Early almond yield forecasting by bloom mapping using aerial imagery and deep learning

Abstract

Our study aims to measure bloom density in almond trees and, in turn, predict crop load at the tree level from aerial RGB images. Almond yield forecasting early in the season has become necessary in California due to the new legislation that requires growers to determine their nitrogen fertilizer rate based on soil tests and crop yield forecasts. Predicting almond yield by estimating crop load from bloom density would be the earliest in-season yield estimation. It would help growers with precision management, saving inputs, and complying with the nitrogen limit mandate. Drone flights were conducted during the almond bloom period for six orchards. An end-to-end deep learning architecture, U-Net, was used to segment and determine bloom density per tree. Our model was able to achieve high levels of accuracy, with precision and recall values of 81% and 65%, respectively, which means that the model was able to identify a significant proportion of the bloom pixels with high precision. We conducted a simple linear regression analysis to compare bloom density determined by manual hue-saturation-value (HSV) thresholding (as labels) to bloom density predicted by the U-Net, which resulted in a coefficient of determination (R2) of 0.78 and a normalized root mean square error (NRMSE) of<1 percent (0.69%). The results confirmed that the model is robust and independent from lighting and environmental conditions. End-of-season yield data at the row level were collected from an experimental almond orchard for three consecutive years and used as ground truth for verifying the association between the bloom density and crop load. Our tests revealed a significant positive Pearson’s correlation coefficient (R = 0.65–0.9, p < 0.05) when comparing crop load and the estimated bloom density for Nonpareil cultivar every year, which growers could use to predict expected in-field yield variability in regard to bloom density.