Delair.ai platform enables an automated and integrated experience from uploading data to generating an actionable report analytic for crop management - such as .csv file with traits at the microplot level with your customized id to link it to your business systems. Thanks to delair.ai for Agriculture & Forestry, you can :
- Map and Scout to visually analyse the fields
- Run field inventory to characterise the crops and monitor them
- Drive precision Ag practices
In short, building and managing Digital Twins of your crops and plantations.
1. Supported Sensors
The list of supported Multispectral Sensors usually employed for Agriculture purpose can be found here.
2. What information shall I provide to process Agriculture datasets ?
- The image files including calibration panels if a 5-bands camera is used (if using a 3-bands camera, a calibration panel is not required)
- The GCPs file (if provided) in .txt or .csv format, with the coordinates system (WGS84)
- If for a Microplot Vectorization: the experimental plan with explanation of the microplots numbering
- If for Gaps & Count or Row Vectorization : the row spacing, intra-row spacing, and canopy diameter
- The project CRS (Coordinate Reference System)
- The deliverable CRS
3. Scouting Analytics
Delair.ai provides a wide range of basic and advanced scouting maps for crop monitoring. Please note that the interpretation of the Analytics (NDVI, NDRE, MCARI2...) relies on the crop parameter that could be analysed with the analytics. This will lead sometimes to direct analysis, and in other cases it will require an on-site specification. For sure, local knowledge and agronomy knowledge (type of soil, variety, agronomic practices) will help to strengthen the analysis.
During project creation step, after having browsed a multispectral dataset, and prior to launching upload, the available outputs for the used sensor will be displayed - in example below, the Photochemical Reflectance Index (PRI) will not be computed -.
Base Scouting Maps listed as Outputs are all computed by default if the used sensor and data allows for, and will be generated as layers. During this step, it is possible to check boxes to order the optional computation of Advanced Scouting Maps : CCCI, CIR, and MSAVI2 additional vegetation indices :
Once the photogrammetry process is complete, the scouting maps are made available for consultation through the Layers panel.
3.1. RGB (Bird Eye View)
To view the orthomosaic, enable RGB from the Layers left panel :
3.2. DSM (Digital Surface Model)
DSM represents Earth’s surface with the natural and built features.
Enable DSM from the Layers left panel. Use Value at Point from the right panel in order to browse height at a certain location of the map :
3.3. NDVI (Normalized Difference Vegetation Index)
NDVI gives indication of the photosynthetic activity, crop vigor and helps spot anomalies. NDVI of dense vegetation canopy will tend to positive values (0.3 to 0.8).
- Soils generally exhibit a near-infrared spectral reflectance somewhat larger than the red, and thus tend to also generate rather small positive NDVI values (0.1 to 0.2).
- Moderate values represent shrub and grassland (0.2 to 0.3)
- High values indicate temperate and tropical rainforests (0.6 to 0.8)
Enable NDVI from the Layers left panel. Examinate Statistics, or Value at Point, from the right panel :
3.4. MCARI2 (Modified Chlorophyll Absorption Ratio Index 2)
MCARI2 is employed for Green biomass evaluation. Enable MCARI2 from the Layers left panel :
3.5. NDRE (Normalized Difference Red Edge)
NDRE is used to assess field chlorophyll content, and thus nitrogen content or stress detection (at the earliest stage possible, before NDVI).
- Soil typically has the lowest values
- Unhealthy plants have intermediate values
- Healthy plants have the highest values
A sensor with the Red Edge band is required to compute this indice.
Enable NDRE from the Layers left panel :
3.6. VARI (Visible Atmospherically Resistant Index)
VARI serves to isolate soil from plant. Usually utilized if RGB sensor only, less accurate than NDRE or NDVI
Enable VARI from the Layers left panel :
3.7. PRI (Photochemical Reflectance Index)
PRI serves to identify areas of the field suffering from water stress, evident by clusters of pigments (such as chlorophyll and carotenoids).
