Convolutional neural networks for accurate estimation of canopy cover

Canopy Cover (CC) is a key variable in agriculture, providing critical information on crop growth and health. RGB cameras can estimate CC, but conventional methods depend on vegetation characteristics, with color thresholds needing manual adjustments based on visual conditions. This manual intervention poses a challenge for integrating CC into automated cropping systems without human involvement. In this study, we present a breakthrough in solving this challenge by developing a Convolutional Neural Network (CNN) designed to compute CC autonomously. Deploying an RGB camera across various crops (wheat, alfalfa, and sweet pepper) in Spain and California, we harnessed a dataset comprising 283,000 images, each with dimensions of 256 × 256 pixels, for network training. Optimizing the neural network was achieved through a Genetic Algorithm (GA), employing an objective function focused on minimizing parameter density while maintaining accuracy. This approach ensures adaptability across various devices, making it especially suitable for low-power edge-computing applications in agriculture. Remarkably, the results indicate that our CNN, with fewer than 200,000 parameters, achieved an R2 of 98 %, an MSE of 0.0024, and an MAE of 0.038. This work demonstrates the feasibility of utilizing a low-density CNN for automatic CC calculation using RGB images. This approach offers significant potential for integrating CC into automated cropping systems, ultimately enhancing scalability and cost-efficiency in precision agriculture.