On the Effect of Multiseasonal Earth Observation Availability for Assimilation-supported Modelling of Winter Wheat Yield
Hank, Tobias1; Frank, Toni1; Bach, Heike2; Mauser, Wolfram1
1Dept. of Geography, LMU Munich, GERMANY; 2VISTA - Remote Sensing in Geosciences, GERMANY

Predominantly driven through the continuing growth of the world population, the competition between food, energy and environmental demands for bioproductive land surface is gradually becoming more severe. Sustainably increasing the efficiency of agricultural production, e.g. through smart farming practices, might level this conflict to some extent. Reliable information on crop status thereby is the key to improved crop management. Above all, site-specific management approaches are based on the awareness of spatial heterogeneities of growth conditions. Satellite-based earth observation (EO) currently represents the only technological solution able to provide spatially continuous information on the status of plant physiological variables. In order to allow for the generation of agronomical application maps, spatially explicit information on crop status is required one or two days in advance of the scheduled execution of the management measure (seeding, fertilization, plant protection, harvesting etc.). The availability of information therefore is very time-critical. Due to various restrictions associated with optical remote sensing (weather conditions, sensor availability etc.), the number of observations that may be acquired of a specific target during a single growth period normally is rather small. However, the monitored surface processes (e.g. plant growth) are highly influenced by dynamic variables (weather conditions, human interference, pests, diseases etc.) and therefore cannot be assumed to follow a linear development. Bridging the gaps between satellite observations through interpolation consequently will not allow for an adequate representation of crop growth. To overcome these temporal constraints, advanced information systems have been developed, which are based on the assimilation of EO data into process-based models of agricultural production. Model-based approaches are able to mechanistically combine spatially explicit information derived from EO data with temporally dynamic information, such as hourly weather data. They thus may provide the desired information in hourly intervals mostly independent from the date of the satellite data acquisition. Nonetheless, also assimilation-based model approaches depend on high-quality satellite observations. The conflict between the general necessity of routinely acquiring high quality earth observation in order to produce high quality agronomical information products on one hand and the difficulties of generating highly frequent (daily) monitoring series using optical sensors on the other, leads to two important research questions:

1. How many EO acquisitions actually are required during the course of one single growing season to allow for the generation of reasonably accurate agronomical information products?

2. Are there preferred periods of time during a growing season, where EO acquisitions may contribute more information to the final result as during other periods?

The presented study is limited to two consecutive growth periods (2010 and 2011) of winter wheat, culti-vated on a large farm in Northern Germany, where an exhaustive data set consisting of RapidEye and Landsat TM imagery is available. Plant physiological properties were derived from the optical EO data through inversion of the canopy reflectance model SLC (Soil-Leaf-Canopy). The land surface process model PROMET (Process of Radiation, Mass and Energy Transfer) was used for the hourly simulation of crop development, while biophysical variables derived from remote sensing were assimilated during run-time. The quality of the simulation was assessed by comparing the modeled winter wheat yield to calibrated spatial yield measurements. By gradually reducing the number of satellite observations included in the assimilation process and repeatedly comparing modeled against measured yield, the impact of observation frequency was assessed. Additionally, all possible combinations of the available EO dates were applied to the model and the model outputs were ranked according to the achieved overall accuracy. The 20 % best performing constellations were analyzed with respect to the included EO dates in order to identify observations with the strongest positive impact.

The results indicate that the most reliable match between modeled and measured data can be obtained by including the maximum number of available observations (2010: 5; 2011: 7). The average accuracy was significantly reduced, when less than four observations per season were used. It could further be detected that the highest impact on the model results of winter wheat yield was exerted by observations from the end of June, while the least impact could be observed for images recorded during the middle of May. Average impact was detected for observations from the middle of July, as well as for early images recorded towards the end of April.

Although the model results could only be validated through the variable yield, it can be assumed to some extent that correctly modeled yield is likely to be the result of equally correctly modeled intermediate variables. This is mainly due to the fact that nearly all growth influencing situations that may occur during the modeled season are affecting yield formation. Yield therefore may be considered as a variable that memorizes the course of the whole growing season. According to our findings, the research questions posed above consequently may be answered as follows:

1. At least four observations per season are required for the generation of a reliable information product on crop status, when a whole growing season shall be covered. If more than four observations are available, they should also be integrated into the assimilation process to further improve the model performance.

2. For the modeling of winter wheat yield in northern Germany, satellite observations from the end of June are ideal, followed by middle of July and end of April. Both investigated seasons agree on the minimum information content of observations from middle of May.

Our study emphasizes that global efforts of improving agricultural efficiency may strongly be supported by satellite monitoring. According to our results, global earth observation activities should aim at reliably providing at least four spatially continuous datasets during the growth cycle. Thinking globally, this would mean aiming for at least eight cloud-free global data sets per year. With a potential revisit time of 5 days, the twin configuration of ESAs Sentinel-2 will strongly contribute to this goal. With the help of continuous multisensoral earth monitoring data streams, it will also become possible to more carefully assess the growth stage dependent impact of scene selection on agronomical information products.