Multi-Spectral Imaging for Weed Identification in Herbicides Testing

2.1 Imaging System

The imaging system was composed of a four-band Parrot Sequoia multi-spectral camera.1 Due to its small size and light weight, this camera can be mounted on several platforms (aircraft, terrestrial vehicles, or unnamed aerial vehicles (UAV)) to carry out radiometric measurements. Collected data can be extracted from the camera in three different ways: via USB, WiFi and SD card. Parrot Sequoia camera has a sensor resolution of 1280 × 960 pixels, 1.2 megapixels, a size of 4.8 × 3.6 mm, and collects data in discrete spectral bands: Green (GRE, 550 nm, 40 nm bandwidth), Red (RED, 660 nm, 40 nm bandwidth), Red Edge (REG, 735 nm, 10 nm bandwidth) and Near Infrared (NIR, 790–40 nm bandwidth), with a 10-bit depth. A sunshine sensor can be mounted together with the camera for accurate radiometric correction. The device is a fully-integrated sunshine sensor that captures and logs the current lighting conditions and automatically calibrates outputs of the camera so measurements are absolute. Figure 2 shows a Sequoia camera with the imaging sensors (NIR, RED, GRE and REG) and the sunshine sensor.

OPEN IN VIEWER

Fig. 2

Parrot Sequoia camera (the imaging and sunshine sensors are shown).

Pixel-based classifications have difficulty adequately or conveniently exploiting expert knowledge or contextual information. Object-based image-processing techniques overcome these difficulties by first segmenting the image into meaningful multi-pixel objects of various sizes, based on both spectral and spatial characteristics of groups of pixels (Flanders et al., 2003).

eCognition Developer2 is a powerful development environment for object-based image analysis. It is used in earth sciences to develop rule sets (or applications for eCognition Architect) for the automatic analysis of remote sensing data. Trimble eCognition software is used by Geographic Information System professionals, remote sensing experts and data scientists to automate geospatial data analytics. This software classifies and analyses imagery, vectors and point clouds using all the semantic information required to interpret it correctly. Rather than examining stand-alone pixels or points, it distils meaning from the objects’ connotations and mutual relationships, not only with neighbouring objects but throughout various input data (Geospatial, 2022).

In this work, eCognition developer version 9.5.0 has been used to classify pixels using Object-based image-processing techniques.

2.3 Field Experiment

The study site was located in Córdoba, Spain (349042, 4198307 UTM coordinates, zone 30). It was a sunflower (Helianthus annuus, L.) crop when the plants had about four to eight leaves and a height of approximately 8 to 20 cm. Moreover, the soil contained some weed species in early development (Chenopodium album, L., Convolvulus arviensis, L.; and Cyperus rotundus, L.). A rectangular frame of 57 × 47 cm was used as reference to scale and identify equivalent points in all bands. The frame was placed on the ground wrapping some sunflower and weed plants up (Fig. 3). A picture was taken with the multi-spectral camera, which was mounted on a platform, including the power supply, voltage regulators and a monitor. This platform was transported by an operator who framed the scene, while another operator was in charge of shooting the Sequoia camera through the mobile phone’s WiFi. The distance from the camera to the plants was approximately between 2 and 3 m, which is equivalent to distances used when terrestrial vehicles or UAV at low flight altitude are used as platforms (Louargant et al., 2018). A total of 26 multi-spectral images of the sunflower crop were considered in the experimental study (each one consisting of 4 bands).

OPEN IN VIEWER

Fig. 3

Raw image where the frame used in the experiment is shown.

Band co-registration of the images from Sequoia camera is not accurate enough for weed detection. A rectification methodology based on lens distortion and perspective distortion corrections has been considered to solve this problem.

