Estimating Straw Cereal Plant Density at Early Stages Using Reflectance Based and Image Segmentation Based Methods Under Different Spatial Resolutions

Abstract

Estimating straw cereal plant density at early stages is important for field crop management and phenotyping. Usual plant density estimation methods include manual counting and image-based counting, both of which have limited throughput, e.g., due to the need for high spatial resolution images that must be acquired at low altitudes. In this study, we explored the potential of a high-throughput alternative based on the exploitation of spectral information instead of spatial information. A large and diverse dataset was collected on microplot field experiments, encompassing six sites, three leaf stages and four species of straw cereals. Canopy spectral reflectance was acquired with a spectrometer, both in nadir view or in 45° view zenith angle perpendicularly to the row direction. Two reflectance-based approaches were then tested to estimate plant density. In the direct approach, density was directly estimated from reflectance using gaussian process regression (GPR) and spectral band selection based on Akaike’s information criteria. In the indirect approach, green fraction derived from high spatial resolution RGB images (GF_rgb) was first estimated from reflectance using GPR and band selection, and then linearly related to density. These reflectance-based methods were compared to a classical image-based, baseline method, which estimates density directly from GF_rgb.An ablation study first showed the superiority of 45° observations, and the necessity to calibrate one model for each site, growth stage and species. Next, the band selection process recommended using no more than four bands as inputs to the GPR models, so that the models have a good balance between accuracy and parsimony. The resulting direct and indirect estimations had similar overall relative errors of 30% (RMSE = 77 plants/m2), reaching a minimum of 27% (RMSE = 71 plants/m2) when focusing on three-leaf stage. On the other hand, the image-based, baseline method had a lower overall relative error of 22% (RMSE = 54 plants/m2) for submillimetric spatial resolutions. However, this error increased when degrading the image spatial resolution to mimic an increase in sensor altitude, exceeding that of reflectance-based methods when the ground sampling distance was larger than a threshold size between 1 to 2 mm depending on growth stage. These results show that higher spectral information can compensate for lower spatial information, and that spectral methods can provide high-throughput and reasonably accurate estimates of straw cereal plant density.

Keywords: Plant density, spectral reflectance, spatial resolution, wheat, barley