This Startup’s Deep Learning Algorithms Enable Testing Cannabis Quality From a Phone Camera

The market for cannabis is gaining global acceptance, for medical and recreational purposes. The total market size is currently estimated at $340 billion, with over 200 million people globally consuming cannabis weekly. Progress has been made with regulations and progressive thinking for decriminalization and legalization, catapulting cannabis as a high-growth market, but some aspects haven’t been tinkered with enough, like the ability to test cannabis quality at home. HiGrade, an Israeli startup, developed deep learning based technology to cover the full spectrum of cannabis quality testing, all from a smartphone camera.

HiGrade, founded in early 2017, developed image processing technology using deep neural networks to test the quality of cannabis bud and crops by measuring potency and grade, and providing cultivation aid. HiGrade allows all value chain members, from growers to consumers, to test cannabis quality in a fast, cost effective and non destructive way. Their mobile app allows users to capture three images of cannabis buds or plants to measure total THC percentage (Tetrahydrocannabinol) levels, get overall quality, and provide cultivation aid to growers - an agronomist in your pocket.

The startup was founded by Asaf Levy and Assaf Gavish, experts in the agronomy and bio-agricultural domains. Gavish, CTO, has a masters degree in plant biology from the Weizmann Institute of Science with specialization microscopy and image analysis. Levy, CEO, holds an MBA from Tel Aviv University and grew up working in the Golan Heights, farming fruit crops. The agronomist duo decided to channel their expertise to the cannabis market because of the infancy of its analytical community. “The knowledge gap in the cannabis industry is of the widest” explained Levy. “Unlike wheat and tomatoes with decades of cultivation knowledge, for more than 80 years, there's no research on how to cultivate cannabis. It's all been passed on informally.” Moreover, the economics of cannabis is lucrative. “Cannabis is over 2,000 fold more profitable per square meter per year than general grains.”

Typically, cannabis quality testing is done in a laboratory setting with High Performance Liquid Chromatography (HPLC) equipment. The test is accurate, but the process is lengthy, destroys the product, and most importantly is a sample for a batch, with variability from the potency of the flower sample from the rest of the plant, or crop. That margin of error results in serious disinformation to growers and consumers, especially since THC percentage is a major variable by which the product is priced.

HiGrade’s mobile app is capable of analyzing the plant’s grade with a standard camera. With their kit, containing a small mountable mobile lens with 30X magnification, users can test for potency, or get cultivation aid features.

For cannabis flower potency analysis (off-plant), their patent-pending technology is built off a training set of cannabis flowers, with corresponding images tested individually in HPLC labs, providing the ground truth of potency. With their dataset, they trained deep neural networks in an ensemble of 30 models to predict the chemical analysis for the subject image. With more users and cannabis flower image submissions, their network graph learns what features of the flower matter, and to what extent. “Our algorithm found over 2000 such features, and reached a predictive ability of 85% accuracy” explained Gavish.

For cultivation aid, they employ classification and regression algorithms, and for plant health, they utilize object detection with classification for insect and disease detection.

Their bud potency analysis is fully autonomous, while cultivation aid, including harvest timing optimization, and pest and mold detection, is semi-autonomous, wherein they apply agronomical knowledge and expertise to their cultivation advice. “With this relationship, we can assess which problems to apply our research by understanding which problems our users struggle with” explained Gavish.

HiGrade is targeting consumers and small-scale cannabis growers. Their app has grown organically from 7,000 at the beginning of 2019, to over 55,000 users (almost 8x!), and they’re processing over 100,000 images per month. These images are constantly annotated by HiGrade’s agronomic team to be used as datasets for the next generation of algorithms, which the company intends to deploy in industrial cannabis cultivation and processing.

The startup plans to develop applications dedicated to greenhouse crop growers for continuous monitoring of camera feeds with agronomist AI based oversight and are currently in pilots with industrial farms worldwide.