The way that we currently use imaging systems in agriculture involves capturing images (manually or automatically) in the field and processing them later in the lab, which is very time-consuming. I always wanted more. We have the Raspberry Pi, fancy machine/deep learning algorithms, computer vision libraries (e.g. OpenCV, PlantCV), etc. So why not combining all together and have a system that does everything from capturing images to feature extraction and analysis onboard in real-time?
Visible-bands Raspberry Pi camera is easy, cheap to assemble, however, there is only so much we can do with RGB images. So I decided to make a low-cost high-quality multimodal camera system with Raspberry Pi at its core that combines multispectral (NOT w/ a cheap blue filter!!) and thermal imaging (thermal-VIS-NIR) plus light sensor and a range finder. You can connect the ClimaVUE50 all-in-one meteorological sensor (Campbell Scientific) directly to the imager's SDI-12 port. The system has a GUI with embedded computer vision and crop models.
The software automatically generates and displays three image layers, and each layer has several options to choose from:
Layer 1 (multispectral): R+G+NIR / R / G / NIR (filter dependent)
Layer 2 (index): Red NDVI / GNDVI / Normalized Red / Normalized Green / Normalized NIR / Segmented Canopy
Layer 3 (thermal + microclimate): ΔT / CWSI / actual Transpiration / Stomatal Conductance
In addition, it measures plant canopy height, and automatically calculates canopy cover, crop coefficient, and crop ET. The imager output is a CSV file with timestamped, geotagged data.
This is an ongoing effort and I'm currently working on improving its image segmentation algorithms. To date, my focus has been mainly water and crop loss management in terms of application area, but I'm expanding it. I'm sure those in phenomics and interested in high throughput phenotyping can also benefit from this system.
I have summarized the system features and its history in the presentation below. Please free to watch.
Comments