Résumé:
The detection and classification of plastic waste have become critical priorities amid escalating environmental
concerns, particularly in aquatic and terrestrial ecosystems. Conventional imaging and manual inspection
techniques often lack the accuracy, efficiency, and scalability required to address these challenges in dynamic,
real-world conditions. Hyperspectral imaging (HSI), with its ability to capture detailed spectral signatures
across hundreds of narrow wavelength bands, emerges as a promising solution for identifying and distinguishing various plastic polymers from natural elements. However, the practical application of HSI is hindered
by computational complexity, spectral redundancy, environmental variability, and the limited availability of
labeled data.
This thesis presents a comprehensive, adaptive framework that synergizes deep learning and classical machine
learning techniques to overcome both environmental and computational barriers. We propose and evaluate
four complementary methodologies : (1) a Principal Component Analysis (PCA) followed by Support Vector
Machine (SVM) classifier for efficient low-dimensional spectral classification ; (2) a Random Forest model
utilizing full-band statistics to enhance robustness and interpretability ; (3) a 2D U-Net convolutional neural
network capable of direct spectral-spatial segmentation without prior dimensionality reduction ; and (4) a hybrid two-phase pipeline that couples U-Net-based RGB segmentation with hyperspectral SVM classification
to achieve fine-grained polymer-level discrimination.
These methods are benchmarked on a range of diverse, real-world hyperspectral datasets, including aerial drone imagery and controlled flume-based laboratory acquisitions, which feature varying environmental
backgrounds (e.g., water, sand, algae), lighting conditions, and polymer types. Results demonstrate that the
hybrid pipeline outperforms standalone models by leveraging the spatial precision of convolutional architectures and the spectral sensitivity of kernel-based classifiers. Specifically, the segmentation phase achieved an
F1-score of 99.42%, accuracy of 99.35%, Intersection over Union (IoU) of 98.75%, and mean Average Precision (mAP) of 87.91%, with a validation loss of only 0.016. In the classification phase, the SVM attained an
overall precision of 91.5% across six polymer types—PET, HDPE, LDPE, PP, EPSF, and weathered plastics,
with perfect (100%) precision for EPSF and LDPE.
The full framework processes 10 files in just 12.7 seconds on GPU, highlighting its potential for real-time
deployment in embedded systems. This work confirms the feasibility of automated plastic detection in both
field and lab environments and lays the groundwork for scalable applications in UAV and satellite-based
monitoring, smart recycling infrastructures, and embedded environmental sensing platforms.
