Noninvasive system for weight estimation in cactus crops: A YOLOv5-decision tree approach based on interval type-2 fuzzy sets

This study proposes a noninvasive system to estimate the weight of a prickly pear cactus (Opuntia ficus-indica), which combines a model based on deep learning to detect and estimate its area and a model based on supervised learning and developed under a type-2 interval fuzzy sets approach to estimate its weight. YOLOv5 was used for detection, which achieved results with an accuracy of 0.999, recall of 0.999, mAP_0.5 of 0.995, and mAP_0.5:0.95 of 0.990. Area estimation is performed based on the object's coordinates in the image. To perform the weight estimation task, a comparative analysis of models based on supervised learning was performed to identify the model that best described the response variable, selecting a decision tree model. The training and adjustment process included the uncertainty in a weight evaluation system using a granatary scale to enhance the weight estimation capability. Using a trapezoidal membership function, the type-2 interval fuzzy set approach was used to convert the collected data into fuzzy numbers. The model findings for the weight estimation task were also favorable, with a model-explained variance R2 of 0.98, a root mean square error (RMSE) of 10.02, and a mean absolute error (MEA) of 6.33 g. Finally, both models were incorporated into a graphical user interface to facilitate their use in the real weight estimation process. A real case application with the proposed system demonstrated its effectiveness in detecting and estimating weight dynamically with an error of 9 g. Therefore, its viability as an alternative system for weight estimation in cactus crops was demonstrated. Furthermore, a comparative analysis demonstrated the superiority of the proposed hybrid approach over similar approaches in the literature.