Gawain Jones

Modelling agronomic images for weed detection and comparison of crop/weed discrimination algorithm performance

A new method for weed detection based on modelling agronomic images taken from a virtual camera placed in a virtual field is proposed. The aim was to measure and compare the effectiveness of the developed algorithms. Two sets of images with and without perspective effects were simulated. For images with no perspective, based on Gabor filtering and on the Hough transform, the performance of two crop/inter-row weed discrimination algorithms were tested and compared. The method based on the Hough transform is, in any case, better than the one based on Gabor filtering. For images with perspective effects only, an algorithm based on the Hough transform was tested and an extension to real images is discussed. These tests were done by a comparison between the weed infestation rate detected by these algorithms and the true one. This evaluation was completed with a crop/weed pixel classification and it demonstrated that the algorithm based on a Hough transform gave the best results (up to 90%).

Field radiometric calibration of a multispectral on-the-go sensor dedicated to the characterization of vineyard foliage

Development of a multispectral on-the-go system (visible and NIR) to characterize vineyard in natural light.Implementation of a radiometric method to calibrate the images using linear and spatial interpolation.Computation of vegetation index (NDVI) from reflectance images and strong correlation with Greenseeker data.Proximal multispectral imaging system can be used to assess foliage vigor at green berry stage. The accurate assessment of the vigor and disease impact is a major challenge in precision viticulture. It is essential for managing phytosanitary treatments. Up to now, some remote sensing techniques such as aerial imagery and handheld optical sensors have been applied to grapevine characterization. However each technique provides limited, specific information about foliage. To broaden the characterization of the foliage, we developed a proximal integrated, multispectral imaging sensor that operates in the visible and near-infrared bands. It is mounted on a track-laying tractor equipped with a Greenseeker-RT-100, coupled with a GPS-RTK. As the sensor is very sensitive to the ambient light, a radiometric calibration is required: it allows producing absolute reflectance images, using a color chart. If the chart is hidden by leaves, for instance, the images are corrected using the linear interpolation method. The adaptive radiometric method is evaluated as a function of the number of neutral patches selected on the color chart during the linear regression process and the efficiency of the spatial interpolation method is assessed using a leave-one-out-cross-validation (LOOCV) method.The radiometric calibration is validated by comparison of NDVI maps produced by imagery and by the Greenseeker, a commercial system. In the early stage of berry formation, we examined and quantified the spatial patterns and demonstrated a low-cost imagery method that is capable of analyzing correctly the vigor. This corroborates the efficiency of the calibration method encouraging the use of multi-spectral imagery for other vineyard applications, such as the characterization of physiological status.

Crop/weed discrimination in simulated images

In the context of site-specific weed management by vision systems, an efficient image processing for a crop/weed discrimination is required in order to quantify the Weed Infestation Rate (WIR) in an image. This paper presents a modeling of crop field in presence of different Weed Infestation Rates and a set of simulated agronomic images is used to test and validate the effectiveness of a crop/weed discrimination algorithm. For instance, an algorithm has been implemented to firstly detect the crop rows in the field by the use of a Hough Transform and secondly to detect plant areas by a region based-segmentation on binary images. This image processing has been tested on virtual cereal fields of a large field of view with perspective effects. The vegetation in the virtual field is modeled by a sowing pattern for crop plants and the weed spatial distribution is modeled by either a Poisson process or a Neyman-Scott cluster process. For each simulated image, a comparison between the initial and the detected weed infestation rate allows us to assess the accuracy of the algorithm. This comparison demonstrates an accuracy of better than 80% is possible, despite that intrarow weeds can not be detected from this spatial method.

Simulation of agronomic images for an automatic evaluation of crop/weed discrimination algorithm accuracy

In the context of precision agriculture, we present a robust and automatic method based on simulated images for evaluating the efficiency of any crop/weed discrimination algorithms for a inter-row weed infestation rate. To simulate these images two different steps are required: 1) modeling of a crop field from the spatial distribution of plants (crop and weed) 2) projection of the created field through an optical system to simulate photographing. Then an application is proposed investigating the accuracy and robustness of crop/weed discrimination algorithm combining a line detection (Hough transform) and a plant discrimination (crop and weeds). The accuracy of weed infestation rate estimate for each image is calculated by direct comparison to the initial weed infestation rate of the simulated images. It reveals an performance better than 85%.

Unsupervised Classification Algorithm for Early Weed Detection in Row-Crops by Combining Spatial and Spectral Information

In agriculture, reducing herbicide use is a challenge to reduce health and environmental risks while maintaining production yield and quality. Site-specific weed management is a promising way to reach this objective but requires efficient weed detection methods. In this paper, an automatic image processing has been developed to discriminate between crop and weed pixels combining spatial and spectral information extracted from four-band multispectral images. Image data was captured at 3 m above ground, with a camera (multiSPEC 4C, AIRINOV, Paris) mounted on a pole kept manually. For each image, the field of view was approximately 4 m × 3 m and the resolution was 6 mm/pix. The row crop arrangement was first used to discriminate between some crop and weed pixels depending on their location inside or outside of crop rows. Then, these pixels were used to automatically build the training dataset concerning the multispectral features of crop and weed pixel classes. For each image, a specific training dataset was used by a supervised classifier (Support Vector Machine) to classify pixels that cannot be correctly discriminated using only the initial spatial approach. Finally, inter-row pixels were classified as weed and in-row pixels were classified as crop or weed depending on their spectral characteristics. The method was assessed on 14 images captured on maize and sugar beet fields. The contribution of the spatial, spectral and combined information was studied with respect to the classification quality. Our results show the better ability of the spatial and spectral combination algorithm to detect weeds between and within crop rows. They demonstrate the improvement of the weed detection rate and the improvement of its robustness. On all images, the mean value of the weed detection rate was 89% for spatial and spectral combination method, 79% for spatial method, and 75% for spectral method. Moreover, our work shows that the plant in-line sowing can be used to design an automatic image processing and classification algorithm to detect weed without requiring any manual data selection and labelling. Since the method required crop row identification, the method is suitable for wide-row crops and high spatial resolution images (at least 6 mm/pix).

