Grapevine Nitrogen Status identification by aerial multispectral imaging

Table Grape Nitrogen Remote Sensing Digital Ag Lab

Future agricultural and food production systems must make better use of limited resources to ensure that farmers can economically produce more high-quality food for the growing world population while minimizing impact on the environment. Grapevines have lower mineral nutrient requirements than many other crops, still vine nutrient status should be monitored to help prevent deficiencies that could reduce vine capacity, yield, and fruit quality. Monitoring nutrient status can also help growers to prevent unnecessary fertilizer applications which needlessly increase production cost, and could cause unintended nutrient imbalances or environmental contamination.

A nitrogen (N) management plan helps growers make the best use of N fertilizers, and an important part of such plans involves monitoring vine N status. Tissue sampling is one of the best way to monitor nitrogen in grapevine. However, it is an imperfect monitoring method because it is based on laboratory assessment which is destructive, costly and time consuming. This is a laborious practice that, and most of the time it provides an average value for the entire vineyard. In contrast, remote sensing has the potential to provide more detailed data about entire vineyards, potentially for every vine.

UAS Digital Ag Lab UC Davis Alireza Pourreza
One of the aerial imaging systems used for multispectral image acquisition

The recent introduction of sophisticated, more affordable and user-friendly drone-based aerial imaging equipment has made it possible for growers to consider collecting frequent large-scale data for monitoring vineyards. However, what’s lacking is a well-established aerial image processing pipeline, a comprehensive calibration procedure, and accurate data interpretation models. Until those exist, it will be difficult for growers to fully capitalize on aerial imaging applications. In this project, we aim to develop a data-driven decision support tool that will enable growers to accurately monitor vine nutrient status. This tool will enable growers to use drone imagery to obtain site-specific insight about nutrient status throughout the entire vineyard. Over the last few seasons we have been pairing aerial imaging and proximal sensing with laboratory analysis of tissue samples to develop calibrated models that should enable us to use remote sensing for estimation of vine N status. The tools that will be developed in this study will be accessible via a web application that will automate most of the processing and data analytics steps and provide per-vine nutrient status at different stages of the growing season. 

In 2018, we created an experimental plot design with three different N application rates of (0, 19.3, and 48.3  g N/vine) applied in either two (Slugs) or 10 (Spoonfeed) separate applications throughout the season and with 5 repetitions. In total, we had 30 plots, each included 5 vines.

Experimental site selma digital ag lab uc davis grape n
Different treatments of nitrogen application in the experimental block located in the north side of a table grape vineyard in central valley of California

Tissue samples were collected from vines in the experimental plot and assessed in laboratory for true nitrogen status. The nitrogen content percentage were used as labels to calibrate multispectral data interpretation model. 

Grapaevine tissue sampling for measuring nutrient status
Grapaevine tissue sampling for measuring nutrient status in an experimental site in Selma, California

Study in 2018

Our main objective in 2018 was to investigate the potential of aerial multispectral imaging for detecting grapevine Nitrogen status. We flew over a Flame Seedless vineyard in Selma, CA once during the bloom and once before harvest. Tissue samples collected from each plot (of 5 vines) were mixed together and the lab analytics provided us with 30 nutrient labels for 30 experimental plots (1 label per 5 vines in a plot). Various potential vegetation indices (identified by other researchers for monitoring N) were extracted from aerial multispectral data and tested to check if different Nitrogen levels in vine will leads to distinctive vegetation indices values. The following figure lists the vegetation indices tested in this study:

Vegetation indices used in this study
List of vegetation indices extracted from 5-band aerial imagery

Correlation between leaf nitrogen and vegetation indices were calculated and plotted in the figure bellow. Blue columns show R-squared value and orange line shows p-values for each index. The best correlation was achieved when R-squared was high and p-value was low, this means the model significantly explains most of variation within the data.

