Artificial intelligence (AI) describes the capacity of computers to simulate human intelligence. AI attempts to mimic human reasoning (and action) by digital information assimilation and processing.
Cognitive processes that are relevant for AI are knowledge representation, automated reasoning and machine learning (ML). Deep learning (DL) is a subfield of ML and provides deep neural network (DNN) algorithms that are a powerful tool for computer vision domains.
In Earth observation (EO), DNNs have been successfully applied in image preprocessing, feature detection, image classification, and change detection. The DL approach uses sample data to label image objects and then to train a DNN on its capability to recognize different types of land cover. The trained DNN is then applied to a full set of remote sensing images to map the land cover in the covered area of interest. A share of sample data independent from the training data is then used for validation of the output. Typical applications of DNN include automatic extraction of buildings, of built-up areas and their change and the mapping of land cover, agricultural fields or crop types. Despite being relatively new to EO, DNN algorithms now are among the top performers in most of the applications.
Potential & Requirements
EO processing platforms like CODE-DE enable the exploitation of the vast archives of pre-processed multi-temporal and multi-sensor satellite observations, especially when they are supported with equally large processing capabilities for AI applications. For any DNN to undertake any intended operation (e.g. classification and regression), it is often required to perform a huge computational process that demands a powerful physical or virtual computing environment equipped with fast processing units such as they are available with Graphical Processing Units (GPUs).
New CODE-DE capabilities
CODE-DE users now have access to GPU-accelerated Virtual Machines. In response to high computational requirements of machine learning-based processing chains - that cannot be feasibly executed using CPUs - CODE-DE has introduced two new types of Graphical Processing Units to its infrastructure. Both the Nvidia A100 GPU and the Nvidia RTX A6000 graphics card are top-of-the-line tools in their respective categories. Nvidia A100 is currently the most powerful GPU for data centers, with architecture dedicated for AI and High Performance Computing (HPC) applications. It allows users to execute most demanding tasks. Whereas Nvidia RTX A6000 offers perfect capabilities for workflow prototyping and diverse real-world applications.
The FAQ section on CODE-DE provides manuals on how to enable GPU. A user can access the GPU capabilities by requesting a quota in the dashboard of CODE-DE. Once the quota is granted, it is possible to set up a virtual machine that includes the GPUs in its processing environment. The virtual machine is equipped with a programming environment that hosts interactive code development tools like Jupyter notebooks, DL frameworks like Tensorflow and Pytorch and other third party libraries. This enables easy development and execution of scripts to process EO data on CODE-DE and provide application relevant digital maps.
GPUs are available to CODE-DE users via a "passthrough" mechanism, meaning that users do not share a selected GPU with others. This mechanism also assures that the hypervisor overhead for computational operations is virtually eliminated. Consequently, the fullest capacity of the selected GPU is available to the user. GPU accelerated Virtual Machines can also be equipped with very fast SSD (NVMe) storage to guarantee sufficient data flow during operations. GPU accelerated Virtual Machines are available in two new flavours: One is based on the Nvidia A100 series, the other on Nvidia RTX A6000 series. Both are equipped with 24 vCPU and 118 GB of RAM.
To support the efficient use of this highly innovative DL environment for various AI applications, the CODE-DE Team collaborates with EO experts of the Department of Geoinformatics at the University Salzburg in generating training material and documented use cases.