We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Soil Microplastic Multiscale Image Dataset (Microscopic and Macroscopic)
AI summary Read the abstract
Researchers created an open image dataset of microplastics in farm soil samples, combining smartphone photos and microscope images, to train computer programs to detect, classify and count plastic particles. The dataset includes thousands of labeled images of fibers, films and fragments, and was used to test a deep learning system that identified particles with up to 96.83% accuracy.
Soil Microplastic Multiscale Image Dataset (Microscopic and Macroscopic) OverviewThe Soil Microplastic Multiscale Image Dataset is an open-access image collection created to train and test computer vision models for detecting, classifying, and counting microplastics in agricultural soil. The dataset combines two complementary imaging scales: Smartphone Macro Photos: Overlapping photos taken around an entire circular filter paper. These can be stitched into a full panorama to count all particles on the sample and estimate soil contamination in particles per kilogram (particles/kg). Optical Stereo Microscope Images: High-resolution images that capture sharp surface textures, colors, and edges, allowing deep learning models to identify and measure individual particles with high accuracy. The dataset covers the three most common microplastic shapes found in agricultural fields: Fibers: Thin, elongated, thread-like synthetic plastic strands. Films: Thin, flexible, sheet-like plastic pieces. Fragments: Rigid, angular, irregular plastic pieces with rough or broken edges. Why This Dataset Was CreatedMicroplastics in farm soils can damage soil structure, harm earthworms and microbes, reduce crop yields, and enter our food chain. Testing soil for microplastics today is a major challenge: Manual microscope sorting takes hours of tiring visual search under a lens and easily leads to human counting mistakes. Chemical laser methods (such as FTIR and Raman spectroscopy) are accurate, but the machines are expensive, require trained specialists, and are too slow for routine monitoring. To make microplastic monitoring scalable and affordable, the EcoLens project uses ordinary smartphone cameras for wide-area counting and standard optical stereo microscopes for detailed particle inspection. Laboratory Sample PreparationAll samples were prepared under controlled laboratory conditions using genuine soil samples: Real Soil Matrix: 38 independent agricultural soil samples were collected from farmland fields. Using real soil is important because it introduces natural soil grains, minerals, and organic debris that challenge computer vision models. Spiking: Soils were dried, sieved, and spiked with 0.5% to 1.0% (w/v) microplastics prepared through mechanical milling (0.5 mm to 5 mm size range). Centrifugation: The soil-plastic mixture was centrifuged repeatedly at 9,000 rpm for 30 minutes in ultra-pure water to separate lighter plastics and organic residues from dense soil sediments. Vacuum Filtration: The separated liquid was vacuum-filtered through circular 0.47 um GF/C glass microfiber filter paper, collecting the microplastics on a clean, uniform white surface. Technical Lighting Replicates: To teach AI models to handle realistic lighting changes in labs, the 38 physical filter papers were photographed under different controlled lighting setups, yielding 61 original macro filter-paper image sets. Camera & Microscope Hardware Smartphone Macro Camera: Samsung Galaxy A35 5G dedicated macro lens (2560 x 1440 pixels, f/2.4 aperture, ISO 50). Captured 6 to 13 overlapping photos per filter paper under diffuse ambient lighting. Stereo Microscope: ZEISS Stemi 305 Stereo Microscope (Carl Zeiss, Germany) equipped with a ZEISS Axiocam 105 color digital camera and Carl Zeiss ZEN Lite software (10x total optical magnification, calibrated at 2.2 um x 2.2 um per pixel). Original Carl Zeiss XML metadata files are included for all microscope captures. Dataset StructureThe dataset contains 61 sample folders and 1,278 total files (~3.98 GB uncompressed): macroscopic_images/: 61 folders (551 photos) of overlapping smartphone photos covering primary samples (S1–S38), validation samples (VS01–VS17), and clean blank filter negative controls (VSN01–VSN06). microscopic_images/: 55 folders (362 images + 365 XML files) of calibrated stereo microscope captures and Carl Zeiss acquisition metadata. samples_metadata.csv: Table indexing all 61 samples with sample IDs, categories, file counts, filter paper specs, and preparation notes. manifest.csv: Complete file list with MD5 and SHA-256 checksums to verify downloaded file integrity. datapackage.json: Machine-readable Frictionless Data Package schema. Annotations & Dataset StatisticsImages were annotated in Label Studio v1.21.0 with bounding boxes (detection), class tags (classification), and brush masks (segmentation) assisted by the Segment Anything Model (SAM) with manual quality checks: Original Labeled Images: 61 macro