Correlation between leaf nitrogen and vegetation indices at bloom and harvest
Correlation between leaf nitrogen and vegetation indices at bloom (mid-May) and harvest (early-July)

Based on the correlation results, NDRE, CGM, and MTCI showed best correlation to vine nitrogen in table grape. This graph also showed that NDVI, i.e. the most commonly used vegetation index, was not as good as NDRE, CGM, or MTCI for detecting N content. Among the 20 indices tested in this study, NDRE was the best vegetation index at both growing stages (bloom and harvest). 

Prediction of Leaf Nitrogen in Table Grape by NDRE UC Davis Alireza Pourreza
Prediction of Leaf Nitrogen in Table Grape by NDRE

The following figure shows top three vegetation indices as well as NDVI for an experimental flame seedless vineyard at harvest time. Almost all four indices indicated the Nitrogen deficient zones, but based on the regression results, NDRE among all is the most reliable .

Vegetation indices for Nitrogen stress in grapevine
Top Vegetation indices for detecting N stress in table grape

These results confirmed that vineyard managers can obtain an insight into spatial variability of nitrogen throughout the entire vineyard using drone-based multispectral imagery.

Study in 2019

Our main objective in 2019 was to further investigate nitrogen distribution within each vine. So we collected and assessed tissue samples from each vine (instead of combining 5 vines into one sample). Similar to 2018, we have 30 plots, each included 5 vines, and the lab analytics provided us with 150 nitrogen labels. In 2019 we had 5 times more labeled samples compared to 2018, that would potentially improve the robustness of the N estimation model.

We collected very high resolution multispectral imagery (altitude ~10 m) from each plot. The images were processed to segment each vine vegetation from soil and adjacent vines. We used thresholds in NDVI and EGI indices for removing background pixels, then separated each vine using the markers we placed on the ground at the time of aerial imaging.

table grape vine nitrogen by image aerial uc Davis Alireza Pourreza digital agriculture lab
Vine vegetation segmentation process using NDVI and EGI indices

The following histogram shows the distribution of N content for the 150 vines in this experiment. This histogram suggests that N application rate is not necessarily correlated to nitrogen uptake in vine since we had N deficient vines that received maximum N application rate, and also vines with excess N that did not receive any N fertilizer.

histogram N
Histogram of nitrogen content in 150 grapevine samples; colors show the application rates

Using the N content labels, we trained several supervised learning methods with samples that had either excess N or were extremely deficient. Then we tested this trained models with all other vines in the middle. The following figure shows the prediction results using top two supervised learning methods: XGBoost, and Random Forest, both provided very similar R-squared and RMSE.

Machine learning prediction of N multispectral
Prediction of vine nitrogen content by two supervised learning methods

We compared the performance of a supervised-learning method and vegetation indices for identification of N content and determination of spatial distribution of N in vines. Figure bellow shows the results of supervised-learning method [P(excess|X)] compared to NDVI and NDRE. In the top, the comparison of histograms of two vines (one deficient and one excess) shows a clear separation between the two vines, while in NDVI and NDRE, we see many overlaps of pixel values. Also, P(excess|X) clearly illustrates the distribution of N at different parts of each vine, while these details are not visible in vegetation index representations.

Comparison between vegetation indices (NDVI and NDRE) and a supervised learning method for determination of N distribution throughout vines with various N contents
Comparison between vegetation indices (NDVI and NDRE) and a supervised learning method for determination of N distribution throughout vines with various N contents


Study in 2020

Our objective in 2020 is to understand temporal variabilities in aerial data and how they correlate to actual N content at three important stages of the growing season: bloom, version, and harvest. We are working on development and calibration of a robust N content estimation model that can be used by vineyard managers in a practical manner. 


Principal Investigator: Alireza Pourreza

Project Team:  Ali MoghimiGerman Zuniga-Ramirez, Ryan Omidi

Collaborator: Matthew Fidelibus

Project Sponsor: California Table Grape Commission