Leveraging Time-Based Spectral Data from Uav Imagery for Enhanced Detection of Broomrape in Sunflower

Abstract

Sunflower broomrape (Orobanche cumana) poses a severe threat to sunflower crops, parasitizing their roots and hindering plant growth. Current control methods, which typically rely on uniform herbicide applications, are economically inefficient and environmentally damaging. This study investigates the use of unmanned aerial vehicle (UAV)-based multispectral imaging to detect broomrape-infected sunflowers by analyzing temporal patterns in spectral vegetation indices (VIs). Over four imaging campaigns conducted during early subsoil parasitic stages, multispectral data were collected and processed to compute ten VIs. These VIs, reflecting changes in canopy reflectance over time, were then analyzed using various machine learning models, including a pattern recognition neural network (PRNN). Results showed that the PRNN model, trained on time-series data, achieved an overall accuracy of 84.8% and a true positive rate of 80.4% in detecting broomrape infection, emphasizing the strength of utilizing temporal data for enhancing detection accuracies. Pixel-level reconstruction maps revealed varying spectral responses within infected canopies, highlighting the importance of accounting for this heterogeneity. This study demonstrates the potential of UAV-based multispectral imaging combined with advanced machine learning (ML) techniques for early detection of broomrape infestations in sunflower crops, offering insights for managing similar infestations in other crops.

Keywords: Broomrape, Classification, Multispectral, site-specific weed management, pattern recognition neural network, time-based data