Enable PRI from the Layers left panel.
3.8. CCCI (Canopy Chlorophyll Content Index)
CCCI serves for Chlorophyll concentration evaluation, to correlate nitrogen and chlorophyll concentration responses or feed nitrogen variable rate maps.
Enable CCCI from the Layers left panel :
3.9. CIR (Colored Infra-Red)
CIR serves to isolate vegetation from the surrounding environment.
Enable CIR from the Layers left panel :
3.10. MSAVI2 (Modified Soil-Adjusted Vegetation Index 2)
MSAVI2 is for use with early stage crops or where the canopy is not closed to help analyze crop vigor and identify anomalies (similar to NDVI, this excludes soil).
Enable MSAVI2 from the Layers left panel :
3.11. Exporting Scouting Maps
All generated maps can be exported from the Download section. They are available under .tif format in Survey section :
3.13. Single Point Data Extraction from the Scouting Maps
Draw a Point Annotation in order to extract then display indices values at that point :
Extracted values (Geospatial & Ag indices data) appear under the INFO tab :
3.13. Extracting Statistics from the Scouting Maps
In order to get Statistics on a trait, a vector first has to be created for the Area of Interest among the field, in which the statistics must be extracted. This can be achieved for example by uploading a kml file containing a polygon describing the boundaries for this area of interest. From the Uploads section, add the kml to the project pressing New / Import a New file, then select how the file shall be attached :
Select the appropriate CRS options :
After this upload operation is completed, kml becomes visible in Layers, and a Statistics icon will appear as available to be run under the Analytics menu :
Input a Deliverables suffix (optional) if wished, and tick for the traits to be computed :
During this step, you can choose to use and define a value for the Data filter factor :
Once completed, the statistics (geoJSON & csv) can be found in the Download section :
Statistics Output Examples (geoJSON and csv) :
4. Field Inventory Analytics
4.1. Microplot Vectorization
This Analytic serves to digitalize and georeference micoplot boundaries, delineating them as a first setp before measuring, benchmarking, and monitoring traits of varieties, and evaluating field trial responses at microplot level.
It automatically defines the geolocation of microplot boundaries that can then be measured and benchmarked to monitor trial responses or traits of each plant variety. The microplot customer id is linked to each microplot to enable seamless data analysis ; linkage of traits and outcomes. Input required from the customer is the experimental plan (xls format).
4.1.2. Requirements on Input
The base scouting maps have to be generated first.
All the microplots must have the same design. The experimental design shall provide the breakdown of the trial :
- Block_row_id: customer row id of the microplot, within a block or a field
- Block_col_id: customer column id of the microplot, within a block or a field
- block_plot_id: customer id of the microplot
- block_id: customer block or field name (optional)
- Number of row per microplots
- Number of row of analysis
- Plant and row spacing
- Row length
- Alley size
Example of Experimental Plan :
The output format is vector formats .geoJSON. The output contain :
- parent_id: block or field Id
- block_id: customer -block or field name
- block_row_id: customer -row id of the microplot, within a block or a field
- block_col_id: customer column id of the microplot, within a block or a field
- block_plot_id: customer id of the microplot
Enable Microplots from the Layers panel in order to display them onto the map brackground.
Go to the Download section to retrieve geoJSON file if you need to use them with third-party software :
4.2. Field Vectorization
This Analytic serves to digitalize the exact contour of cropped area within a field to know precisely the geolocation of field boundaries and its cropped surface.
4.2.2. Requirements on Input
Field boundaries need to be distinguishable from the sky (tracks, roads, alleys, bare soil). Input to provide are the rough boundaries of the field, to know where is the field in the orthomosaic.
The field boundaries with the field area as an attribute.
Enable Field Boundaries from the Layers panel :
4.3. Statistics by plots
This Analytic automatically extracts statistics around scouting maps and traits at the microplot or plot level (min, max, average, standard deviation, variance).
4.3.2. Requirements on Input
The basic scouting maps have to been computed first, plus the microplots boundaries defined. Quantitative scouting maps, as CIR and RGB are out of scope.
