorithms of GeoAI before. We are going to see how GeoAI works in this article.

GeoAI workflow include data acquisition and preparation, model building and management, and model publishing and application.

Figure GeoAI Workflow

1.     Data acquisition and preparation

Taking deep learning as an example, the dataset that needs to be prepared contains image data and label data for the specific location of the feature you want to identify. For example, to identify buildings, you need the image of the target area and the label of the vector outline of the building, and then generate training data in the corresponding format.

Figure Data acquisition and preparation

1.1 Data Labeling

Data labeling generally includes raw image data and vector labeling data. Take economic tree identification as an example:

•  The image data are high-resolution aerial image or satellite image. Here are the image requirements:

Table Image Data Requirements

•  Vector labeling data is spatial polygon data with attribute information (labels). Users can use the GIS desktop to draw rectangular outlines of image object and assign attribute labels.

Figure SuperMap iDesktopX Generate Vector Label Data

labeling information is generally used to describe the content description of different functions. For example, for object detection (aircraft), the labeling information needs to identify whether it is an aircraft, the pixel coordinates of the aircraft, and the aircraft type.

1.2 Training Data Generation

Depending on the size of the training data, SuperMap provides training tools include SuperMap iDesktopX and SuperMap iServer DataScience services. The latter is more suitable for the large-scale training data generation in batches. Based on the original remote sensing image data and vector labeling data, the service is generated training data used by deep learning.

Figure Training Data Production

For the training sample data, it is necessary to generate the training sample data and train the model through the basic data provided by the user, so as to ensure the accuracy of the model as much as possible.

Figure Training Sample Data

2.     Model construction and management

Model construction and management is to train the neural network model based on the training data samples generated in the previous data preparation process, and at the same time, iteratively evaluate the training model through the validation dataset and test dataset to achieve the actual application accuracy and precision requirements. For example, SuperMap provides object detection of image based on Faster-RCNN model. User can also choose a more suitable deep learning network model for object detection according to the specific application scenario. Since the model training process involves complex numerical calculations, it is recommended to use a server environment that supports GPU computing.

Figure Model Training Tools

SuperMap has built-in open source frameworks such as TensorFlow and Keras. It combines different dataset categories to train machine learning or deep learning models. The overall training process uses multiple iterations (epoch) to obtain network models with better training results. The training model uses a pre-trained model based on large-scale basic training data to reduce training time and improve model training efficiency and accuracy through hyperparameter tuning (learning rate, Batchsize, etc.). According to the length of the model training cycle, SuperMap also provides SuperMap iDesktopX that support short-period model training and SuperMap iServer DataScience training tools that support long-period model training.

Figure SuperMap iDesktopX Model Training Tool (Short-period)

Figure SuperMap Service Training Tool (Long-period)

Currently SuperMap supports multiple network models including U-Net and Faster R-CNN.

The trained model is used to output the results through the validation dataset and the test dataset. At the same time, we perform quantitative statistical verification on the analysis results to evaluate the accuracy of the model application and provide a reference for practical business applications. Finally, the training engine outputs model files and model definition files.

Currently SuperMap supports a variety of model files for different training frameworks, such as TensorFlow's * .pb format. SuperMap encapsulates model files of different AI frameworks into a unified model file * .sdm, which is compatible with multiple frameworks, multiple network structures, and multiple functions, and at the same time shields the complex AI framework concept, making it easier for users manage, iterate, and update models.

3.     Model release and application

Service release is achieved through the SuperMap iServer machine learning module, providing functional coverage including object detection, binary classification, land use classification, and scene classification.

Figure SuperMap Model Application (Target Detection) Results

Figure SuperMap Model Application (Binary Classification) Results

Figure Original Data

Figure Service Release of Model Application

We are looking for distributors, resellers, and partners all over the world. For any inquiry, please contact us at: [email protected].