Deep learning has been revolutionizing various aspects of our lives. Now, it might equip cannabis farmers to identify crop diseases in their genesis, thus helping to prevent sweeping losses in yield.
To that end, a paper  from 2020, the collaborative effort of Indian and U.S. scientists, tested a new algorithm that builds on existing models in that it “considers some temporal structure, i.e. it can learn the progression of the disease and the extent of damage over time.”
The models used were Fast Region Convolutional Neural Network (F-RCNN), MobileNet Single Shot Multibox Detector (MobileNet-SSD), You Only Look Once (YOLO), and Residual Network-50 Layers (ResNet50).
The study employed two datasets, a real one and a sample. The sample dataset was designed to be similar to the real one, at least from an artificial intelligence (AI) standpoint, in terms of the curveballs and tricky parts.
The real dataset spanned a one-and-a-half year period (plants from Oregon) and split into several dynamic categories, including day of growth, disease presence and stage, cure recommendation, light intensity, carbon dioxide (CO2), soil moisture, and soil pH. In addition, the real dataset featured “unstructured” image data as visual examples of the diseases for the AI to evaluate.
Six diseases were identified and tracked throughout the plants’ growth cycles: powdery mildew, calcium deficiency, heat stress, water issues, humidity, and potassium deficiency. Each of these diseases went through three broad stages (according to consultations with farmers), making for 18 image categories. However, there weren’t enough images for calcium deficiency, so the real dataset featured 15 image categories for plant diseases overall.
The sample dataset, on the other hand, featured six categories of flowers and six categories of playing cards.
“Introducing playing cards images into the sample dataset is essential because the difference between different playing cards is very small, just like how the structure of the cannabis leaf will be the same but the diseases will be different on the leaf.”
The results were more than promising.
MobileNet-SSD achieved 85% accuracy for all five categories of diseases with three stages.
ResNet50 had 88% accuracy, the highest score, achieved when classifying two diseases (powdery mildew and water issues) across the three stages of disease. “This is mainly attributed to having more image data in those two classes of diseases. This proves that the model can be scaled up to identify more diseases along with their stages provided we give it more training data.”
The scientists add that the results are all-the-more impressive when you take into account how subtle the differences between the different stages of the diseases are. 
- Pathak K, et al. A study of different disease detection and classification techniques using deep learning for cannabis plant. International Journal of Computing and Digital Systems. 2021;10(1). Impact Factor = 0.810; Times Cited = n/a