Advancements in precision agriculture have driven the development of spray drones for herbicide application, offering the potential to address challenges associated with current application methods and improve weed management. This review synthesizes current research on spray drones to develop broad-use recommendations and identify challenges and knowledge gaps. Although spray drones use lower carrier volumes than ground-based sprayers (high-volume backpack or tractor-mounted sprayers), studies report comparable or superior weed control as well as herbicide cost savings. However, spray drone performance is highly sensitive to operational parameters, as spray distribution and coverage/deposition are strongly affected by flight height and speed, carrier volume, nozzle design, crop growth stage, weed, and weather conditions. The bell-shaped curve of a single-pass spray pattern, which results in most spray deposition occurring directly under the unmanned aerial vehicle (UAV), coupled with advances in imaging, remote sensing, and machine learning, demonstrate the strong potential of spray drones for site-specific weed management. Vegetation indices, multispectral imagery, canopy height models, and Light Detection And Ranging (LiDAR) technology have enabled crop-weed discrimination, though accuracy varies with species, growth stage, and image resolution. Deep-learning models such as ‘You Only Look Once’ (YOLO), Residual Neural Network (ResNet) and Mask Region-based Convolutional Neural Network (Mask R-CNN) achieve high performance for weed detection and/or segmentation but remain limited by training data quality and reduced accuracy with small, overlapping, or dense weed populations. Spray drone-based offline mapping has enabled substantial herbicide savings by delineating weed patches, whereas real-time weed detection is constrained by onboard processing limits, battery life, and lower spatial resolution at operational flight heights. Ground-based smart sprayers offer higher real-time detection precision but lack the field accessibility advantages of spray drones. Despite their potential, spray drones face challenges, including limited payload, off-target movement of pesticides, short battery life, regulatory challenges, and extensive license and complex software and calibration requirements. The downwind spray drift potential of spray drones is greater than ground applications but smaller than manned aerial applications. An upwind swath offset is an ideal best management practice to reduce off-target pesticide movement to susceptible areas from both manned and spray drone equipment. Future research should evaluate spray drones within integrated weed management systems, focusing on preemergence and foliar-applied contact herbicides, adjuvant use, environmental and operational interactions to develop spray drone-specific guidelines and optimize spray performance.

Precision weed detection and mapping in vegetable crops are beneficial for improving the effectiveness of weed control. This study proposes a novel method for indirect weed detection and mapping using a detection network based on the You-Only-Look-Once-v8 (YOLOv8) architecture. This approach detects weeds by first identifying vegetables and then segmenting weeds from the background using image processing techniques. Subsequently, weed mapping was established and innovative path planning algorithms were implemented to optimize actuator trajectories along the shortest possible path. Experimental results demonstrated significant improvements in both precision and computational efficiency compared with the original YOLOv8 network. The mean average precision at 0.5 (mAP50) increased by 0.2, while the number of parameters, giga floating-point operations per second (GFLOPS), and model size decreased by 0.57 million, 1.8 GFLOPS, and 1.1 MB, respectively, highlighting enhanced accuracy and reduced computational costs. Among the analyzed path planning algorithms, including Christofides, Dijkstra, and dynamic programming (DP), the Dijkstra algorithm was the most efficient, producing the shortest path for guiding the weeding system. This method enhances the robustness and adaptability of weed detection by eliminating the need to detect diverse weed species. By integrating precision weed mapping and efficient path planning, mechanical actuators can target weed-infested areas with optimal precision. This approach offers a scalable solution that can be adapted to other precision weeding applications.

