Greenhouses require increased disease control and need to closely monitor their plants to ensure they're healthy. In particular, the project aims to capitalize on the recent cannabis interest.
What it Does
It's a sensor system composed of cameras, temperature and humidity sensors layered with smart analytics that allows the user to tell when plants in his/her greenhouse are diseased.
How We built it
We used the Telus IoT Dev Kit to build the sensor platform along with Twillio to send emergency texts (pending installation of the IoT edge runtime as of 8 am today).
Then we used azure to do transfer learning on vggnet to identify diseased plants and identify them to the user. The model is deployed to be used with IoT edge. Moreover, there is a web app that can be used to show that the
Challenges We Ran Into
The data-sets for greenhouse plants are in fairly short supply so we had to use an existing network to help with saliency detection. Moreover, the low light conditions in the dataset were in direct contrast (pun intended) to the PlantVillage dataset used to train for diseased plants. As a result, we had to implement a few image preprocessing methods, including something that's been used for plant health detection in the past: eulerian magnification.
Accomplishments that We're Proud of
Training a pytorch model at a hackathon and sending sensor data from the STM Nucleo board to Azure IoT Hub and Twilio SMS.
What We Learned
When your model doesn't do what you want it to, hyperparameter tuning shouldn't always be the go to option. There might be (in this case, was) some intrinsic aspect of the model that needed to be looked over.