We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Multi-Angle Scattering and Image-Derived Feature
Summary
Scientists are testing whether a laser-based technique combined with AI can quickly identify what type of microplastic is in a water sample (like nylon vs. PVC) and measure how much is there. This matters because microplastics are showing up in our water, food, and even bodies, but figuring out exactly what kind and how much currently requires slow, expensive lab work—faster tools like this could make widespread monitoring of these potential health hazards more realistic. Note that this is a technical dataset and methods paper demonstrating the tool works in controlled samples, not a health study on microplastics' effects on the body.
This repository contains the complete dataset, raw data, and reproducibility code supporting research into the automated classification and concentration quantification of microplastics using optical scattering and image-derived features. The work establishes a dual-pipeline framework designed to characterize Nylon and PVC microplastics efficiently: Concentration Regression Pipeline: Utilizes multi-angle scattering intensity profiles ($S_1, S_2, S_3, S_4$) across a concentration range of $0.100$ to $9.983$ to predict sample concentrations via non-linear regressors (e.g., Support Vector Regression with RBF kernel and Random Forests). Polymer Classification Pipeline: Leverages image-derived morphological features to classify the polymer type, evaluated against computer vision benchmarks. File Descriptions RAW_DATASET.zip: NYLON and PVC imaging final_dataset.csv: The clean, curated dataset containing aligned multi-angle scattering values ($S_1$–$S_4$), morphological image-derived parameters, concentration target values, and ground-truth polymer labels (NYLON / PVC). (Note: The companion executable Jupyter Notebook final_code.ipynb handles the end-to-end execution of these files, from preprocessing to model tuning). Methodology & Machine Learning Framework The framework addresses a rigorous validation protocol to ensure robustness and reproducibility: Hyperparameter Tuning & Validation: Implements systematic K-Fold Cross-Validation alongside dedicated hold-out testing to monitor variance and prevent data leakage. Metrics Evaluated: Regression models are assessed via RMSE, MAE, $R^2$, and calculated Limits of Detection/Quantification (LOD/LOQ). Classification pipelines are benchmarked using Confusion Matrices, Precision-Recall Curves, ROC-AUC, and mean Average Precision (mAP). Explainable AI (XAI): Features feature importance mappings via tree-based MDI (impurity), permutation importance, and SHAP (SHapley Additive exPlanations) values for model transparency. Computer Vision Integration: Includes evaluation structures for real-world deployment compatibility testing using YOLO object detection models.