Site-specific weed management (SSWM) provides precise weed control and reduces the use of herbicides, which not only reduces the risk of environmental damage but also improves agricultural productivity. Accurate and efficient weed detection is the foundation for SSWM. However, complex field environments and small-target weeds in fields pose challenges for their detection. To address the above limitations, we developed WeedDETR, a real-time end-to-end detection model specifically designed to enhance the detection of small-target weeds in unmanned aerial vehicle (UAV) imagery. WeedDETR incorporates RepCBNet, a backbone network optimized through structural re-parameterization, to improve fine-grained feature extraction and accelerate inference. In addition, the designed feature complement fusion module (FCFM) was used for multi-scale feature fusion to alleviate the problem of small-target weed information being ignored in the deep network. During training, varifocal loss was used to focus on high-quality weed samples. We experimented on a new dataset, GZWeed, which contains weed imagery captured by a UAV. The experimental results demonstrated that WeedDETR achieves 73.9% and 91.8% AP0.5 (average precision at 0.5 intersection over union threshold) in the weed and Chinese cabbage [Brassica rapa subsp. chinensis (L.) Hanelt] categories, respectively, while achieving an inference speed of 76.28 frames per second (FPS). In comparison to YOLOv5-L, YOLOv6-M, and YOLOv8-L, WeedDETR demonstrated superior accuracy and speed, exhibiting 3.5%, 6.3%, and 3.6% higher AP0.5 for weed categories, while FPS was 14.9%, 12.9%, and 1.4% higher, respectively. The innovative architectural design of WeedDETR significantly enhances the detection accuracy of small-target weeds, enabling efficient end-to-end weed detection. The proposed method establishes a solid technological foundation for UAV-based precision weeding systems in field conditions, advancing the development of deep learning–driven intelligent weed management.

Automatic detection and removal of weeds is a challenging task that requires precise sensors. While crops and weeds possess similar features in terms of appearance, they can be discriminated based on spectral information. This can be done because any object has its own specific spectral signature based on its physical structure and chemical contents. This study examined the use of wavelet transform and deep learning for discrimination of weeds from crops. A total of 626 spectral reflectances in the range of 380 to 1,000 nm were obtained for three crops (cucumber [Cucumis sativus L.], tomato [Solanum lycopersicum L.], and bell pepper [Capsicum annuum L.]) and five different weeds (bindweed [Convolvulus spp.], purple nutsedge [Cyperus rotundus L.], narrowleaf plantain [Plantago lanceolata L.], common cinquefoil [Potentilla simplex Michx.], and garden sorrel [Rumex acetosa L.]). Morse wavelet was employed to decompose the spectra and extract the scalograms, which are the RGB representations of the spectral data. Two deep convolutional neural networks (i.e., GoogLeNet and SqueezNet) were employed for the recognition of crops and weeds. In addition, six common classifiers, including linear discriminant analysis, quadratic discriminant analysis, linear support vector machine, quadratic support vector machine, artificial neural networks, and k-nearest neighbors classifier (KNN), were used for the task of crop/weed discrimination to build the comparison with the proposed method. The error of prediction gradually decreased, and a 100% correct classification was achieved after 258 iterations. Analysis showed that SqueezNet provided classification of 100% accuracy, while GoogLeNet’s accuracy was 97.8% for the test set. Among the common classifiers, KNN provided the highest accuracy (i.e., 100%). This study showed that the proposed method can be successfully utilized for classification of crops and weeds.

The utilization of remote sensing in agriculture has great potential to change the methods of field scouting for weeds. Previous remote sensing research has been focused on the ability to detect and differentiate between species. However, these studies have not addressed weed density variability throughout a field. Furthermore, the impact of changing phenology of crops and weeds within and between growing seasons has not been investigated. To address these research gaps, field studies were conducted in 2016 and 2017 at the Horticultural Crops Research Station near Clinton, NC. Two problematic weed species, Palmer amaranth (Amaranthus palmeri S. Watson) and large crabgrass [Digitaria sanguinalis (L.) Scop.], were planted at four densities in soybean [Glycine max (L.) Merr.]. Additionally, these weed densities were grown in the presence and absence of the crop to determine the influence of crop presence on the detection and discrimination of weed species and density. Hyperspectral data were collected over various phenological time points in each year. Differentiation between plant species and weed density was not consistent across cropping systems, phenology, or season. Weed species were distinguishable across more spectra when no soybean was present. In 2016, weed species were not distinguishable, while in 2017, differentiation occurred at 4 wk after planting (WAP) and 15 WAP when weeds were present with soybean. When soybean was not present, differentiation occurred only at 5 WAP in 2016 and at 3 WAP through 15 WAP in 2017. Differentiation between weed densities did occur in both years with and without soybean present, but weed density could be differentiated across more spectra when soybean was not present. This study demonstrates that weed and crop reflectance is dynamic throughout the season and that spectral reflectance can be affected by weed species and density.

