📄 agriculture

Application of Unsupervised Clustering Techniques to Somatic Embryos for Automated SE Fluidics System

This study demonstrates that a machine learning pipeline combining unsupervised clustering techniques, specifically Constrained Unsupervised Learning (CUL) with Principal Component Analysis, can achieve near-perfect accuracy and significantly faster processing speeds than traditional morphological methods for the automated sorting and categorization of somatic embryos in industrial fluidics systems.

Original authors: Chaitanya S. Nayak, Punnag Chatterjee, Cyrus Aidun

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chaitanya S. Nayak, Punnag Chatterjee, Cyrus Aidun

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a massive, high-speed factory that grows tiny, microscopic trees (specifically pine and spruce embryos) in a liquid solution. This process is called Somatic Embryogenesis. The goal is to grow millions of these embryos to replant forests, which helps soak up carbon dioxide from the air.

However, there's a problem: these embryos don't all grow at the same speed. Some are perfect, ready-to-go "babies" with all their little leaves (cotyledons) intact. Others are misshapen, incomplete, or just blobs of cells that won't survive.

In the past, humans had to stand by a conveyor belt, look at every single embryo through a camera, and manually decide: "Keep this one" or "Throw that one away." This is slow, tiring, and prone to human error.

This paper is about teaching a computer to do this sorting job automatically, but with a clever twist. Instead of just telling the computer exactly what a "good" embryo looks like, the researchers let the computer figure out the patterns itself first, and then taught it how to sort.

Here is how they did it, broken down into simple steps:

1. The Setup: Taking the "ID Photos"

First, the embryos flow through a glass tube. A camera takes a black-and-white photo of each one as it passes.

The Cleaning Crew: Before the computer can look at the photos, the researchers had to clean them up. The photos had random text, smudges, and spots on them. They used digital tools to crop out the embryo, remove the text, and smooth out the edges, leaving just a clean picture of the embryo against a black background.

2. The Challenge: Orientation vs. Reality

Imagine taking a photo of a person. If they turn their head left, right, or upside down, the photo looks very different, even though it's the same person.

The Problem: The embryos were floating in liquid, so they were captured at random angles. If the computer just looked at the pixels, it might think two embryos are different just because one is upside down and the other is right-side up.
The Solution: The researchers used a trick called Image Augmentation. They took every photo and created "twins" of it by rotating it and flipping it. Then, they averaged all these versions together. This forced the computer to ignore the angle and focus only on the shape (like how many leaves it has or how long its body is).

3. The Sorting Methods: Three Different Ways to Group

The researchers tried three different "brainy" ways to sort these embryos without human help.

Method A: The "Guess the Number" Game (K-Means)

This is like asking a computer to sort a pile of mixed socks into 3 piles.

The Result: When they asked for 3 piles, the computer got confused. It mixed up the "good" and "bad" embryos.
The Fix: When they changed the request to just 2 piles (Good vs. Bad), the computer did much better. It successfully separated the healthy embryos from the rejects with about 92% accuracy. It realized that the most important difference wasn't the specific type of "good" embryo, but simply "good" vs. "bad."

Method B: The "Nudge" System (Constrained Unsupervised Learning)

This is the most clever part. Imagine you are sorting a pile of rocks, but you don't know the rules.

The Nudge: The researchers gave the computer a few tiny hints. They said, "Hey, these two rocks look exactly the same, so they must go in the same pile" (Must-link). Then they said, "These two look totally different, so they must go in separate piles" (Cannot-link).
The Result: With just these small hints, the computer became a genius. It sorted the embryos into three perfect groups:
1. Well-defined embryos: Perfect shape, clear leaves. (100% accurate)
2. Elongated blobs: Long bodies but no leaves. (100% accurate)
3. Circular blobs: Just round lumps with no shape. (87.5% accurate)
Why it matters: The computer didn't just follow a rigid rule; it used the hints to learn the concept of a good embryo on its own.

Method C: The "Tree of Life" (Hierarchical Clustering)

This method is like building a family tree. The computer starts by treating every single embryo as its own unique family. Then, it starts merging the most similar ones together, step-by-step, until everything is one big group.

The Discovery: They looked at a "family tree" diagram (called a dendrogram). They found that no matter if they started with 500 photos or 4,000 photos, the tree structure looked the same once they reached a certain point.
The Takeaway: You don't need to look at millions of photos to teach the system. Looking at just 500 representative photos was enough to understand the whole population.

4. The Final Step: The Speedy Sorter (Online Classification)

Once the computer learned how to group the embryos using the methods above, the researchers trained a final "Speedy Sorter" (a Supervised Learning model).

How it works: The computer took the groups it discovered and learned to recognize them instantly.
The Speed: This new system was 32% faster than the old way of measuring physical features manually. It could process an image in about 0.004 seconds. That's faster than a human can blink.

Summary

The paper shows that by letting a computer discover patterns first (Unsupervised Learning) and then teaching it to sort quickly (Supervised Learning), we can automate the process of growing trees.

