Deep Learning based flower detection and counting in highly populated images: A peach grove case study

Abstract

Farmers and producers need to estimate crop yield in order to plan and allocate human and economic resources during the harvesting season. For many crops, such as peach groves, the number of fruits is correlated with the number of flowers produced by each tree. Therefore, estimating the number of flowers in peach groves can serve as a good indicator of crop yield, disregarding climate hazards. However, in peach groves, tree images present several challenges, including a high number of flowers, interference from distant trees, and occlusion between elements. These issues pose a difficult task for computer vision and machine learning techniques. In this study, we propose the utilization of state-of-the-art deep learning techniques for image detection purposes; namely the YOLO architectures on its versions 5, 7, and 8 and their different size models (n, s, m, l, x); as well as predicting object density using multi-column in densely populated images, using a multi-column deep neural network. The methodology was tested on a new dataset comprising 600 images of peach trees during the blooming season, in the region of Catalonia, Spain. Out of these, 400 images were used to train the model, while 100 were allocated for testing and another 100 for validation. The counting results were evaluated using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and percentage error (%Err). For the detection algorithms, metrics such as accuracy, precision, recall, and mean average precision were utilized, alongside metrics for evaluating the counting process.The experiments demonstrated that predicting the density map yielded better results in the counting process, achieving an MAE of 39.13, RMSE of 69.69, and a percentage error of 9.98. The detection algorithm that exhibited superior performance was YOLOv7x, with metrics of MAE 152.7, RMSE 212.9, and a percentage error of 29.7 %. These results indicate that, for counting purposes, predicting the density map produced better overall outcomes.