Classification of weed species in the paddy field with DCNN-Learned features

Abstract

Weed is one of the main factors diminishing the yield of rice. A common approach toward weed control is the broad spraying of chemical herbicides, but the abuse of which can cause the damage to our environment and ecosystem. Fortunately, site-specific spraying, based on automatic weed identification systems, can substantially reduce the herbicide dosage. This study investigates a new automatic weed identification method that learns features from original images using Deep Convolutional Neural Networks (DCNN). The DCNN-learned features can avoid the complicated designing process designing and lengthy tuning procedures required for obtaining traditional hand-crafted (HC) features. The classification results of six weed species were compared between DCNN-learned features and HC features, using Support Vector Machine (SVM), Ensemble, and K-Nearest Neighbor (KNN) classifies. The classification accuracy of HC features using SVM classifier was 87.8% and that using DCNN-learned features was 94.5%. Hence, our DCNN-learned features demonstrated better classification accuracy compared to HC features. Furthermore, the classification results of weed images with complex background (CB) and background removed (BR) were compared to verify the sensitivity of DCNN-learned features with respect to changes in background and illumination. The average classification accuracy of DCNN features in the 6th and 7th fully connected layer, F6, F7, using the three classifier didn't make much difference with CB images and BR images, which shows that the segmentation of weed image background had a small effect on the accuracy of classification using DCNN-learned features. The results demonstrated that DCNN-learned features can improve the classification accuracy of weeds with the complex background and variable illumination.