Statistical Characterization of Plant Biosensor Data: Distributional Properties and Inter-Variable Relationships

Plant Biosensor Data: Distributional Properties and Inter-Variable Relationships

Abstract

Precision agriculture is now increasingly being monitored with continuous biosensors. But many datasets are analyzed for predictive modeling without any basic statistical properties of the data being characterized, which is required to be able to determine whether it is appropriate for downstream use. Objectives: To characterize the distributional properties of the key environmental and soil biosensor variables, to assess the linear relations among variables, and to determine if the biosensor readings vary systematically across individual plants of a biosensor-monitored dataset. Methods: A de-identified dataset of plant biosensor readings was used, consisting of 120 readings per plant from 10 plants spanning an approximate 6-hour time interval. The soil moisture, ambient temperature, soil temperature, humidity, light intensity, soil pH, nitrogen level, and phosphorus level were analyzed. Descriptive statistics were conducted, normality was tested, and ANOVA was used. Results: All variables showed very peaked and leptokurtic distribution shapes (skewness and excess kurtosis values close to zero). The patterns here are not a formal goodness-of-fit test against a uniform distribution, so they are not definitive statements of uniformity, but rather hypotheses that should be considered. Conclusions: The distributional patterns of the biosensor variables were very symmetrical and flattened (indicating uniformity) and had very low inter-variable correlation. The statistical characteristics in these readings indicate that single-timepoint readings alone might not be suitable for predicting an outcome in a meaningful way without extra labeled outcome data and engineered temporal properties. The results provide a baseline of statistics to be used in subsequent research with this data.