SustaiN: A Decision Support System for Sustainable Nitrogen Management in Corn and Sorghum using Satellite Remote Sensing

The overall methodological framework for this study is illustrated in Figure 1.

N-Status index development

Experimental sites

The goal of this project is to develop a suitable N-status index to accurately explain N-variability from crop canopies. Our collaboration with the Donald Danforth Plant Science Center granted us access to their corn experimental sites in Ofallon, MO in 2021 and 2022. These experimental fields consisted of two corn fields, each established with around 50 distinct genotypes. Each field included a 3m by 2m plot with two rows of corn. Additionally, we will have access to the experimental sites of corn and sorghum in Fischer Farm, MO, for the 2023 growing season. Our collaborators will establish several variable N-treatment plots to understand the N-variability from the analysis.

Data collection

UAV data collection

We collected UAV-hyperspectral data above the experimental fields in 2021 and 2022 using the Headwall NanoHyperspec 12 mm push-broom sensor mounted on the DJI Matrice 600 Pro hexacopter. This sensor provides high spatial and spectral resolution data across the very near-infrared region (400 nm to 1000 nm). We selected the hyperspectral sensor for our study since the new PS satellite has 8 bands and the UAV hyperspectral bands can be easily matched with the PS data. Table 1 provides the dates of data collection from different fields and corresponding sample data information.

Table 1: Data collection dates in the 2021 and 2022 season

Nitrogen data collection

To validate the different spectral features from remote sensing data and crop nitrogen status, we measured the amount of nitrogen in the plant leaves using the Dualex Scientific 4 instrument. This instrument accurately measures the Nitrogen Balance Index, an indicator of overall nitrogen status in plant leaves. We randomly selected different sample plots in the same UAV data collection dates and collected four Dualex readings from well-sunlit leaves. We then averaged the four sample readings and corresponded them to the plot average N-status.

Spectral sensitivity analysis

To evaluate the spectral response to N-variability, we selected 8 channels that best represented the wavelength of PS scenes and calculated different spectral indices based on the combination of these bands. We computed the difference index (DI), ratio index (RI), and normalized difference spectral index (NDSI) for all pairwise combinations of the bands using the following equations:

DI = ρ_a - ρ_b

RI = ρ_a / ρ_b

NDSI = (ρ_a - ρ_b)/(ρ_a + ρ_b)

where, ρ is the reflectance of wavelengths a and b. To identify the potential N-Status index for the decision support system (DSS), we calculated Pearson's correlation coefficient between the spectral indices and the corresponding NBI data collected from the field. This analysis allowed us to identify the most promising spectral indices for accurately determining N-variability and developing the N-Status index for our DSS.

Decision support system

The decision support system (DSS) for SustaiN was developed using Python and utilizes different frontend and backend technologies. The frontend of the DSS provides an interface for user interaction, while the backend is the actual computation platform that performs all the calculations.

PlanetScope (PS) images

The key idea behind the DSS is the automated download and processing of PS images. The PS images have eight narrowband channels (Table 2) available from August 2021, providing high spatial resolution data (3-m) that allows researchers to monitor crops more accurately. PS images are collected by a constellation of approximately 185 mini satellites known as Planet Doves that together collect images of the entire earth almost daily. However, the use of daily images is often hindered by cloudy conditions during data collection. Still, PS offers one of the best solutions specifically for crop-related studies.

Access to the DSS

Access to the DSS is limited to registered farmers and researchers only. To register, farmers can fill out a questionnaire available on the project website. Based on their answers, the user will receive login information via email (login ID and a randomly generated password). Each time the user intends to access the DSS, the login credentials are required. The user IDs and passwords are securely hosted in the database services provided by deta.sh. Since this project is a pilot approach to democratize remote sensing data, open access for everyone is out of the scope of this project.

Frontend

The frontend of the DSS is built using the Streamlit framework and provides an interface for user interaction (Figure 2). Once logged in, users provide two major pieces of information: 1) a boundary defining the location of their intended field or area of interest (AOI), and 2) the date for which the information is requested. The boundary can be drawn on an interactive map powered by the Folium maps. An address finder in the top right corner of the map helps pinpoint the location of the farmer's field, powered by the Open Street Map. The boundary must be less than 5 square kilometers; otherwise, the process cannot be started. The date is limited to the years 2022 and 2023, so the focus can be given only to recent farming practices. Once the two inputs are defined by the user, the process can be submitted, and the backend of the application handles the rest of the work.

Backend

The backend of the DSS was developed using a combination of open-source geospatial libraries in Python, enabling seamless integration with the frontend of the application. Upon receiving the user's inputs, i.e., geometry and date, the backend uses the Planet API and Python to automatically identify suitable PS satellite images by filtering out images that do not meet the user's criteria. The API searches for images within a 15-day timeframe from the user-provided date and selects the best image based on visibility percentage provided by the PS satellite, which is calculated based on the availability of cloud and haze within each scene. It should be noted that the highest visibility score does not guarantee that the requested geometry will not contain any cloud or haze.

After downloading the image, the backend calculates two vegetation indices, the Normalized Difference Vegetation Index (NDVI) and the N-status index developed earlier, from the PS image. The NDVI is commonly used to highlight healthy vegetation from remote sensing imagery and can be used to mask out unwanted objects within the user-defined geometry, such as roads, buildings, waterbodies, and forests. However, the range of NDVI values for healthy vegetation can vary depending on the specific crop and environmental conditions. Therefore, in the results section, the user can dynamically adjust the NDVI range to mask out unwanted objects within the result.

Using the adjusted NDVI range, input geometry, and date, the backend generates an interactive map of the N-Index overlayed on top of the basemap imagery. The user can toggle between the generated map to highlight major N-deficient areas suggested by the N-Index. The user can also interactively adjust the NDVI range to adjust the number of unwanted pixels within the AOI. Finally, the information can be downloaded as an HTML page in the local drive by clicking the download button. The output will include the interactive map along with other metadata information.

Cloud integration

The DSS is hosted within the Heroku framework, which is a popular cloud platform that allows easy deployment of web applications. The Heroku platform provides a wide range of services, such as automated deployment, scalability, and database integration. Therefore, when the traffic increases for the application, we can easily scale it up to match the demand. In addition, the Heroku platform provides built-in monitoring and logging services, which allow for easy identification and resolution of any issues that may arise during the deployment and operation of the DSS.