Weed Warden Updates
Author: Liam Duncan
Update
This week the goal was to do tests at 12" and 16" that can be easily replicated by someone else who is trying to use the Weed Warden. The method to make the test as replicable as possible was to get a bag of Potting soil and spread it to make a consistent dirt surface that is 4' x 7' and to use a measured grass vegetation sample.
The first step is to spread the dirt to make a 4' x 7' square. A picture of this is in the figure below.
Next step is to make a consistent vegetation sample for the user. The type of vegetation chosen for this was common grass because it is available for most people. The grass was cut and zip tied in the middle to make a sample that is about a 3" diameter when looking at it from above. Some pictures with measurements are found below.
The sample is then "Planted" in the square of dirt and the cart with the Weed Warden is set up at the desired height so that the entire setup looks like the picture below.
Once the setup is finished the tests can be performed. To perform the tests the steps below must be followed:
- Position the Weed Warden above the dirt with no vegetation in view below the sensor.
- Turn on the Weed Warden and wait for roughly 30 seconds while the system calibrates (if the buzzer is hooked up to the system the Weed Warden will beep loudly for 5 seconds when the calibration is done).
- Now that the calibration is done, the cart can be rolled over the grass sample to trigger the 5V/12V relay.
- If the cart is rolled back and forth over the sample, and the NDVIB value is plotted it will generate a graph similar to the one seen below.
Author: Liam Duncan
Update
This week we got some Bull Thistle, Sow thistle, and Narrow Leaf Plantain from the Lewis Brown Farm to test the sensor with. As per the requirements from last week, the weeds were cut into roughly 3" x 3" squares so that they would be about the size of a baseball. It has also been decided that the ideal height for the sensor is at 2 feet so that is where we did all our testing today. We have also started to do rolling tests to better simulate a real life scenario. This means that the cart is rolled between an area with no vegetation and an area with vegetation to simulate real life. An example of the testing setup can be found in the image below.
The tests done with the narrow leaf plantain used the sample below that measured roughly a 3 inch square.
The results from the testing can be found in the graphs below.
The narrow leaf plantain sample was easily detected at 2 feet as can be seen in the graph above.
The tests done with the Sow Thistle used the sample below.
The results from the testing can be found in the graphs below.
The Sow Thistle sample was easily detected at 2 feet as can be seen in the graph above.
The tests done with the Bull Thistle used the sample below.
The results from the testing can be found in the graphs below.
The Bull Thistle sample was not as easily detected as the other samples at 2 feet. This may be because the sample has a slightly smaller area or it could be because of the type of weed. Further testing will need to be done to determine the exact reason for the difficulty detecting this sample.
Next steps for the project include adding a 12V->5V buck converter to the system, doing some more tests with even smaller weeds, and solidifying the project documentation.
Author: Liam Duncan
Update
Some Highlights from this week include running some tests at 1 foot off the ground with a 5-inch grass bundle and a meeting with one of our stakeholders to get a better idea of what the requirements for the system are. The 1-foot tests showed that the system can detect exceedingly small vegetation when at lower heights and the stakeholder meeting gave a better idea of how the system will be deployed and what type of terrain it will need to tackle.
The picture below shows the sample 5-inch piece of grass that was used to do the 1-foot tests.
This picture shows the testing setup for the Weed Warden at 1 foot off the ground.
The Final picture is the graph generated from the 1-foot testing showing that Weed Warden has the capability for detecting small weeds when it is closer to the ground. On the graph, the grey line represents the NDVI indicator line, the green lines represent the presence of vegetation underneath the sensor, and the orange and blue lines represent thresholds for classifying the sample as vegetation. Simply put, if the grey line crosses the blue line or orange line then the Weed Warden will classify it as vegetation.
After the meeting with our stakeholder, we have gained some valuable information about how this system will be used.
Some important points from the meeting are:
- The desired height for weed detection will be at 2 feet for a tractor boom and 1 foot for a rover
- The Weed Warden should be able to detect plants that are the size of a baseball
- The most common plants that will need to be detected are Russian Thistle, Prickly Lettuce, Kochia, and Tumble Mustard
- Dust will be a problem and there should be a way to detect if there is dust covering the sensor
- If attached to a tractor, the minimum speed will likely be about 4mph
- The system should be shock and dust resilient
- The system should be powered by a 12V DC jack
- Want to think about ways to attach an array of sensors to the main controller
With the latest information about how the Weed Warden will be used it will allow us to fine tune the system for the desired requirements. This means that we should obtain some of the plants that are most common weeds, fine tune the algorithm to work at 2 feet with baseball sized weeds, and install a 12V jack with a DC-DC 12V to 5V converter. We would also like to investigate an effective way to attach multiple spectral triad sensors to the main controller. This should be fairly easy because they are all controlled by I2C.
Author: Liam Duncan
Update
Since the last update, the graphs of the data have been revised to better show the workings of the Weed Warden, there was a Mock stakeholder update presentation that can be found here, and the abstract as well as a few other parts of the HardwareX paper have been written.
