Vaishali Swaminathan, Nithya Rajan, J. Alex Thmoasson, Karem Meza Capcha, Amrit Shrestha, Tianyi Wang, Jeffrey Siegfried
Abstract
In places where plant emergence rate and distribution are not uniform, early season aerial images can help track the emergence of individual plants in the field. The structure of the cotton plant makes it easier to detect and locate them within the field earlier than later. In this study, four sequentially captured UAV-based aerial images, obtained within a span of one-and-half months during the early cotton growth season, were used to locate plants that had spatiotemporal variabilities in emergence. An algorithm was developed by combination of image processing and geographic information system (GIS) tools to extract plant counts and locations from the images. The two image transformations tested for detecting newly emerging plants were hue-saturation-value (HSV) and excess green-minus-red (EXGR), which were extracted from the red (R), green (G), and blue (B) bands. The ExGR is a 1-dimensional color space that is defined by the equation G – B – 2.3 R. Since the HSV transformation resulted in a 3-dimensional color space, it was transformed into a 1-dimensional color space using a custom transformation (indexed HSV= 3.387 H – V). Appropriate thresholds were assigned for each date based on the histogram in order to separate the plants from soil, and the locations of newly detected plants were appended. Further, row lines were detected using morphological pruning to eliminate weed and noise in the background. While the results showed similar performance between HSV (R2 = 0.96) and ExGR (R2 = 0.91), the HSV transformation certainly performed better in discerning plant pixels from soil background when subjected to varying light conditions. The precision (true positive/total detected) and recall (true positive/actual count) values were approximately 0.92 and 0.86 respectively. The detected plant locations and counts were used to determine the height and yield per plant in plots that had non-uniform distribution.
Comments