Drones for Plant Health Monitoring.

Data Labelling
Object Detection using YoloV4
Object Tracking/Counting using DeepSort
Saved Snapshots of Individual Crops

Inspiration.

One very prominent problem that has plagued medium to large scale farming (especially crop production) is the issue of adequate plant monitoring and proper data collection. When crops are planted across a very large expanse of land, It becomes very challenging to monitor these crop fields as they grow from the seedling stage up until harvest. Currently, the existing method for plant monitoring in most parts of Africa is Manual Inspection, which involves walking the lengths of large acres of cultivated land fields and manually collecting data and noting observations, which is a lot of stress and a waste of useful time.

What it does

One of the recent developments in modern agriculture is the use of drones, as they’re expected to become an integral part of future precision agriculture (PA). Our solution combines the use of Artificial intelligence and drone technology to automate the process of disease detection and plant health monitoring while simultaneously collecting valuable plant data. This will provide a countermeasure to pests and diseases, timely pesticide/herbicide application and ultimately, increased yield at the end of the day.

Our attempt to solve the problem stated involves using a drone programmed to fly along a specified path, the drone transmits the video stream to a laptop for processing via wifi and the following is done in real-time

Object detection is performed on each frame of the video stream to locate the crops by drawing bounding boxes.
Next is Object Tracking which associates a unique id to each crop and also keeps track of the number of crops on its path i.e counting individual crops.
lastly, the program takes a snapshot of the detected crops for data collection and further examination by the farmer.

How we built it

Data Gathering and Annotation.

To get enough images for this task, we took frames from videos of the beans field used for testing taken by our drone (Dji Tello). Image Labelling was done using LabelImg to prepare the images for custom object detection.

Custom Object Detection using YoloV4.

The object detection algorithm used was the YoloV4 which offers great training speed and impressive performance on small datasets.

Object Tracking using DeepSORT.

Deep SORT is a recent algorithm for tracking and has shown remarkable results in Multiple Object Tracking(MOT), where each frame has more than one object to track.

Challenges we ran into.

Limitations

Storing crop location as metadata
Constant crashing due to high wind speed.
Low video quality
Compute resources to train models.

How do we plan to solve these limitations.

Since most of the limitations are from the drone used, To solve this we already started building our own custom drone, using 3D printed parts. This will help us;

Add GPS capabilities, which will help us get location data
Mount better cameras.
Better wind resistance because of its weight and larger motors.

Accomplishments that we're proud of

We feel very excited to have built a solution out of the little resources we had, It took us weeks of frequent visits to our university farm to put all these together. And we'll definitely keep improving until we have a market-ready prototype.

What we learned

We learnt to collaborate with farmers on how to identify needs, strategise solutions, measure effectiveness and implement solutions using modern technologies like drones and AI.

What's next for Drones for Plant Health Monitoring.

We plan to perfect what we have and commence testing for a full crop cycle, with cowpea (the only crop our model can recognize for now) which takes 60 - 90 days to harvest. Doing this will help with concrete figures and statistics, and will also help to find flaws and limitations in our design choices.