GiRoS: A computer vision approach to estimate sunflowers performance

SPETALE FLAVIO EZEQUIEL; MURILLO JAVIER; CACCHIARELLI, PAOLO; GUSTAVO RODRIGUEZ; TAPIA ELIZABETH

Quilmes

Congreso; XI CAB2C; 2021

Background:Sunflowers (Helianthus annuus L.) are one of the three most important crops in Argentina, along with soybeans and corn. Correct identification of the sunflower yields contribute to the efficientcharacterization of the techniques and products utilization that can be a reference for future planting processes. Sunflower yield is performed considering the difference between the number of achenes and the number of seeds, not all achenes contain seeds inside. Traditionally, this task is manually performed over crop samples. Besides being extremely time consuming, the accuracy tends to be poor. In this work, we face the problem of identifying and counting the number of achenes through a computer vision method. A web interface, the source code, test image datasets and a batch script is freely available at www.cifasis-conicet.gov.ar/bioinformatica/Girasol/.Results:The final estimation of sunflower yield in our proposal consists of four steps: i) detection of sunflower contour; ii) identification and counting of achenes; iii) identification and counting of seeds and iv) results report. In this work, we focus on the first two steps considering three useful flower development stages in the sunflower (R7 and R8, both post-anthesis reproductive stages) where achene colors change from white, to brown and finally to black. To tackle contour identification problem, a mask is generated using edge detection techniques with different vegetation indices, i.e. metrics generated to distinguish plant parts by means of three channels (R, G, B), that are then applied to the original image. The second step performs an individual achene identification considering two vegetation indices, area and a series of decision rules based on the color of the achene. The dataset to test our method has 50 white, 30 brown and 70 black real images with different characteristics such as shape, size, shadows, bright taken from aprivate field in Santa Fe - Argentina during the 2020 harvest where each jpeg image has 1280x720pxresolucion.Conclusions:The general average accuracy over the dataset is 85%. In particular, the accuracy obtained for black,white and brown achenes was 85%, 89% and 81% respectively. We note that the intense bright andshadows from the border leaves in sunflower borders affect method precision. Without those effects, the proposed method performance may increase. Time reduces from hours to seconds for 10 images, moreover the result is objective, always providing the same result for the same photo, a situation which is not always true for two different persons performing the task.