Statistics/Microplots are delivered for each selected scouting maps with the following attributes :
- standard deviation
Enable the statistic to be displayed from the Layers panel, and open right panel displaying properties, sectiing in map the microplot or plot to examinate.
From the Download section, following formats can be exported:
4.4. Row Vectorization
This Analytic automatically identifies rows within fields in order to assist with field counting or help guide farming equipment.
4.4.2. Requirements on Input
This Analytic is available for crops planted in row and where bare soil is visible in between the rows.
The base scountign maps have to be computed first, theoretical boundaries of field and row spacing must be provided.
Rows digitization with the following attributes :
- row id
- parent id (microplots) and vegetation length in case of microplots
4.5. Plant Height
This Analytic automatically determines plant height for phenotyping or volume estimation in forestry.
- Vegetation height map
- Extrapolated DTM if no bare soil option selected
- Statistics at field / microplots level if option selected
4.5.3 Requirements on Input
- If no bare soil survey is already available, an area of bare soil must be visible in the survey (such as alleys or tracks).
- Plants should be taller than 50 cm
- Overlap and sidelap shall be of mininum 80%
In this example, we enable Vegetation Height from the Layers panel :
Go to the Plant Height category in the Download section, that contain the geotiff and extrapolated DTM :
4.6. Plant & Gap Counting
This Analytic automatically determines plant count and gaps.
4.6.2. Requirement on Input
- It is mandatory to see bare ground in between two plants on the map, or only visible gaps will be detected
- For row crops, with 6cm of GSD and a multispectral camera, optimal corn stage will be 4 leaves, and starbud for sunflower
- With enhanced resolution, earlier stages can be targeted. The basic rule is : size of GSD <= half canopy size (ie: plant width seen from above)
- For orchard / vineyard: plant must be planted in row. if canopy is closed, only gaps will be detected. Height of the plant must be > 50cm
- For palm trees, plant must be > 50 cm
Gaps are only for plants in rows and are delivered with the following attributes :
- Gap length
- Position of gaps at line end or not
Enable Layers with trees and plant count :
Go the All other files section from the Download panel. Geojson files for counts (statistics and individual positions will be available there. Plots-stats.json file contains the count, while trees.json, in the example below, contain the individual detection positions.
4.7. Flowering Characterization (beta, not visible in UI, available on order)
Flowering automatically calculates the percentage of flowering vs. no-flowering — currently suitable for row crops with yellow flowers such as canola and sunflowers.
- Flowering mask (in app and downloadable)
- Flowering % per microplot (in app and downloadable)
4.8. Emergence Characterization (beta, not visible in UI, available on order)
Emergence automatically calculates the percentage of green (or leaves) to characterize seedling vigor — suitable for early stage crops.
- Emergence mask (in app and downloadable)
- Emergence % per microplot (in app and downloadable)
4.9. Stay Green (beta, not visible in UI, available on order)
Stay Green provides an automatic assessment of the ability for the crop to remain as a green growing plant late in the season.
- Stay green mask (in app and downloadable)
- Stay green % per microplot (In app and downloadable)
4.10. FCOVER (beta, not visible in UI, available through order)
FCOVER, standing for Fraction of green vegetation COVER, is a measurement of the fraction of ground covered by green vegetation. It therefore expresses the ratio between biomass and soil, between these two extrema :
- 0 = only soil
- 1 = we only see biomass
Practically, it quantifies the spatial extent of the vegetation.
- Biomass mask (in app and downloadable)
- FCOVER per microplot (in app and downloadable)
5. What are the deliverables that I can download locally and use in other softwares ?
You can export all the quantitative analytics such as the plant count and the height, in vector formats or CSV. The large plantations often use ArcGIS. The seed companies use the CSV because they work a lot with Excel or other softwares that are compatible with Excel.
You can also export all the maps (NDVI, NDRE, etc.) in vector formats.
