Every grower has done it. You hold the flower up to a loupe, tilt it toward the light, and try to decide how many trichomes have turned amber. Is it thirty percent? Fifty? The loupe doesn't have a percentage counter. Your eye makes its best guess, your experience fills in the gaps, and you make a call. The same thing happens with stigmas — you eyeball the ratio of orange to green and trust your read.

This is not a criticism of the method. It has worked for decades, for good reason. Trichomes and stigmas really do change colour as the plant matures. The problem is reproducibility. Two experienced growers looking at the same plant can reach different conclusions. The same grower can reach different conclusions on different days. What the field has been missing is an objective, automated system that can extract a precise measurement from an image — and then tell you what that measurement actually means for cannabinoid content.

A study published in Agriculture in February 2026 by researchers at Ben-Gurion University of the Negev, working in collaboration with RCK Science-Based Cannabis Genetics, built exactly that. Their system uses a consumer smartphone fitted with a clip-on macro lens, a multi-stage AI pipeline running Faster R-CNN for trichome detection and YOLOv8 for both trichome classification and stigma segmentation, and a correlation framework that links visual measurements to HPLC-verified cannabinoid data. The study ran two separate experiments across two growing seasons, and the key signals held in both.

Why Two Separate Tools Were Needed

The researchers built a dual-path pipeline because trichomes and stigmas are different problems. Trichomes are small — the tiny glandular structures that produce cannabinoids — and they cluster densely on the flower surface. Stigmas are the larger thread-like structures that receive pollen during fertilisation and visibly shift from green to orange as the plant moves through its flowering period. Each indicator requires its own detection and analysis strategy.

Both paths aggregate their outputs at the flower level — so the system doesn't just tell you about one trichome or one stigma, but about the overall profile of the whole flower. Those flower-level statistics were then correlated with HPLC-measured cannabinoid concentrations taken on the same measurement days, giving the team a ground truth to validate against.

The Imaging Setup — What the Smartphone Actually Did

One of the more remarkable aspects of this research is how deliberately low-tech the imaging hardware is. The team used an iPhone 14 Pro and an iPhone 12, each fitted with a Moment Macro Lens M-series clip-on attachment. The iPhone 14 Pro achieved an effective magnification of around thirty times; the iPhone 12 reached twenty times. Both were used with the CameraPixels app, which handled auto-exposure, ISO, and focus settings.

Images were captured both inside a working greenhouse — with natural light variation — and in a controlled laboratory environment after harvest, with an adjustable LED lamp. The CameraPixels app's auto-exposure functionality handled the lighting differences, giving the model a diverse and realistic training set rather than one calibrated to ideal lab conditions.

Each high-resolution image was divided into non-overlapping 512 × 512 pixel patches. A pre-processing sharpness filter — based on Canny edge detection — then discarded blurry or out-of-focus patches before passing the remaining regions to the AI models. This filtering step proved critical: one of the key findings from the failure analysis was that image quality, not model capacity, was the primary source of missed detections.

What the AI Actually Detected — and How Well

Trichome detection is the harder of the two problems. The structures are small, they cluster together, and the distinction between a clear and a milky trichome is genuinely subtle. The team tested a wide range of state-of-the-art object detection architectures and found that their two-stage pipeline — Faster R-CNN for detection, then a separate YOLOv8 classifier for the maturity class — outperformed all single-stage alternatives.

The gap between the two-stage pipeline and the single-stage models is worth understanding. The detection stage specialises in finding trichomes — just localising them. The classification stage then focuses entirely on reading the colour of each detected trichome. By separating these two tasks, the system avoids the compounding of errors that happens when a single model tries to do both at once. This is particularly important for the subtle distinction between clear and milky trichomes, which look almost identical to the human eye but carry opposite implications for harvest timing.

Experiment One — What the Correlations Showed

The first experiment ran across six cannabis cultivars during a spring growing season, imaging flowers at seven time points from day 51 to day 79 after flowering initiation. HPLC measurements were taken on the same days, providing a cannabinoid profile to correlate against the AI-derived visual metrics.

The trichome results were mixed. Milky trichomes showed moderate positive correlations with total cannabinoid levels in some cultivars — supporting the biological expectation that milky trichomes reflect the active biosynthesis phase. Clear trichomes generally showed negative correlations, consistent with immature tissue. But amber trichomes were inconsistent across genotypes: in some cultivars, rising amber ratios accompanied rising cannabinoid levels; in others, they did not.

Stigma colour told a more consistent story. Across all six cultivars, green stigma ratios declined while orange ratios rose over time. The point where those two curves crossed — the visual transition from green-dominant to orange-dominant — closely coincided with the peak in total cannabinoid concentration in most cases. This was especially clear in four of the six cultivars (805-12, 1416-3, 611, and 616), where the crossover nearly aligned with the cannabinoid maximum.

When the team formalised this into a harvest day prediction model — using the first day where the orange stigma ratio exceeded 40% as the threshold — they found it outperformed both milky and amber trichome predictors. Excluding four outlier plants affected by imaging quality issues, the orange stigma predictor achieved an RMSE of 1.83 days for total cannabinoids, compared to 2.57 days for simply predicting the average flowering day.

Experiment Two — Confirming the Signal

The second experiment was conducted during winter 2024–25 using three of the cultivars from Experiment 1, with an improved imaging and annotation protocol designed to address the consistency issues identified in the first round. The results were notably cleaner.

The improved sampling protocol in Experiment 2 mattered. Many of the outlier predictions in Experiment 1 were traced back to image quality problems — blurry frames, shallow depth of field, or imaging angle variability that caused the same plant on the same day to yield very different trichome counts depending on how the camera was positioned. When these acquisition variables were better controlled, the biological signal became more consistent.

What the Results Actually Mean for Harvest Decisions

The study is clear that this is a proof-of-concept. Six cultivars in one greenhouse, two smartphone models, and a dataset of thirty plants does not produce a model ready for commercial deployment. The authors say so explicitly. But the directional findings are robust enough to draw practical conclusions.

What Is Still Missing — and What Comes Next

The study's limitations are worth naming clearly. The stigma segmentation model was trained on only 115 images — a small dataset that constrains how well it generalises to different cultivars, lighting conditions, and camera hardware. The correlation experiments covered six cultivars in Experiment 1 and three in Experiment 2, which is not enough to build genotype-independent maturity prediction models.

The trichome-to-cannabinoid correlations were also inconsistent across genotypes in Experiment 1, which the authors attribute to both biological variability and sampling limitations. Some cultivars showed strong patterns; others showed weaker or inverted ones. This genotype-dependency is a known challenge in cannabis phenotyping and is not specific to this study — it was also a finding in last week's Tran et al. paper on stigma colour staging across 25 genotypes.

The authors identify the path forward: larger datasets across more cultivars and environments, better image acquisition protocols with automated quality filtering built into the capture step, and direct integration testing in live greenhouse workflows. They also note the possibility of combining the visual pipeline with additional inputs — days after flowering, temperature data, cultivar-specific calibration — to build a multi-feature model once sufficient training data exists.

The long-term vision is a field-ready tool that any cultivator can use on a smartphone: point it at a flower, capture a sharp macro image, and receive an AI-derived maturity score alongside a harvest recommendation window. That tool doesn't exist yet at commercial scale. But this paper shows it is not a distant possibility. The biology is consistent. The AI can read it. The question now is how many more flowers, cultivars, and seasons of data it takes to make the system robust enough to trust.