Efficient Classification Of Weeds In Crops For Autonomous Spray Application

Abstract

eeds have an impact on crop health because they share water and minerals from the soil, hence, lowering crop production. Human health is endangered by manual weedicide spraying. Localized weedicide spraying by aerial spraying units can help save water, weedicide chemicals, and have a lower impact on human health. To categorize crop, weed and soil surface, multi-spectral cues are required. In order to undertake autonomous tasks in agricultural settings, efficient crop, weed, and soil background classification is required. Pixel-wise segmentation approaches are outperformed by semantic segmentation as seen in much research. Improved background classification is witnessed in the crop-weed-background three-class classification situation, however crop and weed classification is low in literature when utilizing semantic segmentation. This observation is theoretically justifiable because the background is very different from the other two classes, and crop and weed are more similar. This study addresses this issue by suggesting the W network, which consists of two encoder-decoder architectures of varying sizes. The first encoder-decoder structure distinguishes between background and vegetation, removing background pixels. The second encoder-decoder structure retrains on crop, weed, and background, with this retraining the model learns improved features to efficiently categorize crop and weed classes. Three significant dataset contributions are made in this research to aid further research in this field. A dataset of Sesamum indicum collected with a Phantom 3 drone, a Nicotiana tabacum dataset was collected with a Mavic Mini drone and a Beta Vulgaris dataset was collected with a Phantom 4 drone. To demonstrate the efficiency of our proposed W network, we trained it from scratch using our captured tobacco dataset. When compared to state-of-the-art three-class segmentation models, it performed better on test data. We fine-tuned our W network using a sesame crop dataset to further validate our proposed model and demonstrated its generalizability, it demonstrated consistency with good results. Any other agricultural crop-weed scenario could easily be finetuned with our proposed W network. In addition to the above this study also proposes an approach for classifying patches of images depending on dataset grouping and model ensembling, as well as smart convolutional neural networks (CNNs) that can identify patch images more quickly. The proposed system was tested in a variety of lighting settings, as well as circumstances of both dry and wet soil conditions and at various growth phases. Test experiments have shown up to 10 percent increase in mean intersection over union (MIOU) in pixel-wise results using our proposed W network as compared to UNet and SegNet, maximum MIOU of 0.92 and 0.90 are achieved using tobacco and sesame fields datasets respectively, qualitative pixel-wise results also confirm this visually. Similar trend is seen in patch-wise results using our proposed patch-wise method where maximum accuracy of 92% is observed using RGB and CIR combination of images as input. Sensitivity analysis of different configuration of patches generation have shown that our proposed patch-based system is stable with different configurations of patches.