The graph above shows testing data with the sensor at 3 feet off the ground using a large bundle of grass. The NDVI value from the spectral sensor is in grey, the threshold value in blue, and the grass indicator in green. This graph represents the information much better than previous graphs by adding the threshold calibration value and the grass indicator so that the accuracy of the model can be observed.
Author: Liam Duncan
Update
Since the last update, a calibration sequence has been added to the software to make readings adaptable to different environments. The calibration sequence takes ten scans of a bare patch of dirt, averages their NDVI values, then adds .02 to the average value.
The idea behind this is that the spectral sensor is sensitive to changes in lighting conditions and environment, so it is necessary to calibrate it in its current environment. The rationale for adding .02 to the baseline NDVI value is to prevent false positive triggers from small fluctuations in NDVI values.
The graph above shows testing data with the sensor 2 feet above the ground. The high points on the graph are when a bundle of grass was placed under the sensor and the low points are when there was no grass under the sensor. This data is promising for the future of Weed Warden because the difference between grass and no grass are clearly defined. This is an ideal dataset and not all dataset are this well defined.
The System is not perfect. As can be seen in the picture above, the NDVI values are not always clearly high or low. This could be due to a number of factors such as the effective area of the sensor at 3 feet, interference from moving the grass bundle in and out of the field of view of the sensor, or changes in environment because this data was taken while it was raining. A more consistent testing procedure should be developed to eliminate as many external factors as possible. Also the effective area of the sensor should be tested at three feet and at 2 feet.
Author: Liam Duncan, Brendan Miller, Colton Shaw, and Alyssa Berquist
Update
Lead by our new member Liam, the software has been updated to the latest release of Loom. The SD stores data for the current light state, NDVIB, EVI, and PSND values. The current light state allows us to know if the IR light was on or off. The device is being tested on a wagon with a shade cover. We are only using one IR light, but that may change with more testing. Details on the NDVIB, EVI, and PSND can be found here: Possible Indices Used on Weed Warden
We are testing the minimal amount of green that the sensor is able to detect over a dirt landscape. We are also testing if an interchanging an IR light and testing other possible indices for any notable differences.
An image of a recent testing setup can be seen above. We are using a large whiteboard to provide additional shade cover. The sensor is elevated three feet to match the requirements. The IR light is turned on for one data collection cycle then is turned off for a data cycle and repeats. Currently, we are unable to detect plant life a quarter in diameter.
Author: Alyssa Berquist, Brendan Miller, Colton Shaw, and Zachary McLaughlin
Update
Due supply chain issues caused by COVID-19, many of the necessary components needed to upgrade our second Weed Warden unit have been delayed. Until the shipments for these parts come in, this unit will be on standby. Similarly, the PCBs have yet to arrive. Access to the machine shop has been restricted until recently, so adjustments to the enclosure will be made once permission is granted. Without the ability to further prototype our units, our focus shifted towards en masse data collection and testing polarized film in combination with the triad sensor.
In this experiment, we examined the effect of polarized film on the spectral triad sensor. Interest in the use of polarized film for the Weed Warden started due to the use of edgebands in NDVI calculations, with a secondary interest in preventing saturation. Polarized film blocks out certain electromagnetic spectrums (light) by absorbing different wavelengths depending on the angle of the films in relation to each other. The figure above animates this process. Theoretically, by adjusting the angle between two pieces of film, the amount of light blocked can vary.
The most surprising finding was that the edgebands (blue and NIR) spectrums on the 45° film were increased, whereas on the other films they decreased the edgebands but were still the closest values to 0. The rest of the visible color spectrum was substantially blocked from the spectral sensor, with one of the more absorbed wavelengths (730nm) being reduced by ~80%, ~50%, and ~65% in the 0°, 45°, and 90° films, respectively.
These films should be used in conjunction with the spectral sensor when in need of higher weighted spectral value for the green wavelength of 585nm, less weighted spectral value for some red, NIR, and non-585nm green wavelengths, and/or inflating the outermost edgebands: 430nm, 850nm, 900nm, and 940nm. When choosing which of the three films to use, the average absorptions should be considered, with ~20% increments from lowest to highest film absorption: 45°, 90°, 0°.
Author: Brendan Miller, Zachary McLaughlin, Alyssa Berquist, and Colton Shaw
Update
Creating a watertight sensor housing has been the focal point of mechanical development. The 3D model, pictured to the right, is in the process of being manufactured and has had a series of iterations with minor tweaks 3D printed before testing for water tightness.
The current iteration of the case, once assembled, will be tested for water tightness. The lid of the case is to be made out of acrylic, and requires a routed groove to fit an o-ring to prevent water from leaking in through the lid. Routing this groove into the plastic requires a milling machine, which we do not have access to during the current COVID-19 pandemic. Nonetheless, we are still testing the housing with a temporary 3D printed version of the lid, pictured next to the acrylic lid.
Overall, progress on the sensor housing is promising despite some setbacks due to COVID-19. Temporary workarounds are allowing us to move forward on making the best design possible for the Weed Warden project. Pictured to the right is the housing unit assembled without the lid and electronic furnishings.
