Assimilation of remote sensing data into agro-ecosystem model for a real-time prediction of grassland productivity

Abstract

Grassland management in intensive agriculture environments is increasingly constrained by environmental regulations, aiming for more sustainable production. Real-time prediction of grassland productivity and timely monitoring of grasslands are essential in order to make these agricultural activities more sustainable and ensuring food security. Finding the best time for mowing is a function of biomass productivity and quality. Agroecosystem models have been proven as suitable tools for simulating grass growth under the influence of soil, climate and management. However, due to factors such as heterogeneity in plant driving variables, lack of information on soil and management data, and grass species, as well as due to uncertainty in agroecosystem models, predicting grassland productivity using these models has remained a challenge. Remote sensing (RS) have the potential to provide timely and frequent observations of the land surface at a range of spatial scales, which can be useful for measuring the productivity of grasslands. However, the approach suffers from gaps such as revisiting frequency, cloud cover, measuring limited sets of variables, and ability for future prediction. Therefore, it is desirable to combine crop growth models and RS observations using data assimilation techniques. In this study, we exploited the best properties of RS data and combined it with the predictive and explanatory abilities of crop growth models. We implemented particle filtering algorithm to assimilate LAI (Leaf Area Index) values driven from Sentinel-2 into an agroecosystem model at four case studies covering different parts of Germany. We estimated the initial uncertainty ranges of model parameters from a multi-objective uncertainty-based calibration algorithm. Additionally, we compared the suitability of different spatial resolutions in reducing the uncertainty in estimated parameters and improving the performance of grassland models for predicting grassland productivity during different time within growing seasons. We found that reducing spatial resolutions of LAI from original 10 m to over 100 m, decrease the accuracy in the estimates of the grassland productivity substantially. However, the computational time decreased which was an advantage for data assimilation at national scales. We concluded that it is important to consider reasonable spatiotemporal scales and to obtain trade-offs between accuracy and effectiveness depending on the scale of the area under study.