Speaker
Description
Introduction
Crop canopy reflectance is often used as a proxy for crop vitality. While it relatively easy to identify low vitality spots through vegetation indices (e.g. NDVI, WDVI, etc.) automating the identification of the causes of the low vitality spots remains an unsolved challenge. In fact factors that can cause a drop (or an increment) in vegetation indices, for example water and nitrogen abiotic stress, and biotic stresses like soil and air-borne diseases, and weeds. The objective of this project is to create a model to detect the presence low vigor (e.g. poor spots on an NDVI map) and identify its cause for potato crops.
We are developing a hybrid model (Scientific ML) composed a recurrent neural network trained on a synthetic data set generated using a potato growth model (Tipstar), coupled to a canopy reflectance model(PROSAIL). The model will consume time series data of multispectral signatures as well as data on crop management (e.g. fertilization, water stress, maturity class), weather and soil to facilitate the identification of the anomaly. The appearance of different stresses at different times in the season will be a major driver of the predicted stress factor, for example low emergence will cause an initial decrease in canopy vigor indicators — like NVDI — that will decrease as the season proceeds and canopy will close. The model will be validated on experimental data described below.
Materials and methods
Field experiment dataset
In the first year two cultivars were cultivated over 6 plots without replicates and canopy reflectance was monitored bi-weekly through a hyperspectral sensor and a drone multispectral images, along with final yield and bi-weekly SPAD measurements. In the second year (2024) we are repeating the experiment with three replicates and two sites. Preliminary results from the second year of experiment will be presented.
Synthetic dataset
We generated a synthetic dataset of crop growth and its spectral canopy under multiple stress factors. Crop growth was simulated using Tipstar — a potato growth model — whereas canopy spectra was simulated using Prosail, a radiative transfer model.
The first dataset is a factorial in-silico experiment with the following factors: two sites, 10 different growing seasons, two cultivars (one early and one late cultivar), 5 stress factors.
No interaction between the stress factors was simulated (e.g. no low emergence and water stress in the same simulation).
The factors that were accounted for in the model are the following:
-
Control was simulated as potential yield (no stress factor) by adding ample nitrogen fertilization and irrigation events
-
Water stress was simulated by removing irrigation events
-
Nitrogen stress was simulated by removing nitrogen fertilization events
-
Low emergence was simulated by halving the number of emerged plants
-
Weeds — a process that is not present in the crop model — was simulated by increasing the leaf area index input in the prosail parameter at the beginning and at the end of the season.
Modelling
The models are trained on the synthetic dataset and will be validated on the observed dataset. A first model serving as baseline for model comparison was developed based on difference between the observed NDVI and the NDVI of the simulated causes of anomalies.
Results from more complex models, based on recurrent neural network classification and multiple inputs (time series of reflectance at different wavelengths instead of crude indices like NDVI, management information) algorithms will be presented.
Results
Data from the first year experimentation (Fig.1) indicate that the time development of the different stresses, weed stress results in higher NDVI at the beginning of the season, nitrogen stress results in lower NDVI through the season, lower emergence in an initial decrease of NDVI that is recovered as the season proceeds, and individual plants expand.
Preliminary results from the baseline model indicate a good model fitness, with more difficulties in identifying the cause of stress in the middle of the season because of the saturation of the NDVI signal. Results from more complex models will be presented.
Figure 1: Results from year 1 of experimentation. NDVI (top) and NDVI percentage change compared to control for the two cultivars (Avenger, a late cultivar and Frieslander an early cultivar).
| Keywords | hybrid ML;crop anomalies;crop growth models; AI; potato |
|---|
