Effect of varying training epochs of a Faster Region-Based Convolutional Neural Network on the Accuracy of an Automatic Weed Classification Scheme

Abstract

Site-specific weed detection and management is a crucial approach for crop production management and herbicide contamination mitigation in precision agriculture. With the advent of unmanned aerial vehicles (UAVs) and advances in deep learning techniques, it has become possible to identify and classify weeds from crops at desired spatial and temporal resolution. In this research, a faster region based convolutional neural network was implemented for the automatic weed identification and classification using a mixed crop farmland as a case study. A DJI phantom 4 UAV was used to simultaneously collect about 254 image pairs of the study site. The images were annotated before transferring them into google colaboratory where they were trained over five epochs; 10,000, 20,000, 100,000, 200,000, and 242,000 with the aim of detecting the point when the model flattens out in the process of automatically identifying and classifying the weeds. The neural network identified and classified five classes which are; sugarcane, spinach, banana, pepper, and weeds. Finally, the accuracy of the automatic weed classification was evaluated with the aid of the recorded loss function and confusion matrix, and the result shows that the implemented model gave a classification accuracy of 52.5%, weed precision of 50%, weed recall of 7.7% and F1 score of 71.6% at 10,000 epochs, classification accuracy of 67.8%, weed precision of 67%, weed recall of 52.4% and a F1 score of 85.9% at 20,000 epochs, classification accuracy of 97.2%, weed precision of 96.2%, weed recall of 97.5% and a F1 score of 99% at 100,000 epochs, classification accuracy of 98.3%, weed precision of 98.1%, weed recall of 99.1% and a F1 score of 99.4% at 200,000 epochs, and classification accuracy of 97%, weed precision of 95%, weed recall of 99% and a F1 score of 99% at 242,000 epochs. It was observed that the model's performance improves significantly with increase in the number of epochs but got saturated at 242,000 epochs. The findings showed that the faster RCNN is robust for automatic weed identification and classification in a mixed crop farm.