Multispectral Images for Differentiation of Lettuce Plants Inoculated With Phosphate-Solubilizing Fungi Under Different Phosphorus Doses

Advanced technologies, such as remote sensing, have become prominent tools for monitoring various agricultural crops. However, few studies have investigated the potential of low-cost multispectral sensors for discriminating treatments involving bioinput applications, particularly in lettuce (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Lactuca sativa</i>) cultivation. This study aimed to evaluate the potential of multispectral images acquired by low-cost sensors, combined with machine learning techniques, for discriminating treatments applied to lettuce plants, considering different inoculation methods of the fungus <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Aspergillus niger</i> and variations in phosphorus fertilization doses. The experiment was conducted in 2021 under a 6 × 3 factorial design, including six inoculation methods of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A. niger</i> (no inoculation; seed treatment with conidial suspension; granular inoculation with 4 × 10² and 4 × 10⁶ conidia g⁻¹ at transplanting; and inoculation with conidial suspensions at 4 × 10² and 4 × 10⁶ conidia mL⁻¹ at transplanting) and three phosphorus doses (0, 50, and 100%). During plant development, images were captured with an RGB camera and a MAPIR Survey 3W multispectral camera on days 5, 12, 19, 26, and 33 after transplanting. For treatment discrimination, average reflectance values of spectral bands were extracted, and vegetation indices were computed. Classification was performed using machine learning algorithms in the Weka software, considering three distinct scenarios. The Random Forest model showed the best performance, particularly when combining RGB and RGN sensor attributes, achieving overall accuracies of 65.35%, 79.57%, and 94.24% in the respective scenarios. This study reinforces the potential of spectral imaging for crop monitoring and management.