101304 Early Season Corn Stand Quantification Using UAS and Computer Vision.

Poster Number 454-809

See more from this Division: ASA Section: Climatology and Modeling
See more from this Session: Agricultural Remote Sensing Poster

Wednesday, November 9, 2016
Phoenix Convention Center North, Exhibit Hall CDE

Sebastian Varela, Agronomy, Kansas State University, Manhattan, KS and Ignacio A. Ciampitti, Kansas State University, Manhattan, KS
Abstract:
Remote sensing offers a wide range of sensors in agriculture. From low resolution satellite to cm level UAS. Current UAS technology unleash ultra-high spatial resolution data to end users. From an agronomic site-specific management perspective; soil zoning, plant population and yield potential are linked. Operationally, in general the farmer implements uniform seeds/acre target to a field. From a site-specific management practice higher yield environment in a field require higher inputs to achieve higher potential yields. In the other direction, for low yield environments inputs can be optimized. An accurately and timely spatial quantification of plant population can lead to a more efficient performance of the crop. Spatial distribution of plants is the outcome of complex interactions of factors such as soil type, seeding depth, soil temperature, and water available between planting and emergence. We evaluate different analytics approaches for early plant population estimation. During 2016 season an experimental area located in Ashland Bottoms Experimental Station (KS, US) was utilized. The experiment includes 7 population ranges and 5 repetitions of Dekalb 61-88 hybrid.  We consider 3 early season dates and 2 UASs altitudes in order to test timing for goodness-of-fit of the analytic procedures applied to different population ranges. Data processing workflow was implemented using Photoscan and Matlab software. The workflow procedure includes: 1) feature space transformation of RGB imagery; 2) imagery classification: K-Means, SVM (Support Vector Machine), DT (Decision Tree) and ANN (Artificial Neural Network); 3) Hough-transformation method and filter procedures were applied to erode misclassified pixels as plant candidates and 4) performance evaluation of classification was implemented. Accuracy, misclassification rate, external supervision needed and computer performance were considered. Lower flight altitude, intermediate planting date and K-means classification presented the best model in terms of accuracy, external supervision needed and computer performance for plant count estimation. UASs ultra-high resolution imagery and machine learning allow new insights of early season corn performance.

See more from this Division: ASA Section: Climatology and Modeling
See more from this Session: Agricultural Remote Sensing Poster