Many upland pastures and forest meadows in the western United States contain significant infestations of yellow hawkweed and oxeye daisy. Documentation of infestations is necessary in order to plan and assess control tactics. Previous work with an airborne charge coupled device (CCD) with spectral filters indicated that flowering yellow hawkweed with at least 30% cover was detectable at 1 m resolution. A single image of a large area may not capture all plants in the flowering phase and multiple images are costly. The objective of this paper was to assess the accuracy of images recorded at different phenological stages. We compared three methods of classification: unsupervised classification of a three principal component analysis image, supervised classification of a three principal component analysis image, and supervised classification of a composited image consisting of four bands and normalized difference near infrared (NIR)/red band. Regardless of the classification method, images of yellow hawkweed and oxeye daisy in full bloom had lower classification error than at early bloom or post bloom. The percent error for yellow hawkweed classification was about twice as high at post bloom as at full bloom, but varied slightly depending on the method of classification and cover class. The ability to detect discrete colonies of yellow hawkweed was not affected by phenological stage, but the ability to measure the area of each cluster differed among stages. Less than one-third fo the pixels classified as yellow hawkweed or oxeye daisy in the early bloom image remained in the same class in the full bloom image. About half the pixels in the full bloom image remained in the 90 to 100% cover class at the post bloom image. Seasonal growth of the grasses masked some yellow hawkweed and oxeye daisy plants, and accounted for differences in classification among phenological stages.

Yellow hawkweed infests permanent upland pastures and forest meadows in northern Idaho. Conventional surveys to determine infestations of this weed are not practical. A charge coupled device with spectral filters mounted in an airplane was used to obtain digital images (1 m resolution) of flowering yellow hawkweed. Supervised classification of the digital images predicted more area infested by yellow hawkweed than did unsupervised classification. Where yellow hawkweed was the dominant ground cover species, infestations were detectable with high accuracy from digital images. Moderate yellow hawkweed infestation detection was unreliable, and areas having less than 20% yellow hawkweed cover were not detected.

New or recently introduced weeds are biological pollutants in our natural and agricultural ecosystems. Unlike chemical pollutants, new weeds left unchecked often proliferate and pose problems that may not become apparent until eradication is too expensive or impractical. Management strategies for weeds should include: 1. prevention (from entering foreign commerce); 2. exclusion (detection of weed contaminants in imported products at ports of entry); 3. detection, containment, and eradication of incipient infestations; and 4. perpetual control (of widespread species that cannot otherwise be addressed). Appropriate legislative authority, modern weed technology, funding, and a renewed commitment to the concept of prevention are needed to prevent the introduction of new weeds. A national initiative to prevent the establishment of new weeds would be beneficial by saving on future losses and perpetual control costs. Actions taken now will prevent the continued introduction and spread of new weeds in the United States.

Ground-based surveys to regularly document the size, shape, and location of weed populations usually are not economically feasible. Digital images of vegetation in semiarid rangeland were obtained from four charge-coupled devices with spectral filters mounted in an airplane. The ability to distinguish yellow starthistle and common St. Johnswort from other rangeland vegetation in images with 0.5, 1, 2, and 4 m spatial resolution was assessed. Detection, positioning, and measurements of size and density of yellow starthistle and common St. Johnswort colonies with densities as low as 30% ground cover were possible at all those resolutions. Images indicated yellow starthistle occupied about one-third of a 180-ha study area. Images taken in June indicated 2 to 10 ha less yellow starthistle ranging from 30 to 100% cover than images taken in July for all resolutions. Images indicated common St. Johnswort occupied less than 4 ha of the 180-ha study area. This procedure provides a method to establish baseline plant community composition and a way to monitor species population changes and dispersal over time.

This paper presents a method for automating weed detection in colour images despite heavy leaf occlusion. A fully convolutional neural network is used to detect the weeds. The network is trained and validated on a total of more than 17,000 annotations of weeds in images from winter wheat fields, which have been collected using a camera mounted on an all-terrain vehicle. Hereby, the network is able to automatically detect single weed instances in cereal fields despite heavy leaf occlusion.