2.4.1 Lens Distortion Correction

To correct lens distortion, a lens calibration process which involves calculating the intrinsic parameters of the lens has been considered (Shah and Aggarwal, 1996). A checkerboard of cells of 28 × 28 mm has been taken as the reference object. The reference object has been photographed from different angles. Figure 4 shows an example of a raw image of the reference in the NIR band. As can be noticed, lens distortion is quite evident. Figure 5 shows a scheme of the lens calibration process, where the inputs are four calibration sets of the checkerboard images (one set per band: GRE, NIR, RED and REG). A set of 95 checkerboard images was used for the calibration process.

OPEN IN VIEWER

Fig. 4

Sample photograph in the NIR band using Sequoia, showing a large lens distortion.

The calibration has been carried out using the pinhole calibration algorithm of the Matlab Camera Calibration Toolbox. It is based on the model proposed by Jean-Yves Bouguet (Bouguet, 2015) and it includes the pinhole camera model (Zhang, 2000) and lens distortion (Heikkila and Silven, 1997). The pinhole camera model does not account for lens distortion because an ideal pinhole camera does not have a lens. To accurately represent a real camera, the full camera model used by the algorithm includes radial and tangential lens distortion.

OPEN IN VIEWER

Fig. 5

Scheme of the lens calibration process.

2.4.2 Assessment of Lens Calibration

A Matlab routine has been developed to measure the quality of the lens calibration process described in Section 2.4.1. The scheme of the test calibration process is shown in Fig. 6. The inputs were the intrinsic parameters of each lens (computed in Section 2.4.1) and four test sets of the checkerboard images, one set per sensor.

OPEN IN VIEWER

Fig. 6

Scheme of the test calibration process.

OPEN IN VIEWER

Fig. 7

Test checkerboard images with corrected barrel distortion in GRE band.

OPEN IN VIEWER

Fig. 8

Rectified test checkerboard images in GRE band. Red crosses mark intersections that will be stored in matrix A = ( a i j ) . B = ( b i j ) will be calculated using the theoretical positions of the intersections shown.

The intersections of the rectified image are stored in matrix A. Then, the theoretical points that should be in the intersection points are computed and stored in matrix B. Finally, C = ( c i j ) = | ( a i j − b i j ) | , where A = ( a i j ) and B = ( b i j ) . Matrix C shows the intersection error per point. The mean value of ( c i j ) is the mean error of the image. Figure 7 shows the 16 test images for the GRE band with the lens distortion eliminated. In green colour the four outer intersections are marked. These intersections are needed to bring the image to the main plane and thus to correct the perspective distortion. The output of this step can be observed in Fig. 8. The intersections of the rectified image are stored in matrix A. Then, the theoretical points that should be in the intersection points are computed and stored in matrix B. Finally, C = ( c i j ) = | ( a i j − b i j ) | , where A = ( a i j ) and B = ( b i j ) . Matrix C shows the intersection error per point. The mean value of ( c i j ) is the mean error of the image.

Results of lens calibration will be discussed in Section 3.1.

To obtain a composited multi-spectral image from the four bands (GRE, NIR, RED and REG) a process based on the Computer Vision Toolbox of Matlab has been carried out. Figure 9 shows the scheme of the perspective distortion correction addressed in this work. In this process, raw images of the sunflower crop have been automatically corrected for distortion employing the “undistortImage” function of the Computer Vision Toolbox and the intrinsic parameters of each lens (computed in Section 2.4.1). The rectification process has been performed detecting four reference points (the inner corners) in the frames placed on the photographs and performing a homography to bring them to the main plane. A homography consists of a projective transformation (Goshtasby, 1986, 1988) using the “fitgeotrans” function of the Computer Vision Toolbox.

OPEN IN VIEWER

Fig. 9

Scheme of the steps of perspective distortion correction and multi-spectral image composition.

After correction of lens and perspective distortions, images in each band may be different in size due to errors inherent in the correction processes: error when selecting the vertices of the frame, error in the correction of lens distortions, and error in the correction of perspective. In addition, the co-registration process also involves the commission of an error that accumulates to the previous ones. Image co-registration has been carried out setting as reference the band of the greatest resolution image. Through an affine transformation, the Toolbox of Matlab makes the co-registration of the rest of the bands by fitting every frame vertex to its equivalent in the reference band. The composed multi-spectral images are the inputs of the classification process.

