Initial project situation:Diversification of the range of services to expand the target buyer groups

Supper & Supper is a service company for data science applications, but would like to further diversify its range of services in the future. Therefore, platform-specific products are to be developed in the future, which can be sold to several customers. Supper & Supper is currently focusing increasingly on the field of geo AI and has already built up a strong network of partners in this area. The first geo-related products are also to be developed. Supper & Supper has not yet carried out this type of product development in this form in the past.

Thanks to the established partner network and the broad selection of open source technologies, product development could be considered feasible within this framework. There is also a large amount of freely usable data that can be used for low-risk product development. There is also already experience in the areas of object recognition using TensorFlow and the processing of geodata using ArcGIS Pro software.

Objective of the project:Development of products using the TensorFlow open source framework

The project aims to create a basis for operationalizing artificial intelligence methods and putting them into practice. The aim is to develop a generic procedure for product development in which aerial photographs are used as the basis for generating geodata by means of object recognition. The aim is to develop products that can regularly generate business value. The TensorFlow open source framework and the GIS software ArcGIS Pro will be used for this purpose. Further objectives of the first pilot project are the acquisition of best practices and the development of a process checklist for a systematic approach to future product developments.

The first product to be developed in this context is an automatically generated cadastre for photovoltaic systems, including the assignment of exact addresses and evaluation of the size of the system. This product is to be realized using aerial photographs and a TensorFlow model for object recognition. The data basis for this pilot project are georeferenced aerial photographs from the Open NRW Portal.

Project development:Development and testing based on the photovoltaic systems cadastre pilot project

First, a process for developing and updating a geodata product was modeled based on literature and previous project experience in this area. This process was tested and documented in practice in the pilot project Photovoltaic Systems Cadastre. Several pre-trained neural networks were analyzed in the project and a suitable one was selected for further development. A total of 990 photovoltaic systems were labeled by hand as training data for the development. External data was also tested for inclusion in the training, but this led to a deterioration of the model.

Various methods for pre-processing and augmenting the images were evaluated and the model was then trained on the training data. When applied to a section of aerial images of the city of Bielefeld, a recall value of 84.4 % and a precision of 79.4 % were determined. The entire Bonn area was also processed with the trained model in order to obtain indications of scalability and performance on large amounts of data. The process checklist for the individual process phases was created during the project.

Added value for Supper and Supper GmbH:Development of the photovoltaic systems cadastre product

The modeled product development process was described in detail and successfully tested in practice in the pilot project. The pilot project resulted in the successful development of the photovoltaic plant register product. This and similar products can now be marketed by Supper & Supper. The developed standardized procedure for the development of a data product for geodata offers various possibilities for the development of a multitude of further solutions due to its flexible output. The developed process checklist can also contribute to smooth product development.

By working through the checklist before a new product development, all important decisions for the process should already be thought through. In this way, new products can be developed within a very short time window in future thanks to an effective and efficient approach. The process concept is transferable to a variety of use cases in almost all industries and for many potential customer groups of Supper & Supper. The technological expertise built up during the pilot project can also be used to establish processes for product developments outside the field of geo AI in the future.