- Statistical models for fruit detectability were developed to provide insights into preferable variable configurations for better robotic harvesting performance. The methodology includes several steps: definition of controllable and measurable variables, data acquisition protocol design, data processing, definition of performance measures and statistical modelling procedures. Given the controllable and measurable variables, a data acquisition protocol is defined to allow adequate variation in the variables, and determine the dataset size to ensure significant statistical analyses. Performance measures are defined for each combination of controllable and measurable variables identified in the protocol. Descriptive statistics of the measures allow insights into preferable configurations of controllable variables given the measurable variables values. The statistical model is performed by back-elimination Poisson regression with a loglink function process. Spatial and temporal analyses are performed. The methodology was applied to develop statistical models for sweet pepper ( Capsicum annuum ) detectability and revealed best viewpoints. 1312 images acquired from 10 to 14 viewpoints for 56 scenes were collected in commercial greenhouses, using an eye-in-hand configuration of a 6 DOF manipulator equipped with a RGB sensor and an illumination rig. Three databases from different sweet-pepper varieties were collected along different growing seasons. Target detectability highly depends on the imaging acquisition distance and the sensing system tilt. A minimum of 12 training scenes are necessary to discover the statistically significant spatial variables. Better prediction was achieved at the beginning of the season with slightly better prediction achieved in a temporal split of training and testing sets.