Classification of Tree Genotype using Spectro-temporal Variance in Biophysical Traits Estimated from UAV-based Hyperspectral Time Series Data

Abstract:

The increasing frequency of extreme climatic events has appalling consequences on the tree health and population in global forests. In this context, increasing the biotic and the abiotic stress resistance of trees through selective breeding is a genomic solution to conserve future forests. A major bottleneck in the advancement of large-scale breeding programs is the lack of high-throughput phenotyping methods that can accurately map biophysical traits of individual trees to its genotype. Individual tree genotypes have unique genetic attributes that manifest as phonological difference in biophysical parameters such as the chlorophyll concentration, the leaf water content, and the biomass accumulation rate. Optical remote sensing techniques use vegetation indices as proxies to estimate biophysical parameter, and is hence is a cost-efficient alternative to field-based biophysical parameter estimation. In particular, the recent advancements in the payload capacity of Unmanned Aerial Vehicles (UAV) together with the availability of light-weight hyperspectral cameras allows quick remote data collection of forests at unprecedentedly-high spectral and temporal resolutions. Thus, we propose a method that jointly use the spectral and the temporal information in UAV-based hyperspectral time-series data to classify tree genotypes. We plan to achieve high-throughput tree genotype classification by: a) delineating individual crowns using the Marker-controlled watershed algorithm on the first principal component of the hyperspectral data, derived using the Principal Component Analysis, b) generating a tree-level feature vector comprising of annual variance of the average Normalized Difference Vegetation Index (NDVI), the chlorophyll/carotenoid index (CCI) and the photochemical reflectance index (PRI), in the crown area, c) performing hierarchical K-means clustering in the feature space. We plan to test the performance of the proposed high-throughput phenotyping framework on a mature (20 year old) white spruce plantation with 2000 different genotypes, located in St. Casimir, Quebec. Data acquisition campaigns were conducted three time each in the growing season (between April to October) of 2016 and 2017. Preliminary classification results obtained on a subset of 10 genotypes is promising.