Leaf area index (LAI) prediction using machine learning and UAV based vegetation indices

Abstract

As a critical indicator of plant growth and water use, accurately and promptly estimating leaf area index (LAI) is critical for improved crop management. However, measuring LAI requires substantial effort and time . The main objective of this study was to leverage vegetation indices (VIs) generated from unmanned aerial vehicle (UAV)-based images and machine learning (ML) techniques for estimating LAI of green beans and sweet corn. The research was conducted at the Tropical Research and Education Center (TREC), University of Florida, Homestead, Florida over three seasons from 2020-2023. The experiment for each crop consisted of four irrigation treatments, i.e., 100 % full irrigation (FI), 75 %, 50 %, and 25 % FI, with four replications. Destructive leaf samples were collected by cutting plants from 30 cm row length of two inner plot rows and leaf area (LA) was measured using a LI-3000C transparent belt conveyor. Plant height and canopy width were also measured bi-weekly. Moreover, a UAV-based RedEdge-MX sensor was employed throughout the seasons to collect high-resolution multispectral imageries that consist of five bands. Plant LAI was calculated using the plant density method from measured LA. A calibrated DSSAT model was used to simulate the LAI for both crops.. Simulated LAI from DSSAT was compared against measured LAI, and relationships were established between simulated LAI and 12 VIs generated from UAV images. Additionally, ML algorithms, i.e., random forest (RF), eXtreme gradient boosting (XGB), and light gradient boosting (LGB) models, were trained to predict LAI for both crops using VIs as input features. Results showed DSSAT's perfroamnce in simulating LAI was good for green beans and reasonable for sweet corn. Out of the 12 indices tested, six VIs, i.e., Enhanced Vegetation Index 2 (EVI2), Normalized Difference Vegetation Index (NDVI), Normalized Green Red Difference Index (NGRDI), NIR-RE Normalized Difference Vegetation Index (NIRRENDVI), Red-Edge Normalized Vegetation Index (RENDVI), and Soil Adjusted Vegetation Index (SAVI) showed a good agreement with simulated LAI for both crops. The LGB, RF, and XGB models predicted LAI with acceptable accuracy, achieving r² values of 0.78, 0.90, and 0.90 and RMSE values of 0.43, 0.29, and 0.28 for sweet corn and r² of 0.72, 0.79, and 0.80 and RMSE of 1.01, 0.86, and 0.85 for green beans, respectively. The findings indicate that ML models and VIs derived from UAV imagery could be used to predict LAI with acceptable accuracy for green beans and sweet corn.