A data driven approach for diagnosis and management of yield variability attributed to soil constraints

A data driven approach for diagnosis and management of yield variability attributed to soil constraints

Australian agriculture does not value data to the level required for true precision management. Consequently, agronomic recommendations are frequently based on limited soil information and do not adequately address the spatial variance of the constraints presented. This leads to lost productivity. Due to the costs of soil analysis, land owners and practitioners are often reluctant to invest in soil sampling exercises as the likely economic gain from this investment has not been adequately investigated. A value proposition is therefore required to realise the agronomic and economic benefits of increased site-specific data collection with the aim of ameliorating soil constraints. This study is principally concerned with identifying this value proposition by investigating the spatially variable nature of soil constraints and their interactions with crop yield at the sub-field scale. Agronomic and economic benefits are quantified against simulated ameliorant recommendations made on the basis of varied sampling approaches.

In order to assess the effects of sampling density on agronomic recommendations, a 108 ha site was investigated, where 1200 direct soil measurements were obtained (300 sample locations at 4 depth increments) to form a benchmark dataset for analysis used in this study. Random transect sampling (for field average estimates), zone management, regression kriging (SSPFe) and ordinary kriging approaches were first investigated at various sampling densities (N=10, 20, 50, 100, 150, 200, 250 and 300) to observe the effects of lime and gypsum ameliorant recommendation advice. It was identified that the ordinary kriging method provided the most accurate spatial recommendation advice for gypsum and lime at all depth increments investigated (i.e. 0–10 cm, 10–20 cm, 20–40 cm and 40–60 cm), with the majority of improved accuracy being achieved up to 50 samples (≈0.5 samples/ha). The lack of correlation between the environmental covariates and target soil variables inhibited the ability for regression kriging to outperform ordinary kriging.

To extend these findings in an attempt to identify the economically optimal sampling density for the investigation site, a yield prediction model was required to estimate the spatial yield response due to amelioration. Given the complex nonlinear relationships between soil properties and yield, this was achieved by applying four machine learning models (both linear and nonlinear) consisting of a mixed-linear regression, a regression tree (Cubist), an artificial neural network and a support vector machine. These were trained using the 1200 directly measured soil samples, each with 9 soil measurements describing structural features (i.e. soil pH, exchangeable sodium percentage, electrical conductivity, clay, silt, sand, bulk density, potassium, cation exchange capacity) to predict the spatial yield variability at the investigation site with four years of yield data. It was concluded that the Cubist regression tree model produced superior results in terms of improved generalization, whilst achieving an acceptable R2 for training and validation (up to R2 =0.80 for training and R2 =0.78 for validation). The lack of temporal yield information constrained the ability to develop a temporally stable yield prediction model to account for the uncertainties of climate interactions associated with the spatial variability of yield. Accurate predictive performance was achieved for single-season models.

Of the spatial prediction methods investigated, random transect sampling and ordinary kriging approaches were adopted to simulate ‘blanket-rate’ (BR) and ‘variable-rate’ (VR) gypsum applications, respectively, for the amelioration of sodicity at the investigated site. For each sampling density, the spatial yield response as a result of a BR and VR application of gypsum was estimated by application of the developed Cubist yield prediction model, calibrated for the investigation site. Accounting for the cost of sampling and financial gains, due to a yield response, the most economically optimum sampling density for the investigation site was 0.2 cores/ha for 0–20 cm treatment and 0.5 cores/ha for 0–60 cm treatment taking a VR approach. Whilst this resulted in an increased soil data investment of $26.4/ha and $136/ha for 0–20 cm and 0–60 cm treatment respectively in comparison to a BR approach, the yield gains due to an improved spatial gypsum application were in excess of 6 t and 26 t per annum. Consequently, the net benefit of increased data investment was estimated to be up to $104,000 after 20 years for 0–60 cm profile treatment.

Identifying the influence on qualitative data and management information on soil-yield interaction, a probabilistic approach was investigated to offer an alternative approach where empirical models fail. Using soil compaction as an example, a Bayesian Belief Network was developed to explore the interactions of machine loading, soil wetness and site characteristics with the potential yield declines due to compaction induced by agricultural traffic. The developed tool was subsequently able to broadly describe the agronomic impacts of decisions made in data limiting environments.

This body of work presents a combined approach to improving both the diagnosis and management of soil constraints using a data driven approach. Subsequently, a detailed discussion is provided to further this work, and improve upon the results obtained. By continuing this work it is possible to change the industry attitude to data collection and significantly improve the productivity, profitability and soil husbandry of agricultural systems.