Automatic wheat ear counting in-field conditions: simulation and implication of lower resolution images

Abstract

The number of ears per unit ground area, or ear density, is in most cases the main agronomic yield component of wheat. A fast evaluation of this attribute may contribute to crop monitoring and improve the efficiency of crop management practices as well as breeding programs. Currently, the number of ears is counted manually, which is time consuming. This work uses zenithal RGB images taken from above the crop canopy in natural light and field conditions. Wheat trials were carried out in two sites (Aranjuez and Valladolid, Spain) during the 2014/2015 crop season. A set of 24 varieties of durum wheat in two growing conditions with three dates of measurement were used to create the image database. The algorithm for ear counting uses three steps: (i) Laplacian frequency filter (ii) median filter (iii) Find Maxima. Although the image database was collected at the ground level, we have simulated images at lower resolutions in order to test potential application from cameras with lower resolution, such mobiles phones, action cameras (5 – 12 megapixels), or even aerial platforms (e.g. UAV from 25-50 meters). Images were resized to five different resolutions with no interpolation techniques applied. The results demonstrate high accuracy between the algorithm counts and the manual (image-based) ear counts, higher than 90% in success rate, with a decrease of <1% when images were reduced to a half of its original size, and success rates decreasing by 2.29%, 7.32%, 17.32% and 38.82% for images resized by four, eight, 16 and 32 values, respectively.