Things About Me Series - 05 - Can AI Replace Petrographers?
Artificial intelligence is making waves in every field — but can it really replace the human eye behind the microscope?
The short answer: probably — but not just yet, and not in everything.
I recently reviewed a paper where the authors applied a neural network to identify macerals, claiming that petrography is time-consuming and subject to operator bias. Fair enough. But the issue was that the model completely misidentified the macerals — classifying inertinite (semifusinite and macrinite) as vitrinite, while nearby collotelinite (a type of vitrinite) wasn’t detected at all.
For context, macerals are the organic constituents of coal, grouped into vitrinite, inertinite, and liptinite, and are typically identified under a microscope based on their reflectance (brightness under oil immersion). It's a job that relies on both visual cues and expertise. (see previous blog post on macerals).
I’m not against automation — not at all. If a reliable machine learning method can handle routine analyses, I’m all for it. That just frees me up to focus on the more interesting parts of the work.
I was involved in a study that used Mask R-CNN for maceral identification — a deep learning model that performs instance segmentation detecting and masking each object individually in an image. In theory, this is ideal for identifying individual macerals.
So, we set out to annotate and classify every single particle across hundreds of images, using coal samples ranging from high- to low-volatile bituminous coals from various countries. It took over a year just to annotate the images and build the training set. It was painstaking, detailed work. As expected, the dataset was extremely imbalanced - some macerals were much more abundant than others. The model struggled to segment complex coal particles composed of multiple macerals. And even when we classified macerals by groups (vitrinite, inertinite, liptinite), it still had trouble achieving good accuracy. What we didn’t test, though, was model performance by coal rank.
More recently, a paper by Wang et al. (2022) used semantic segmentation models like U-Net, SegNet, and DeepLab v3+ for maceral group identification. Unlike Mask R-CNN, these models classify each pixel in an image rather than detecting individual objects.
Wang et al dataset included Chinese coals from high- to low-volatile bituminous ranks. Their results — particularly with DeepLab v3+ — were impressive, probably the best I’ve seen so far, with over 90% accuracy (except for inertinite). However, 21% of the inertinite was misclassified as vitrinite, and the authors didn’t offer an explanation. I initially thought low-reflectance inertinite might be the cause, but their figures showed those were still correctly classified.
Now, I’m not a deep learning specialist, so I don’t know exactly how these CNNs make their decisions. But as a coal petrographer, reflectance (i.e. grayscale intensity) is the first thing I use to distinguish maceral groups. That feature is independent of morphology. But CNNs seem to rely most heavily on shape and structure.
I understand why — in many image classification tasks, shape is more informative than colour. For example, identifying vehicles by shape is usually more reliable than using colour. But in petrography, colour (reflectance) is key. If there’s a way to prioritize reflectance as a primary feature, we’re halfway there. Of course, this also requires standardized illumination conditions during image capture.
It may also be more effective to train separate models for different coal ranks or specific mine sites. That way, each model could learn the particular reflectance patterns and textures relevant to that context. But this also means building separate training sets — a major increase in the number of annotated images required.
Still, I think we’re getting close. Most studies now report accuracy above 80%, sometimes much higher, for maceral group classification. The key remains the same: good training data. That means:
Proper boundary labelling
Balanced datasets for each class
Inclusion of edge cases and variable conditions
And above all, correct maceral identification
We’re not at the finish line yet — but we’re on the track. And as both petrographers and machine learning evolve, the finish line might not be about replacing us at all — but working together.
References:
Magalhães Santos, R.B., Augusto, K.S., Iglesias, J.C.A., Rodrigues, S., Paciornik, S., Esterle, J.S., Domingues, A.L.A., 2022. A deep learning system for collotelinite segmentation and coal reflectance determination. International Journal of Coal Geology 263, 104111.
Wang, Y., Bai, X., Wu, L., Zhnag, Y., Qu, S., 2022. Identification of maceral groups in Chinese bituminous coals based on semantic segmentation models. Fuel 308, 121844.