« On-the-go » multispectral imaging system to characterize the development of vineyard foliage with quantitative and qualitative vegetation indices

Over the last years, the literature presents new technologies to optimize vineyard management. In the proximal sensing context, optical sensors are mainly developed to characterize the vegetation and the most famous one is the Greenseeker RT-100 (Trimble, Germany), that provides NDVI. The interpretation of its measurements is complex because it overlaps quantitative and qualitative information. However, it is a robust active sensor especially dedicated to characterize vineyard at early growth stage. To overcome these limits, we developed a multispectral (RGB, NIR) imaging system. We present a first application of spectral imagery, in proximal sensing conditions, to characterize the vine foliage of three grapevine varieties (Meunier, Pinot Noir and Chardonnay) at four phenological stages. The imaging system is embedded on a ground vehicle acquiring images with natural light, and an original radiometric calibration is proposed. From images, three agronomic indices (NDVIimage, NDVIvegetation and “foliage occupation”) are defined. They are computed from entire images and from the area of the grapes. These indices are compared to Greenseeker ones at the beginning of berry formation to be assessed. Whatever the grapevine variety the NDVIimage is in agreement with the index provided by Greenseeker (NDVIGS). At the other stages, the comparison of NDVIGS to the other indices leads to a new interpretation of NDVIGS depending on the phenological stage. The new indices provide a better understanding on the part of quantitative and quantitative information in Greenseeker index and lead to a more accurate leaf quantity estimation (from entire images), or specific physiological status characterization.

Validation of a crop field modeling to simulate agronomic images

In precision agriculture, crop/weed discrimination is often based on image analysis but though several algorithms using spatial information have been proposed, not any has been tested on relevant databases. A simple model that simulates virtual fields is developed to evaluate these algorithms. Virtual fields are made of crops, arranged according to agricultural practices and represented by simple patterns, and weeds that are spatially distributed using a statistical approach. Then, experimental devices using cameras are simulated with a pinhole model. Its ability to characterize the spatial reality is demonstrated through different pairs (real, virtual) of pictures. Two spatial descriptors (nearest neighbor method and Besags function) have been set up and tested to validate the spatial realism of the crop field model, comparing a real image to the homologous virtual one.

Mapping Vineyard Folliage Density with Multispectral Proxidection Imagery

This paper presents a new multispectral imaging system and a new approach to map vineyard leaf development in the context of Precision Viticulture. It is based on the use of a 2-CCD camera (RGB/NIR). This camera is embedded on a track laying tractor and associated with a GNSS system. It works with natural ambient light. A Green seeker is also embedded to validate the first results of imagery. It is a sensor usually used in viticulture to assess the foliage density of vines. The images are calibrated in order to correct geometric distortion and variations of the natural light. The parameters of the geometric correction are stable over time, whereas the radiometric correction needs to include a color checker in the background of all the images. For each of them, a calibration model is computed and reflectance of objects is obtained. The calibrated images can be compared with each other. During 2013, four datasets of about 5000 images were acquired. For all the images, a vegetation index is performed to retrieve agronomic information. Based on these images, different vineyard maps were generated with calibrated images, with non-calibrated images, and with Green seeker values. We present and discuss some of them in order to assess the relevancy of the approach.

Combining spatial and spectral information to improve crop/weed discrimination algorithms

Reduction of herbicide spraying is an important key to environmentally and economically improve weed management. To achieve this, remote sensors such as imaging systems are commonly used to detect weed plants. We developed spatial algorithms that detect the crop rows to discriminate crop from weeds. These algorithms have been thoroughly tested and provide robust and accurate results without learning process but their detection is limited to inter-row areas. Crop/Weed discrimination using spectral information is able to detect intra-row weeds but generally needs a prior learning process. We propose a method based on spatial and spectral information to enhance the discrimination and overcome the limitations of both algorithms. The classification from the spatial algorithm is used to build the training set for the spectral discrimination method. With this approach we are able to improve the range of weed detection in the entire field (inter and intra-row). To test the efficiency of these algorithms, a relevant database of virtual images issued from SimAField model has been used and combined to LOPEX93 spectral database. The developed method based is evaluated and compared with the initial method in this paper and shows an important enhancement from 86% of weed detection to more than 95%.

A crop field modeling to simulate agronomic images

In precision agriculture, crop/weed discrimination is often based on image analysis but though several algorithms using spatial information have been proposed, not any has been tested on relevant databases. A simple model that simulates virtual fields is developed to evaluate these algorithms. Virtual fields are made of crops, arranged according to agricultural practices and represented by simple patterns, and weeds that are spatially distributed using a statistical approach. Then, experimental devices using cameras are simulated with a pinhole model. Its ability to characterize the spatial reality is demonstrated through different pairs (real, virtual) of pictures. Two spatial descriptors (nearest neighbor method and Besags function) have been set up and tested to validate the spatial realism of the crop field model, comparing a real image to the homologous virtual one.