Plant Counts with PlantAI™

Plant AI is a new, machine learning-powered tool, that allows extracting plant-level data in the whole field. The tool provides such insights as total number of plants in the whole field, as well as area, diameter and plant health for each plant or tree. The tool can be used on a variety of crops ranging from corn and soybean to vegetables and trees. The main requirement is that each individual plant must be clearly visible in the imagery. Because of that, the crop growth stage and the altitude at which the imagery is collected must be carefully chosen for most accurate results. Cereal and oilseed crops like wheat or barley usually are not well suited for this tool because the plants are too small and overlap each other even in the early growth stages.

The workflow is divided into 2 steps. The first step is to identify plants in the imagery and extract their shape. The second step is analytics where plants can be classified by either size or plant health to more clearly identify potential variations.

Below is step-by-step guide for the different steps.

Open the processed upload that you would like to run the Plant AI on. Switch to the Analysis tab and toggle "Plant Counts with Plant AI" tool. 

There are 2 options available here:

• Create Counts - use the workflow described further in the article to identify plants in the imagery.
• Request from Solvi - let our support team do the work for you. Once requested, the results will be available usually within 24 hours.

The rest of this article will describe the workflow for the first option "Create Counts"

1. Provide examples

In the first step, you'll need provide at least 10 examples of the plants that you are looking to identify, so that algorithms can learn which plant to identify. 

To do so, first, adjust the blue example area so that it covers at least 10 plants in one or two rows and after that, click on the New Circle-button in the top toolbar to activate the drawing tool and draw the circles on top of each plant by first clicking in the middle of a plant and then again on the edge. 

It is very important that all plants within the outlined area are annotated with a circle. Everything within the area that is not covered by circles will teach the algorithms to ignore, so if there are any weeds or other unwanted objects between the rows, it can be a good idea to include them in the example area.

Once finished outlining the examples. click "Continue" to proceed to the next step.

2. Choose analysis and detect plants

Next, choose what kind of output you would like to generate - basic counts or counts with plant-level data. 

For smaller crops like corn, sugar beets, soybean or early-stage vegetables the exact size of each plant is usually not of interest and choosing "Basic Counts" will usually yield more accurate results if the plants are close to each other and are even touching. 

For pre-harvest vegetables or trees when the diameter of each plant as well as plant health is interesting, choose the "Counts with Plant-level Data"-option.

The larges blue area on the map allows you to choose which part of the field you want to make a test on. Drag it around to choose a part that represents best the area of interest.

finally, click "Detect Plants" to run the detection within the sample area of interest. This process may take up to 5-15 minutes depending on the imagery.

3. Review the results for the sample area

Once ready, you will see the results on the map. Every plant that was identified by the algorithms will be marked with a dot if you chose "Basic Counts" and with a shape outline if you chose "Counts with Plant-level Data".

If most of the plants have been identified accurately you can choose to proceed to the next step where you'll be able to identify the plants in the whole field. 

Usually, some plants may be missed or other unwanted objects like weeds can be detected as plants. In the next step you'll be able to address such issues by providing more examples in the areas with inaccuracies and improve the results.

4. Identify plants in the whole field  

In the next step, adjust the boundaries to cover the whole field and run the detection again. As you adjust the boundaries, the price will update automatically if the boundaries exceed the minimum area of 5 hectares. 

By specifying field boundaries as accurately as possible you can avoid the unwanted objects outside of the field from being identified. 

5. Review the results for the whole field

Now that you have the results for the whole field, you can either proceed to the final step or improve accuracy by adding more examples.

If some of the plants were missed or other objects have been detected mistakenly, click on "Provide More Examples" button to get back to the previous step and add new examples in areas with inaccuracies. 

You can repeat the process described above as many times as needed. Usually, the more examples you provide the more accurate the results will be.

In the “Review Results”-section the two latest detection results will be shown so you easily can see if there is an improvement or not.

6. Proceed to Analysis

Once you get satisfactory results, click on either "Latest count" or "Previous count" to select and then choose "Proceed to Analysis". This will make the results available in the main Plant Counts-tool where you'll be able to see the differences in metrics like size or plant health and export the results as a SHP-file.

7. Review and Analyze

After you've completed all previous steps, the results will appear in the Plant Counts tool. 

If you have chosen "Counts" output type, you'll be able to see the total number of plants and use the "Measure" tool to get the counts in specific areas of the field. 

If you have chosen "Counts & Size Estimations" you will also be able to classify all plants in 3 different classes by Plant Health, Diameter or Area.

Under the Export-tab Plant Counts can be downloaded as a Shape-file format.