Weeds Detection using Deep Learning Methods and Dataset Balancing

Abstract

Weeds have detrimental effects on agriculture and prove costly for farmers because they can quickly spread to fertile areas and reduce the fertility of the soil. Therefore, weed control is crucial for sustainable agriculture, and by detecting weeds and removing them from agricultural lands, we can transfer the limited resources we have to the plants to be grown, which would be a major step forward in sustainable agriculture. This article explores the feasibility of weed detection methods using deep learning architectures. Architectures used in the research are as follows: ResNet152V2, DenseNet121, MobileNetV2, EfficientNetB1 and EfficientNetB7. The F1-Score of EfficientNetB1 is 94.17\%, which is the highest score among those of all architectures. Among all architectures, EfficientNetB1 has the least number of parameters after MobileNetV2. In this research, data augmentation was done using horizontal flip, rotation, width shift, height shift, and zoom.