Application of statistical machine learning algorithms in precision agriculture

Remote sensing can facilitate rapid collection of data in agriculture at relatively low cost. Advancements in unmanned aerial vehicles and sensor technology, along with a significant reduction in the cost of acquiring data, have enabled us to collect and process remote sensing data in real time. Approaches based on remote sensing data are widely used in precision agriculture for estimating crop and soil characteristics such as leaf area index, biomass, crop stress, evapotranspiration, crop yield, and soil organic matter. These approaches typically use predictive models (e.g., linear, quadratic, power or exponential) that are based on ordinary least square (OLS) regression. However, the performance of these predictive models deteriorates when the effects of sun-surface sensor geometry, background reflectance and atmosphere-induced variations on spectral reflectance or spectral vegetation indices are larger than the variations in the crop or soil characteristics of interest. Any errors in the predicted soil and crop characteristics may, in turn, adversely affect farm inputs, farm outputs and thus the net profits. In recent years, machine learning algorithms such as artificial neural networks, support vector machines and Gaussian processes are being explored for developing predictive models for agricultural applications, especially since these algorithms are known to provide more accurate predictions than OLS. In this paper, we describe and experimentally compare the accuracy of OLS and statistical machine learning models for estimating crop water use (or evapotranspiration). We show that models based on machine learning algorithms provide significant improvement in accuracy in comparison with a state of the art energy balance model based on OLS. We use this example to highlight the potential benefits of the use of statistical machine learning algorithms in precision agriculture