Old Way: A human looks at every embryo and decides.
New Way: The computer learns what a "good" embryo looks like by studying a small sample, then sorts thousands of them in a split second with near-perfect accuracy.

This means we can scale up forest farming to help the planet without needing armies of people to sort through tiny, floating tree babies.

Technical Summary: Application of Unsupervised Clustering Techniques to Somatic Embryos for Automated SE Fluidics System

Problem Statement
Somatic embryogenesis (SE) is a critical method for the clonal propagation of economically significant crops, such as Norway spruce (Picea abies). However, scaling SE production for industrial forestry requires automation to reduce manual labor, particularly in the selection and harvesting of mature embryos. A primary challenge in this process is the asynchronous nature of embryo maturation, which complicates the identification of viable embryos. While Machine Learning (ML) has been applied to agricultural image processing, its use in SE remains underexplored. Previous supervised learning (SL) approaches, such as Convolutional Neural Networks (CNN) and Support Vector Machines (SVM), rely on human-defined classes (e.g., "Accept Left," "Accept Right," "Reject"). This manual labeling process often depends on superficial visual features, potentially overlooking significant morphological characteristics, and requires extensive human intervention to define the optimal number of classes. This study addresses the need for an automated, unbiased method to segregate embryo image datasets using unsupervised learning (UL) and semi-supervised techniques.

Methodology
The study utilized a dataset of 720 grayscale images of somatic embryos captured via an SE Fluidics system. The workflow involved several stages:

Image Preprocessing: Raw images were cropped to the embryo's bounding box, resized to fixed dimensions using bilinear interpolation, and cleaned of artifacts (text, smudges) using Gaussian inpainting. Adaptive thresholding, dilation, and erosion were applied to segment the embryo from the background and refine its shape.
Feature Extraction and Augmentation: To ensure clustering was based on intrinsic morphological traits (e.g., body length, cotyledon count, spacing) rather than orientation, image augmentation was applied (rotations and flips). Histogram of Oriented Gradients (HOG) vectors were extracted from these variations and averaged to create a single representative feature vector per embryo.
Clustering Approaches:
- Unsupervised Learning (K-means): Applied with Principal Component Analysis (PCA) for dimensionality reduction. The study tested $k=3$ (mimicking human-defined categories) and $k=2$ (Accept vs. Reject).
- Constrained Unsupervised Learning (CUL): A semi-supervised approach where pairwise constraints (must-link and cannot-link) were generated based on local density and similarity measures. These constraints guided the algorithm to group embryos by morphological traits (e.g., "Well-Defined," "Elongated Blobs," "Circular Blobs") without full manual labeling.
- Hierarchical Clustering: An unconstrained method used to generate dendrograms. This allowed for dynamic cluster determination by analyzing threshold values (e.g., 0.43, 0.53, 0.8) to observe cluster merging and segregation behaviors across datasets of varying sizes (500, 3,500, and 4,089 images).
Hybrid Classification Pipeline: A final pipeline integrated CUL for initial clustering with a supervised SVM model. The CUL-generated clusters served as labeled training data for the SVM, which was then tested on a hold-out set to classify new images in real-time.

Key Results

K-means Performance: With $k=3$ , accuracy was low (47.6% to 83.9%). However, reducing $k$ to 2 (Accept vs. Reject) significantly improved performance, achieving 82% accuracy for rejected embryos and 92% for acceptable embryos.
Constrained UL (CUL) Performance: By introducing pairwise constraints, the CUL method achieved nearly 100% accuracy for "acceptable" (well-defined) embryos and 87.5% for "rejected" (circular/elongated blobs) embryos. The algorithm successfully identified three distinct morphological groups: well-defined embryos (3+ cotyledons), elongated blobs (positive body length, <3 cotyledons), and circular blobs (negligible body length/cotyledons).
Hierarchical Clustering & Dataset Size: Hierarchical clustering with a threshold of 0.8 consistently produced two clusters (Accept/Reject) across all dataset sizes. Analysis of dendrograms revealed that a dataset of 500 images was sufficient to represent the exhaustive morphological traits of the entire population (up to 4,089 images), as cluster convergence occurred after processing this volume.
Hybrid SL Performance: The SVM model trained on CUL-extracted clusters achieved 100% segregation of rejected embryos in the test set. The processing time per image averaged 0.00408 seconds, making the pipeline 32.3% faster than traditional morphological classification methods.

Significance and Claims
The paper claims that the proposed pipeline offers a robust framework for efficient in-situ embryo categorization, facilitating scalable automation in industrial SE applications. By leveraging unsupervised and constrained learning, the system minimizes human bias associated with manual class definition and reduces the reliance on large, manually labeled datasets. The study demonstrates that a small subset of images (hundreds rather than thousands) can establish a base cluster for identifying viable embryos. Furthermore, the integration of CUL with SL provides a pathway for real-time decision-making, overcoming the limitations of traditional UL algorithms in online environments. The authors suggest that this approach can be extended to other datasets requiring automated, high-precision clustering and propose future work involving Hyperspectral Imaging (HSI) to further refine feature extraction.