Leaf area index estimations by deep learning models using RGB images and data fusion in maize

Abstract

Abstract The leaf area index (LAI) is a biophysical crop parameter of great interest for agronomists and plant breeders. Direct methods for measuring LAI are normally destructive, while indirect methods are either costly or require long pre- and post-processing times. In this study, a novel deep learning-based (DL) model was developed using RGB nadir-view images taken from a high-throughput plant phenotyping platform for LAI estimation of maize. The study took place in a commercial maize breeding trial during two consecutive growing seasons. Ground-truth LAI values were obtained non-destructively using an allometric relationship that was derived to calculate the leaf area of individual leaves from their main leaf dimensions (length and maximum width). Three convolutional neural network (CNN)-based DL model approaches were proposed using RGB images as input. One of the models tested is a classification model trained with a set of RGB images tagged with previously measured LAI values (classes). The second model provides LAI estimates from CNN-based linear regression and the third one uses a combination of RGB images and numerical data as input of the CNN-based model (multi-input model). The results obtained from the three approaches were compared against ground-truth data and LAI estimations from a classic indirect method based on nadir-view image analysis and gap fraction theory. All DL approaches outperformed the classic indirect method. The multi-input_model showed the least error and explained the highest proportion of the observed LAI variance. This work represents a major advance for LAI estimation in maize breeding plots as compared to previous methods, in terms of processing time and equipment costs.