The goal of the project was to supply growers with knowledge on how incorporation of machine vision technology can affect the wild blueberry crop, disease pressures, and the overall savings of select agrochemical inputs. A machine vision system was developed and mounted on a rear sprayer boom in front of the sprayer nozzles capable of targeting the agrochemical application on an as-needed basis. Results showed that plants that received the proper fungicide application were less prone to premature leaf drop resulting in larger stem diameters and higher bud counts and harvestable fruit yield. Fungicide application savings using the smart sprayer for spot-application was 12% as compared to a uniform application.

Failure to detect noxious weeds with current survey methods prevents their control and has contributed to their ability to establish and spread in remote range and forest sites. Techniques used in remote sensing can classify plant occurrence on maps, offering a method for surveying invasive species in remote locations and across extensive areas. An imaging hyperspectral spectrometer recorded images on July 19, 1998 in Farragut State Park near Bayview, ID, in the reflected solar region of the electromagnetic spectrum ranging from 440 to 2,543 nm to detect spotted knapweed. The sensor records 128 spectral bands in 12- to 16-nm intervals at a spatial resolution of 5 m. A spectral angle mapper (SAM) algorithm was used to classify the data. Infestations in Idaho with 70 to 100% spotted knapweed cover that were 0.1 ha were detected regardless of the classification angle. However, narrow angles (2 to 8°) did not completely define the extent of the infestation, and the widest angle tested (20°) falsely classified some areas as infested. The overall image error for all classes was lowest (3%) when SAM angles ranged from 10 to 11°. Specific errors for the spotted knapweed class for the 10 to 11° angles showed that omissional and commissional errors were less than 3%. Areas with as little as 1 to 40% spotted knapweed cover were detected with an omissional error of 1% and a commissional error of 6%. Further verification sites were established on August 11, 1998 near Bozeman, MT, using the algorithms developed for Idaho. The omissional error for the Montana sites was 0%, and the commissional error was 10%. The hyperspectral sensor, Probe 1, proved an effective detection tool with the ability to detect spotted knapweed infestations.

Digital imagery from satellites and airborne remote sensing offer an opportunity to accurately detect weed infestations. Image resolution and plant growth stage are critical factors for maximum weed detection with low errors. Data analysis in traditional image assessment has relied on agreement measures, such as Cohen's kappa and asymptotic procedures, that compare what is on the image but not on the ground and what is on the ground but not on the image. Statistical comparisons of multispectral images, however, require some knowledge of the variability of the image classification results to determine significant differences among agreement measures. Bayesian methods were used to develop probability distributions for an agreement measure, conditional kappa, and were then subsequently applied to assess and compare image resolutions and plant growth stages. Results showed that images of a study site known to have yellow starthistle populations could identify the noninfested areas with greater accuracy than infested areas at spatial resolutions of 0.5, 1.0, 2.0, and 4.0 m. The detection accuracy of yellow starthistle in the images taken either prebloom or at flowering with 4.0-m spatial resolution usually was equal to or better than spatial resolutions of 0.5, 1.0, and 2.0 m for the cover classes that were not, moderately (31 to 70%), and highly (71 to 100%) infested. The 0.5-m resolution was better than 4.0-m spatial resolution when detecting the moderate cover class, but both resolutions had high omissional and commissional errors. Contrasting the best detection resolution for finding yellow starthistle colonies across flight times indicated that flying at flowering stage with the 4.0-m spatial resolution provided the best detection of the yellow starthistle cover classes considered. In the cases where different spatial resolutions resulted in equal detection accuracy, the larger spatial resolution was selected because of lower costs of acquiring and processing the data.

Cutleaf teasel is an invasive, biennial plant that poses a significant threat to native species along roadsides in Missouri. Flowering plants, together with understory rosettes, often grow in dense patches. Detection of cutleaf teasel patches and accurate assessment of the infested area can enable targeted management along highways. Few studies have been conducted to identify specific species among a complex of vegetation composition along roadsides. In this study, hyperspectral images (63 bands in visible to near-infrared spectral region) with high spatial resolution (1 m) were analyzed to detect cutleaf teasel in two areas along a 6.44-km (4-mi) section of Interstate I-70 in mid Missouri. The identified classes included cutleaf teasel, bare soil, tree/shrub, grass/other broadleaf plants, and water. Classification of cutleaf teasel reached a user's accuracy of 82 to 84% and a producer's accuracy of 89% in the two sites. The conditional κ value was around 0.9 in both sites. The image-classified cutleaf teasel map provides a practical mechanism for identifying locations and extents of cutleaf teasel infestation so that specific cutleaf teasel management techniques can be implemented.

