Use of AI and Computer Vision in Indoor Farming: Misconceptions

I saw this catchy title a couple of days ago:

AI-enhanced indoor farming takes root at the University of ...

According to the news source, "AI-enhanced indoor farming is taking root through University's Institute for Artificial Intelligence and Data Science researchers are growing plants under an Artificial Intelligence system that is designed to identify any signs of sickness or distress in the plants", and I quote.

It's good to see they are getting involved in agricultural related research considering the fact that the aforementioned university does not have a college of agriculture. 

It's true that AI and computer vision are powerful tools and agriculture can benefit from them. What bothers me though is the fact that in studies like this the domain (agriculture) they're developing tools for is largely ignored. 

There are a number of issues I have frequently seen overlooked by people involved in similar studies (not necessary related to the aforementioned study). 

I frequently see data scientists trying to correlate NDVI (Normalized Difference Vegetation Index) calculated using multispectral/hyperspectral imaging/sensing) to plant water stress. I've also seen the confusion about NIR (near infrared: reflected radiation) & IR thermal infrared: emitted radiation from warm objects). 

I'm not going to get into the details of this, but I will mention a few points that might be helpful:

1) NDVI is good for comparing dead/diseased vs alive/healthy plants. It can only detect severe plant water stress. It cannot be used for water management. Don’t try to correlate NDVI with soil water content! 

2) Normalized thermal data in the form of CWSI (Crop Water Stress Index), on the other hand, is good for detecting plant water stress and irrigation scheduling. The CWSI can detect mild levels of plant water stress. If the right model is used it will show high correlation with soil water deficit.

3) One cannot focus only on biotic or abiotic stress, develop an algorithm and expect results in a real CEA setting. Very often, biotic or abiotic stressors have similar signs. Both need to be monitored at the same time using a combination of relevant sensors.

4) There's limit to how much response we can get from plants by tweaking the light source. At the end of the day, it is our system and algorithms that need to dance to the plants not the other way around! The plant matters and in some cases even the measurements time is determining.

5) To a data scientist w/o the domain (agriculture) expertise, the solution to everything is the black-box approach (AI/ML/DL). This is while there are established theoretical [crop] models, which can be used under any condition.

An AI model might work in the lab under the hood, and fail in the field because of so many important factors that are ignored. Ideally some of those theoretical models can be integrated into an AI approach to make it robust. 

6) AI is not a recipe for magic! For some reason I don't understand, there are people out there under the impression that AI has magical powers and can compensate for low quality sensor data.

No, sensor data is what is used to develop those AI models. Crappy sensor data will only yield crappy models! If your RGB camera cannot see it, your AI model won't! 

I have another post that discusses "how artificial neural networks are convoluting agricultural engineering". I strongly recommend reading that article to see the extent of damage to the agriculture as field due to an AI hype.