YOUME Ousmane

Affiliated project Deep4EnvMonitoring

Title Evolutionary hyperlattice for convolutional neural networks (CNN): application to environmental monitoring

Affiliated CEAs:

CEA SMIA, University of Abomey-Calavi

CEA MITIC Gaston Berger University

Supervision Jean Marie Dembélé, ACE MITIC, Gaston Berger University, Senegal, jmdembele@gmail.com

Co-supervision : Prof. Eugène C. EZIN, ACE SMIA, Institut de Mathématiques et de Sciences Physiques, Université d'Abomey-Calavi, Benin, eugene.ezin@gmail.com

Other contributors to thesis supervision : Prof. Christophe CAMBIER, Sorbonne-Université - IRD

Start date: 01/01/2021
Anticipated date of thesis defense: June-July 2024
ORCID profile: 0009- 0001-8427-8629

• LinkedIn

Project title: Evolutionary hyperlattice for convolutional neural networks (CNN): application to environmental monitoring

Summary of scientific project:

Environmental monitoring is crucial in today's world, faced with growing ecological challenges due to rapid urban expansion, waste management, pollution and the preservation of natural habitats. Advanced technologies, particularly deep learning techniques such as convolutional neural networks (CNNs), can help automate environmental monitoring, enabling better decision-making and ecosystem conservation. Despite their potential, the direct application of convolutional neural networks (CNNs) to environmental monitoring faces considerable difficulties due to the complexity of natural environments, the diversity of data types and the need for accurate, real-time analysis. The main obstacles are obtaining diverse, high-quality datasets, ensuring model generalizability across different ecosystems, managing substantial computational requirements, and addressing privacy, ethical issues and variability in environmental conditions.
This thesis addresses these challenges by integrating CNN techniques with evolutionary algorithms.

Summary of results:
In this thesis, we explored several CNN architectures for remote sensing environmental monitoring. We first developed a model with a Single Shot Detector to identify illegal dumps in Saint Louis, overcoming the difficulties encountered with a basic CNN structure [1]. Our study then focused on automated CNN architecture search, choosing genetic algorithms because of their comprehensive architecture and hyperparameter generation capabilities. This led to the creation of the "Search Length Strategy" and "Search Architecture Networks" algorithms, which significantly reduced computational requirements from 37 to 4 GPU days compared to competitors with the same performance [2].
We also investigated the effectiveness of different segmentation models for coastal plastic waste. We found that Panoptic DeepLab outperformed Mask R-CNN in context recognition and false-negative reduction, although Mask R-CNN outperformed in false-positive minimization [3].

Our data collection focused on various environmental contexts such as coastal waste, large-scale waste landscapes and waste along riverbanks, using advanced methods for improved accuracy. We applied the formalism of our evolutionary algorithm combined with Mask R-CNN to the detection of buildings via satellite imagery, which facilitates the monitoring of urban development in regions lacking open-source data, and can facilitate the generation of trajectories for autonomous drones.