Cutleaf teasel is an exotic weed that infests roadside environments in Missouri. As a growing biennial, the plant develops as a rosette during the first year and bolts during the second. Dense patches contain flowering plants with understory rosettes. The objective of this work was to develop approaches for detecting cutleaf teasel patches with accurate assessment in a complex of species along a roadside. Thus, management of cutleaf teasel could be located at specific sites. Two hyperspectral images (63 bands with 1-m spatial resolution) were analyzed to detect cutleaf teasel along the Interstate Highway I-70 in mid Missouri. Classification of cutleaf teasel reached a user's accuracy of 82 to 84% and a producer's accuracy of 89% at the two sites. The image-classified teasel map provides a practical mechanism for identifying the locations and extents of cutleaf teasel infestation so that specific management techniques can be implemented.

Field experiments were conducted in 1999 at Stoneville, MS, to determine the potential of multispectral imagery for late-season discrimination of weed-infested and weed-free soybean. Plant canopy composition for soybean and weeds was estimated after soybean or weed canopy closure. Weed canopy estimates ranged from 30 to 36% for all weed-infested soybean plots, and weeds present were browntop millet, barnyardgrass, and large crabgrass. In each experiment, data were collected for the green, red, and near-infrared (NIR) spectrums four times after canopy closure. The red and NIR bands were used to develop a normalized difference vegetation index (NDVI) for each plot, and all spectral bands and NDVI were used as classification features to discriminate between weed-infested and weed-free soybean. Spectral response for all bands and NDVI were often higher in weed-infested soybean than in weed-free soybean. Weed infestations were discriminated from weed-free soybean with at least 90% accuracy. Discriminant analysis models formed from one image were 78 to 90% accurate in discriminating weed infestations for other images obtained from the same and other experiments. Multispectral imagery has the potential for discriminating late-season weed infestations across a range of crop growth stages by using discriminant models developed from other imagery data sets.

Multispectral images of leaf reflectance in the visible and near infrared region from 384 to 810 nm were used to establish the feasibility of developing a site-specific classifier to distinguish lettuce plants from weeds in California direct-seeded lettuce fields. An average crop vs. weed classification accuracy of 90.3% was obtained in a study of over 7,000 individual spectra representing 150 plants. The classifier utilized reflectance values from a small spatial area (3 mm diameter) of the leaf in order to allow the method to be robust to occlusion and to eliminate the need to identify leaf boundaries for shape-based machine vision recognition. Reflectance spectra were collected in the field using equipment suitable for real-time operation as a weed sensor in an autonomous system for automated weed control.

Weed-detecting reflectance sensors were modified to allow selective interrogation of the near infrared–red ratio to estimate differences in plant biomass. Sampling was programmed to correspond to the forward movement of the field of view of the sensors. There was a linear relationship (r 2 > 0.80) between actual biomass and crop canopy analyzer (CCA) values up to 2,000 kg/ha for winter wheat sequentially thinned to create different amounts of biomass and up to 1,000 kg/ha for spring wheat sampled at different stages of development. At higher amounts of biomass the sensors underestimated the actual biomass. A linear relationship (r 2 = 0.73) was obtained with the CCA for the biomass of 76 chickpea cultivars at 500 growing degree days (GDD500). The reflectance sensors were used to determine differences in the herbicide response of soybean cultivars sprayed with increasing rates of herbicides. The CCA data resulted in better dose–response relationships than did biomass data for bromoxynil at 0.8 kg ai/ha and glyphosate at 1.35 kg ai/ha. There was no phytotoxicity to soybean with imazethapyr at 1.44 kg ai/ha. The method offers a quick and nondestructive means to measure differences in early-season crop growth. It also has potential in selecting crop cultivars with greater seedling vigor, as an indicator of crop nutrient status, in plant disease assessment, in determining crop cultivar responses to increasing herbicide dose rates, in weed mapping, and in studying temporal changes in crop or weed biomass.