Etude du passage à l'échelle des algorithmes de segmentation et de classification en télédétection pour le traitement de volumes massifs de données

摘要

Recent Earth observation spatial missions will provide very high spectral, spatial and temporal resolution optical images, which represents a huge amount of data. The objective of this research is to propose innovative algorithms to process efficiently such massive datasets on resource-constrained devices. Developing new efficient algorithms which ensure identical results to those obtained without the memory limitation represents a challenging task. The first part of this thesis focuses on the adaptation of segmentation algorithms when the input satellite image can not be stored in the main memory. A naive solution consists of dividing the input image into tiles and segment each tile independently. The final result is built by grouping the segmented tiles together. Applying this strategy turns out to be suboptimal since it modifies the resulting segments compared to those obtained from the segmentation without tiling. A deep study of region-merging segmentation algorithms allows us to develop a tile-based scalable solution to segment images of arbitrary size while ensuring identical results to those obtained without tiling. The feasibility of the solution is shown by segmenting different very high resolution Pléiades images requiring gigabytes to be stored in the memory. The second part of the thesis focuses on supervised learning methods when the training dataset can not be stored in the memory. In the frame of the thesis, we decide to study the Random Forest algorithm which consists of building an ensemble of decision trees. Several solutions have been proposed to adapt this algorithm for processing massive training datasets, but they remain either approximative because of the limitation of memory imposes a reduced visibility of the algorithm on a small portion of the training datasets or inefficient because they need a lot of read and write access on the hard disk. To solve those issues, we propose an exact solution ensuring the visibility of the algorithm on the whole training dataset while minimizing read and write access on the hard disk. The running time is analysed by varying the dimension of the training dataset and shows that our proposed solution is very competitive with other existing solutions and can be used to process hundreds of gigabytes of data.