Improved Canopy Estimation for Greenhouse Spray Applications

Research Poster
Uchit Nair
Category: 
MS
Advisor: 
Peter Ling
Department: 
Department of Food, Agricultural and Biological Engineering
Abstract: 

An intelligent sprayer can improve the efficiency of spray application in plant production and has been demonstrated for field applications. The same principle can be used inside greenhouses (Yan et al., 2018) where a LiDAR sensor is used to generate a 3D map of plant canopies and calculate canopy volumes in order to identify the rate of spray required. This system though has limited accuracy in measuring canopy volumes due to the spatial constraints of a greenhouse. It suffers from problems of occlusion and distortion at greater distances from the laser. Hence, this project proposes to overcome these limitations by introducing a processing algorithm that manipulates the noisy dataset to produce better estimations of the canopy volumes. The algorithm works by first registering or combining multiple scans of the same scene to get as much information as possible and then identifies individual plants by using a clustering algorithm. In order to overcome the problems of distortion and occlusion, each cluster/plant canopy is processed further. Assuming plant canopies are symmetrical, the well-defined side of the canopy is mirrored onto the distorted/occluded side to get a better idea of the overall dimensions and volume. The performance of the processing algorithm was evaluated by calculating individual volumes of objects from the data collected by the LiDAR sensor. The sensor was placed at five different heights (0.25, 0.5, 0.75, 1.0, 1.25m) and scanned 4 rows of regularly shaped objects (boxes, cylinders, basketballs and toy balls) and 1 row of artificial plants, each spaced 0.5m apart within rows and between rows. There was a general trend of increasing accuracy with an increasing sensor height. The lowest mean accuracy achieved was 41.5% for the sensor placed at 0.25m from the top of the objects and the maximum mean accuracy achieved was 79.9% for the LiDAR sensor placed at 1.25m from the top of the objects. The sensor height mitigates the problems of occlusion to a certain degree and hence explains its relationship with the accuracy. The processing algorithm introduced has the potential to not only improve spray efficiencies but also in plant phenotyping and growth monitoring applications.