A Hybrid Multispectral Deep Feature Fusion Framework for Sustainable Solid Waste Classification Using RGB and NIR Imaging

The rapid increase in solid waste resulting from rising consumption has increased the need for effective recycling technologies to reduce environmental impacts and recover valuable resources. This study proposes a hybrid approach based on RGB and near-infrared (NIR) imaging for the automated classification of solid waste. A unique dataset consisting of five waste classes—glass, paper, fabric, metal, and plastic—was created using a conveyor belt system equipped with synchronized RGB and NIR cameras. Waste objects were detected in RGB images using the Segment Anything Model (SAM) and cropped according to the generated YOLO labels. Three spectral channel combinations, NIR-G-B, R-NIR-B, and R-G-NIR, were generated from paired RGB and NIR images. Deep features were extracted from RGB, NIR, and these spectral combinations using DenseNet121. Feature dimensionality was then reduced using PCA, and the resulting features were classified using SVM. The proposed method achieved 98.01% accuracy and recall on the 80/20 hold-out test split, outperforming models based solely on RGB or NIR images, while 5-fold and 10-fold cross-validation yielded mean accuracies of 96.16% and 96.34%, respectively. The results demonstrate that combining the complementary features of visible and NIR spectral information is effective in distinguishing different types of waste. The proposed approach can contribute to the development of smart and sustainable recycling systems by supporting more accurate and automated sorting of recyclable materials.