As part of the AI4GEO project (http://ai4geo.io), industrial and institutional partners have converged on the creation of products on a global scale valuing 2D and 3D spatial data, namely:
- Time series of land cover maps on a large scale and at very high resolution using jointly radiometric and altimetric information.
- Creation of time series of LOD1 and LOD2 models in order to monitor urban development.
- Creation of semantic 3D models calculated from a mesh of classified 3D point clouds (stereo and lidar)
- Detection and characterization of changes in series of 3D point clouds (stereo and lidar) for applications related to autonomous navigation and monitoring of railway tracks.
In this context, several axes of research have been identified which can lead to work with a strong scientific impact:
- Denoising of Digital Surface Models: Digital surface models and point clouds at the output of the photogrametric restitution chain from multi-view satellite images are often very noisy and this noise has an impact on the prediction quality of subsequent models (machine learning and deep learning). This line of research will consist in studying 2.5D or 3D denoising methods. Several solutions are possible: AI and statistical approaches.
- Multi-scale 3D classification: OBIA 2.5D and 3D classification on several semantic scales using contextual information (spatial and temporal). Today, producing an object classification map is largely feasible, but using the spatial arrangement of these objects among themselves to deduce meta-objects (set of objects having semantic meaning, for example a group of houses aligned with swimming pools and gardens is a housing development) is a real challenge. In addition, the use of "true 3D" models (with associated texture) will improve the performance of semantic classification.
- 3D change detection: Change detection in 3D point cloud time series. This line of research is already the subject of a thesis with the IRISA laboratory. However, we already anticipate that it will be difficult to characterize this change (because we notably lack associated ground truths) and to carry out a scaling up of these methods. The objective of the study will be to be in strong synergy with the doctoral student and to increase the maturity of the methodologies proposed (scaling up and qualification of the results).
- Fusion and registration of point clouds at different scales (terrestrial / airborne LIDAR / satellite data).
- Qualification of AI models: Today, to measure the confidence associated with a predictive model (machine learning and deep learning), the classic method remains to divide a set of data of which we have field truths into 2 subsets: learning and test. Metrics such as F1 score, accuracy, Recall, ROC, IOU are evaluated on the test subset. However, when these models are deployed in the operational phase, they predict each new observation and necessarily give a label to it. The idea is therefore to explore approaches making it possible to identify if the new observation is the result of an implementation of a distribution law similar to that which generated the learning data. New methods, such as probabilistic deep neural network approaches, appear in particular for predictive models embedded in critical environments.
- Correction of learning biases: Finally, an axis of research that is also very important and of evaluating and correcting the bias / noise introduced by either human errors or from sensor measurements in reference data(field truths). Indeed, the presence of this noisy data often leads to a loss of generalization of IA models because these then become very sensitive to minor disturbances on the characteristics of new observations which can therefore induce a poor classification.
The post-doc will be conducted at CNES in Toulouse, France as part of the AI4GEO project. It will be supervised by the host research laboratory.The recruited person will be integrated into the AI4GEO team and will be able to interact with various departments within CNES, ONERA and IGN. He/She will also be in contact with the ANITI research teams.