Weed Detection in Capsicum Annuum L. Crops Using Photogrammetry and Deep Learning Algorithms

Cuscuta spp. is a parasitic plant that causes losses estimated at 50 % of the yield of a wide variety of agricultural crops, including vegetables, forages, and trees. In the search for alternatives to solve this problem, small farmers are exploring the application of new technologies to food production. This paper presents a remote-sensing Cuscuta spp. classification model based on aerial images collected by UAVs to generate orthophotos showing infested areas. The proposed model segments the characteristic yellowish color of the stem of Cuscuta spp. in the HSV color space to train a deep Convolutional Neural Network (CNN). The experiments included RGB images of a Chili bell pepper (Capsicum annuum Linnaeus) crop with Cuscuta spp. In addition, we performed model 5-fold cross-validation with different datasets by using images collected during three consecutive weeks to identify the growth of the weed-affected area. According to the metrics used, ResNet architecture proved to be the best model for classifying Cuscuta spp. and non-Cuscuta. The proposed method allows smallholders to detect and locate the weed in the early stages of growth to facilitate removal and mitigation efforts.