As AI technology becomes more and more mature, the technology integration of “AI + GIS” has been applied in more and more practical projects. In order to meet the intelligent demand for deep learning in the geographic information industry, mapGIS 10.5 improves the whole process management of intelligent GIS.
Enhanced spatial information extraction capability of remote sensing images
With the improvement of remote sensing data collection resolution, in order to meet the development needs of intelligent and automatic interpretation of industry data, MapGIS intelligent GIS products are optimized, improved and enriched for deep learning intelligent application algorithm models.
In terms of model function improvement, it enriches the features of remote sensing information extraction and supports remote sensing features such as muitiband, vegetation index, and DSM. It enriches the ground objects supported by semantic segmentation model and adds new features to roads, water bodies, greenhouse, as well as the support of a wide range of villages in the city.
A new network algorithm model is introduced: a new land use all-factor segmentation model is added to segment the all-factor natural resources (as shown in the left figure below: 0 indicates the background, 1 artificial land, 2 agricultural land, 3 woodland, 4 Grassland); New object detection network model can detect ground targets such as vehicles and palm trees.
Image U-Net is one of the earlier neural network algorithms that use full convolution network for binary semantic segmentation. Due to its few parameters, it is easy to converge and has better semantic segmentation effect on small samples, therefore, it is widely used, but in terms of semantic segmentation of remote sensing images, only the native algorithm model is directly used. With the increase of iterations, the accuracy is difficult to improve. After its own optimization, MapGIS has added the Attention gate mechanism and resNet residual mechanism to improve the algorithm. The segmentation of remote sensing image data in a city has significantly improved the accuracy.
Improve data management capabilities
In machine learning, sample data is generally organized in the form of a file, which has the characteristics of various formats and file fragmentation. In order to solve the storage management of large amounts of small files, as well as future cloud-based needs, the MapGIS intelligent GIS sample dataset management module is built based on MapGIS DataStore distributed object storage services (hdfs, Minio, etc.), hosting massive sample datasets into distributed storage, use its scale-out feature to store and manage massive sample small file datasets.
For the management of multi-user multi-model training datasets on the cloud, metadata management capabilities are provided. By extending descriptions of datasets, such as descriptions of application attributes such as adding attribution categories and usage scenarios to datasets, it can quickly and effectively retrieve data sets when facing hundreds of data sets, making the data sets more hierarchical and logical.
Provides python secondary development capabilities
MapGIS intelligent GIS products have built-in MapGIS Objects Python space machine learning development library and provide interfaces related to the full process development of space machine learning.
Provides data set preprocessing related interfaces, mainly providing training data set production, sample standardization conversion, image processing enhancement and other functional interfaces, such as format conversion, random mask, random rotation angle, and Radom noise.
Provides rich deep learning algorithm models, mainly providing conventional neural network models such as alexnet, densenet, googlenet, inception, mnasnet, resnet, shufflenet, squeezenet, vgg, etc, provides Native and optimized models such as classify, unet, fasterrcnn, and maskrcnn.
Provides deep learning algorithm model training, conversion, prediction and other related interfaces, which are mainly used in development scenarios such as algorithm model training to improve accuracy, desktop and mobile model conversion capabilities, meet the application between different platforms.
Provide data science capabilities
Data science is that data scientists use data to assist decision-making and solve practical problems. If they want to complete complex and complicated data collection, storage, analysis and processing, with the help of statistical tools and programming languages, Python, R and other languages are popular. Traditional data science has applied many powerful tools and algorithms, however, there are not many methods and tools for spatial data processing. MapGIS intelligent GIS products inherit the advantages of MapGIS in spatial geography and provide tools for spatial data processing for data science.
MapGIS intelligent GIS products integrate Jupyter, provide integrated python development environment and Notebook tools, provide powerful technical tool support for users in the field of data science, and are mainly used for interactive construction of prototype models by multiple users on the Web, write shared model code for exploratory data analysis and collaborative development scenarios.
Built-in mainstream TensorFlow, PyTorch, Caffe, and Keras deep learning frameworks in the Notebook integrated development environment, built-in MapGIS spatial machine learning library, and NumPy, Scikit-Learning, rasterio, OpenCV, Proj4, matplotlib and other third-party python data processing, machine learning, and data visualization libraries support not only geographic data, but also various non-geographic data types such as videos, text, and images, it can be used in data analysis and mining, deep learning, machine learning and other data-driven application scenarios, providing users with convenient interactive computing design tools, making data scientific analysis easier.