A custom PCB was designed to be able to house connectors for the Triad sensor and the TTL sensor. This PCB would be attached to an Adafruit Feather M0 board and an in-house PCB called Hypnos for power and data logging functionality. The board is fairly simple, as noted by its simple task, but it provides a substantial benefit in easing the complexity of the system. With this board, the Weed Warden would be simply a plug-and-play system that can easily be disassembled, connected, and handled. Before, the Weed Warden had made use of up to six boards with differing functionality and loose connections purely soldered on. There was no relative ease of plugging in sensors and the system itself provided difficulty in understanding what each board did and how to operate them. The form factor of this previous system was also very cumbersome and occupied a large amount of space within the housing. This new system, with the new PCB and the Hypnos/Feather combination should allow for a smaller form factor, ease in understanding and handling, and a more efficient system.
We are currently determining whether a normalized difference vegetation index (NDVI) would help detect live green vegetation. To test this theory, the Weed Warden will be held over areas of grass about three feet off of the ground. The device will be sensing during a long interval, and it will collect data on 30-minute intervals.
The individual wavelengths measured by the triad sensor will be summed up, then each wavelength will be divided by the total wavelength. The NDVI is calculated by summing the infrared bands (730nm and 760 nm) and the blue bands (435nm and 460nm), then subtracting the red bands (645nm and 680nm). This value is then divided by the sum of the infrared, blue, and red bands. The acceptable NDVI range will vary depending on the total amount of light received by the sensor.
Author: Brendan Miller, Colton Shaw, and Alyssa Berquist
Update
An initial design for the case has been printed. We are now adding the TSL 2591 light sensor to measure the amount of sunlight while the spectroscopy sensor continues to read the light reflected from the ground. The data from the TSL will help us calibrate Weed Warden to work under any light conditions.We are also calculating the normalized difference vegetation index (NDVI) and the NDVI with the blue wavelengths. Testing will be done to see if the NVDI will help Weed Warden differentiate between what is a live plant and what is not.
A new member has joined our team! Zachery McLaughlin is an Electrical Engineering student here at Oregon State University, and he will be assisting the electrical tasks of this project.
Author: Brendan Miller, Colton Shaw, and Alyssa Berquist
Update
A string of in-lab testings have been conducted in order to have lighting as a control. The tests concluded in erroneous data collected by the triad. We honed in on the problem which was insufficient lighting. Most tests will be conducted outside.We have created an algorithm that can now distinguish artificial green from plant green!
The weed warden enclosure has been drawn up and is ready for printing. Thanks Alyssa!
Author: Brendan Miller, Colton Shaw, and Alyssa Berquist
Update
Back from the holiday, our team is back in full throttle.In the coming weeks, a new enclosure will be designed and printed using a 3D printer.
Future development of Weed Warden is dependent on the shipment of the UV lights that will be mounted onto the Weed Warden. An unexpected issue arose with the last shipment of UV lights, delaying our testing of our near-IR detection algorithm.
Pseudocode has been developed for the Weed Warden, as shown below.
Author: Brendan Miller and Colton Shaw
Update
This week we have added Alissa Bergquist onto our WeedWarden Project. She will be working on the mechanical aspects of the project. Welcome to the team Alissa!Gathering as much data with the triad and TTL have been the main task this week. At the start of next year we will have another member join us who will be involved in creating an algorithm to enhance our weed detecting ability. In order for this to happen, massive amount of data must be collected.
Author: Brendan Miller and Colton Shaw
Update
* New software and hardware components are being implemented and await testingA set of LEDs will be mounted onto the outside of the enclosure in order to increase the detection of vegetation. With data collected from testing with these new lights, we will be able to calibrate our software.
The aforementioned PCB design has been completed. The near-infrared sensor is still in development.
Author: Brendan Miller and Colton Shaw
Update
A schematic for the WeedWarden is being developed in order to produce a PCB to consolidate wiring for our sensors.Field tests have been conducted successfully and the data collected has been added to our database which will be used for further tailoring of sensing.
The near-infrared sensor is still in development.
Author: Brendan Miller and Colton Shaw
Update
Plants reflect green and near-infrared light, both of which our Weed Seeker has the ability to sense. Instead of using only UV light to sense weeds, we will be collecting data to incorporate the visible and IR spectrum.WeedWarden is now using Loom 2, and after every ten seconds a picture is taken and data is stored. Both the data and pictures are clearly labeled to distinguish them from the data taken every two seconds.
A wagon mount was built to easily use the two sensors side by side and remain at a constant height. The sensors are spread three feet apart and are two feet off the ground.
Author: Brendan Miller and Colton Shaw
Update
We are currently in transition from Loom to Loom 2.The second prototype of the WeedWarden is complete. This second WeedWarden will allow us to field tests more effectively by doubling our ability to gathering data.
Author: Brendan Miller and Colton Shaw
To resolve this issue, I looked at how I split the camera setup and the snapshot function. The issue was that no new pictures were actually being taken, and the pictures would either be repetitive or would not save properly. By splitting up the functions the camera would not reset after each picture and would send the same or corrupted image to be saved. Instead of keeping these split, combining the functions allowed the camera to reset properly in order to take a new picture.
Author: Brendan Miller and Colton Shaw