Скачать или смотреть GeoPython 2024: Enabling and optimizing image-based AI workflows in constrained environment

Gillian Milani
Enabling and optimizing image-based AI workflows in constrained environment

The software used to develop the AI model of the FSO for land cover mapping include multiple pieces enabling their training in a particular environment. The integration of python libraries like h5py or optuna can be replicated in any AI project for improving data and resource use.

The FSO is responsible for producing land cover and land use statistics. An artificial intelligence tool allowing partial automation of the survey has been developed in order to reduce the duration of the survey from 9 to 6 years. The introduction of this tool has already reduced the number of points interpreted by 27%, while maintaining accuracy of over 98%.

Needless to say, the IT environment like those of many companies and institution worldwide, constraints greatly software development. This talk will focus on three points that enabled the development of AI model within a public administration environment:
1. the management of large image library within HDF5 files using the h5py library.
2. the use of the optuna library as a simple communication channel across devices.
3. a set of tools used and adapted to keep track of training trials on the long-term.

We believe that these approaches, in particular the HDF5 packaging, can find many uses across different projects and are worth to be shared with the community. The talk will not be a deep dive into our models and their performances.