filter captures (50 train / 11 val) and 212 microscope captures (169 train / 43 val). Manually Labeled Crops: 718 macro crops (238 fibers, 147 films, 219 fragments) and 628 microscopic crops (250 fibers, 137 films, 241 fragments). Augmented Dataset: Expanded using rotations, flips, brightness, and contrast adjustments to 10,012 macro crops (3,912 fibers, 2,538 films, 3,562 fragments) and 3,000 microscopic crops (1,000 fibers, 1,000 films, 1,000 fragments). EcoLens Deep Learning Pipeline & Benchmark ResultsThe dataset was used to train and evaluate a three-stage deep learning pipeline: Particle Detection (YOLOv8): 82.45% mAP on microscope images; 56.27% mAP on stitched macro panoramas. Morphology Classification (EfficientNet-B3): 96.83% accuracy (96.83% F1-score) on microscope crops; 93.61% accuracy (92.39% F1-score) on macro crops. Instance Segmentation (Mask R-CNN): Morphology-specific models for fibers, films, and fragments, achieving up to 0.8550 IoU for fragments and 0.8478 IoU for films. Soil Concentration Estimation: Calculates total contamination in particles per kilogram (particles/kg) across the reconstructed filter paper: Particles per kg = (Detected Particles / Soil Weight in grams) * 1000 Open-Source Code & Web PlatformThe complete software platform, image stitching tools, and trained deep learning models are available on GitHub:GitHub Repository: GitHub Repository How to CiteIf you use this dataset or code, please cite our research paper: Sewwandi, A., Silva, H., Gunawardhana, N., Edirisingha, I. P., Undugoda, L. J. S., & Lankasena, N. (2026). EcoLens: An Integrated Deep Learning Framework for Detection, Classification, and Quantification of Microplastics in Soil Using Macro and Microscopic Imaging. Faculty of Technology, University of Sri Jayewardenepura, Sri Lanka.
More Papers Like This
Morphological Detection and Classification of Microplastics and Nanoplastics Emerged from Consumer Products by Deep Learning
AI summary Read the abstract
Researchers created a new open-source dataset of microscopy images for training AI models to automatically detect and classify micro- and nanoplastics. The dataset fills an important gap in available tools for microplastic research and provides a foundation for developing faster, more efficient methods to identify plastic contamination across environmental samples.
Application of hyperspectral and deep learning in farmland soil microplastic detection
AI summary Read the abstract
Hyperspectral imaging combined with deep learning was applied to detect and classify microplastics in farmland soil, offering a non-destructive, rapid alternative to time-consuming chemical extraction methods. The model achieved high classification accuracy across polymer types, demonstrating the potential for field-deployable microplastic monitoring in agricultural settings.
Microplastics detection in agricultural soil combining 3D Laser Scanning Confocal Microscopy with machine learning
AI summary Read the abstract
Scientists developed a faster, chemical-free method to detect and measure tiny plastic fragments (as small as 53 microns) in farm soil, using a special 3D microscope paired with computer-assisted image analysis. This matters because microplastics in agricultural soil can end up in our food and water, and better detection tools help researchers track how much plastic pollution is actually building up in the fields that grow our crops. The method worked well in most soils but had more trouble detecting dark-colored plastic bits and fibers in soil with lots of natural organic debris, so there's still room for improvement.
Processed NIR-HSI and self-organizing map dataset for microplastic classification on sandy substrates
AI summary Read the abstract
Scientists have developed a camera-based scanning technique that can spot and map tiny plastic bits (microplastics) hiding in sand, without digging through it or destroying samples. This dataset provides the underlying reference data and computer models used to identify different plastic types, which could help researchers monitor beach pollution more efficiently, an important step since microplastics can end up in the food chain and, eventually, in us.
2T2D spectroscopy and deep learning for soil MPs quantification
AI summary Read the abstract
Scientists have developed a new tool that combines a special light-scanning technique with artificial intelligence to detect and measure tiny plastic particles (microplastics) hiding in soil. This matters because microplastics in soil can end up in the crops we eat and the water we drink, so having a faster, more accurate way to track them helps researchers understand how much plastic pollution is entering our food supply. This dataset and code release makes the detection method available for other scientists to test and build on, which could speed up future research into microplastics and health risks.
Research digests by email
When a large batch of papers lands in the Atlas, we read through it and send a short write-up of what stood out.