Deep Learning based Weed Detection and Classification for Smart Farming

Weeds are unwanted plants that might compete with crops and have a negative effect on agricultural productivity. Conventional weed control methods relied heavily on the indiscriminate spraying of herbicides across entire fields. Herbicide use was not only economically burdensome but also contributed to environmental degradation. The widespread application of herbicides had adverse effects on soil health, water quality, and the overall ecosystem. In this work, we present the classification of weeds using various Convolutional Neural Network (CNN) models trained using two different datasets, namely DeepWeeds and Plant Seedlings. The investigations evaluate the performance of twelve deep learning models, including ShuffleNet, VGG-16, VGG-19, AlexNet, ResNet-101, ResNet-50, ResNet-18, Inceptionv3, Inception-ResNet-v2, NASNet-Large, NASNet-Mobile, and EfficientNet-b0. The evaluation is conducted on both balanced and unbalanced datasets, with the balanced dataset yielding more accurate results. Inception-ResNet-v2 is found to outperform other models, although it requires more training time.