Inverse Meta-modelling of Yield-Monitor Data for Estimating Soil-Available Water-Holding Capacities at a Farm Resolution of 10 m

Abstract

Spatially dense, geo-referenced information is an integral component of precision agriculture (PA) management. Moreover, the value of temporally dense information is gaining recognition. An example of such valuable information is crop yield data. An intuitively appealing response to these information requirements is simulation modelling. In order to meet the temporal density and the spatial extent requirements of PA, simulation modelling is faced with a major challenge: that of capturing yield variation at a spatial resolution relevant to PA. Adequate computer power to run a representative number of simulations (>1,000) and suitable information to populate the models are the motivating challenges behind this study. Inverse meta-models were derived from the agricultural production simulator (APSIM) using neural network modelling to predict soil-available water capacity (AWC). Using as many years of yield data as was available for a dryland grain farm in Australia, ‘effective’ AWC maps with a resolution of 10 m were made by averaging maps estimated from different yield years. The AWC values were validated in terms of value for predicting spatially variable yield. The AWC maps were significantly different, depending on the year of yield data used. This demonstrated that the ‘effective’ component of the AWC values contains information about climate interacting with the soil, the crop, and the landscape. The AWC values proved useful for predicting yield using simple linear models (0.48 < R 2 < 0.80) rather than using APSIM. A conclusion from this study is that the inverse meta-modelling concept is an efficient way of extracting soil physical information that exists within crop yield maps. Further research attempting to enhance understanding about the ‘effective’ components of the AWC values and to improve the temporal consistence of the AWC values is important. A greater number of AWC scenarios, more years of yield data, and the inclusion of additional information into the meta-models are possible ways forward.