By 2050 the world's population is expected to grow to almost 10 billion, an increase that, according to the FAO, will increase agricultural demand by 50% compared to 2013 levels. Given the already volatile state of food security in the world today, increasing demand will only increase this precarity. Decision makers responsible for the adequate supply of food for their populations will increasingly rely on accurate predictions of agricultural production. However, these predictions are reliant on the accurate estimation of crop yields.
One method for creating these estimates is through the monitoring of phenological characteristics — specifically biomass and leaf area index (LAI). These characteristics have been used in a number of previous studies on crop monitoring, but both are labor- and resource-intensive to collect. Fortunately, the use of remote sensing can dramatically reduce these collection costs. Earth observation (EO) data can be collected remotely in a variety of spatial, spectral, and temporal scales. Combining this availability with a number of analytic techniques developed for it, EO data is an ideal solution for estimating biomass and LAI.
NASA Harvest’s Dr. Mehdi Hosseini explored this solution in a paper he co-authored alongside Dr. Omid Reisi Gahrouei [Centre Eau Terre Environnement], Dr. Heather McNAirn [Agriculture and Agri-Food Canada], and Dr. Saeid Homayouni [Centre Eau Terre Environnement]. The paper, Estimation of Crop Biomass and Leaf Area Index from Multitemporal and Multispectral Imagery Using Machine Learning Approaches, used high spatial resolution EO data from the RapidEye sensor to estimate biomass and LAI for different cash crops, specifically corn, soybean, and canola. Previous research has looked at the relationship between coarse resolution EO-derived vegetation indices, biomass, and LAI and found that this relationship could be used to predict crop yield 1-2 months ahead of harvest. However, the coarse resolution of the EO data prevented yield estimations for small fields.
To overcome this shortcoming, the team used RapidEye EO data which monitors ground conditions at a 6.5 m resolution. As vegetation grows, it’s spectral reflectance changes in response to its internal water levels. Understanding how this reflectance of electromagnetic radiation shifts, scientists can create indices that reflect different conditions of the vegetation ranging from pigmentation to sugar and protein content. To account for variations in the measuring capabilities of different vegetation indices (VI), the authors created 11 VIs for this study.
They then used ground truth measurements collected from various corn, soy, and canola fields. In-situ observations were collected from 27 agricultural fields and biomass and LAI measurements were collected from 3 locations within each field. At each location, sample crops were collected, separated into component parts, and measured both fresh (i.e. wet) and after being dried.
The authors then modeled the relationship between given indices’ values and the coincident in-situ observations using both artificial neural networks (ANN) and support vector regression (SVR) methods. Their results showed that the SVR modeled biomass and LAI more accurately than ANN. Specifically, the SVR’s Root Mean Square Error (RMSE) for predicted biomass was only 56 g/m2 for all crops and its RMSE for all crop’s predicted LAI was 0.51 m2/m2. The results of the study show that high resolution vegetation indices can be accurately used to predict vegetation biomass and LAI. Going forward, this knowledge can be used to help construct reliable yield and production models.
For more information on the study’s results, you can read the full publication here.