That is why the link with the “customer id attributes” in delair deliverables is of high importance: to ease and enable their valorisation in customer business software.
Delair strategically partners with industry leaders in order to provide integrations between delair.ai and business software.
- Read about Phenome Networks PhenomeOne partnership.
- We also provide a connection to John Deere MyJohnDeere platform.
Contact us or talk with you sales representative if you need integration with other software.
7. Agriculture & Forestry FAQ
7.1. Is delair.ai able to distinguish weeds from corn ?
The first step before weed identification is to detect the rows and plants location in each row. Then the algorithm will identify the weeds that are located within or outside the rows, taking into account the initial position of the plants (corn for example). No qualification of the weed species is done.
7.2. Can you spot individual off-types, e.g. in a seed production field, where genetic purity is crucial ?
For the moment, we do not propose a specific solution to spot individual off-types but they can be detected with different methods, especially if they have a different color, chlorophyll content, vigor, biomass, etc.
7.3. Would the sensors be able to detect pollen shedding in flowering plants for hybrid crop production ?
We don’t have a solution to detect pollen shedding at the moment. But if the pollen shedding is correlated to a specific and significant plant color that we can detect with the sensor, it would be possible to detect this stage with our software.
7.4. Can we do DTM for irrigation systems ?
Yes, on bare soil DSM = DTM. We can do it but it’s not in the catalogue yet.
7.5. The extrapolated DTM is not in the price list. What does it come with ?
Extrapolated DTM is used when no DTM on bare soil is available to calculate the height of vegetation.
7.6. Does the vegetation height depends on extrapolated DTM (Palm inventory use case) ?
7.7. How does the tree counting model really work ?
It is based on the top and crown of each tree. We are launching a (linear) model that identifies it. The method is based on DSM.
7.8. How can we check the tree counting accuracy ?
We compare the values from the algorithm with the counting from the field (humans walking around the trees and counting them) and the "manual" photo interpretation counting.
Example : for palm trees, ground truth was done with GPS, but no geotagging of each tree (no precision GPS), and two rounds of ground truth was done on each field to assess consistency of ground truth.
Beware that it will only compare two methods results, with their own error (ie : ground truth error and drone error)
7.9. What is the accuracy of the stand counts ?
The accuracy of stand counts is around 95% with good quality images at the right stage.
7.10. Why are there no “Gap counting” examples on the palm inventory demo ?
Because it works only for plants in rows.
7.11. Can you spot the start of an infection ?
The vegetation indexes we provide based on chlorophyll content/concentration, photosynthetic activity or biomass, allow plant stress detection such as a disease. Then ground sampling is necessary to identify the disease. For the moment we do not provide models to detect specific diseases on a specific crop.
7.12. Do we have a spectrum library to identify plant disease ?
No, we don't have any library (NB : this doesn't exist).
7.13. Can you determine yield in corn or soybean by plot? If so, by specific weight of the plot and how accurate would the results be in comparison to current corn/soy plot harvesting yield measurement technologies ?
delair.ai can provide the plant density and plant height for each microplots at different crop stages for a first yield estimation. The accuracy for these indicators is usually around 95%.
7.14. What size of objects, plants, can we identify ?
We can distinguish an object if its size is 2x to 3x its GSD. That is if GSD = 8cm, we can see objects/plants of 24 cm. According to the crop, we can identify the seedling more or less early.
7.15. What is the minimum size of a microplot that can be analyzed ?
Today Delair is working with customers that have different sizes of microplots. There is no minimum size of microplot, but to digitize them we need to distinguish the plants from the alleys. We also need the experimental plan with an explanation of the microplot numbering.
7.16. Can you detect flowering of cereals and maize ?
For the moment, the flowering analytic only works with Sunflower and Canola, but we work on extending the solution to other crops.
7.17. In forestry, is it possible to assess the volume of wood (in cubic metres) ?
So far no. But we can count trees and evaluate their height and crown size, which are the basic parameters that feed volume estimation models.