Figure 10 shows the whole band-to-band co-registration process for the composition of a multi-spectral image from the raw images of the sunflower crop. Concretely, the NIR band is taken as an example. Figure 10(d) is an RGB representation of the three most representative bands (GRE, NIR and RED). These RGB images will be the input for the Deep Learning methods explained in Section 2.8.

OPEN IN VIEWER

Fig. 10

(a) Raw photograph in the NIR band for sample 38; (b) NIR band of sample 38 after lens distortion correction; (c) NIR band after perspective correction taking as reference the four inner corners of the frame; and (d) RGB representation of three overlapping bands (GRE, NIR and RED).

In this section, a novel algorithm for multi-spectral imaging classification based on e-Cognition is described. Figure 11 shows a scheme of the classification processes carried out in this work. The inputs of the process are the previously generated multi-spectral images and the output is the classification of all the pixels of the image in three classes: sunflower, weed and soil. It is worth noting that in the experiments carried out it has not been possible to distinguish the sunflower from the weed only using the spectral signature. The spectral signature in close-up photography (from 2 to 3 metres in our case study) is highly variable, as each leaf has a different level of illumination, there are shadows within the same plant and the plants contain several parts (branches, leaves in different stages of growth, etc.) with their respective spectral signatures. Therefore, it was also necessary to use geometric factors to differentiate sunflowers from weed. These geometric factors are the assumption that the sunflower plant is larger and more developed than the surrounding weed.

OPEN IN VIEWER

Fig. 11

Scheme of the classification process.

Here we focus on identifying sunflower plants using a seedling growth strategy. These kind of algorithms are based on a set of initial points called seeds that grow annexing adjacent regions that have similar properties (e.g. texture or colour) (Baatz and Schäpe, 2000). These seeds have been determined using the most descriptive leaves of sunflowers, taking into account factors of light, shape and size. In the process three main blocks can be distinguished: vegetal matter-soil classification, sunflower node detection and an iterative region growing algorithm. The steps carried out in each block are described as follows.

In the vegetal matter-soil classification, two steps are computed. Firstly, a clusterization based on multi-resolution segmentation with the spectral signatures is carried out, using the four bands (GRE, NIR, RED and REG) and the Normalized Difference Vegetation Index (NDVI). NDVI is an indicator of the presence of vegetal matter (Weier and Herring, 2000), and it can be formulated as follows: NDVI = NIR − RED NIR + RED . According to this index, plants have positive values between 0.2 and 1. In the study case, NDVI is considered to classify every cluster as vegetal matter or soil. Figure 12(a) illustrates the output of this block for sample 38. Any cluster with an NDVI value greater than 0.2 is classified as vegetal matter and printed in green colour. Finally, we group all vegetal matter clusters into nodes and discard false positives (nodes equal or less than 0.03% of the total size of the image) (see Fig. 12(b)).

OPEN IN VIEWER

Fig. 12

(a) Output image for sample 38 after executing the block vegetal matter-soil classification. Vegetal matter and soil are printed in green and black colour, respectively; (b) Image after grouping of vegetal matter into adjacent nodes and removing cluster less that 0.03% of the total size of the image; (c) Image after executing the multi-spectral quadtree segmentation; (d) Image with the square-shaped clusters marked (seeds); (e) Image after the first stage of the region growing algorithm; (f) Image of the second stage of the region growing algorithm. Potential sunflowers are marked in yellow and potential weed in red colour; (g) Image of the third stage of region growing algorithm; (h) Output image of the classification process from the eCognition visor. Yellow areas represent vegetal matter classified as sunflower, red areas represent vegetal matter classified as weed and black areas identify soil; and (i) Output image of the classification process.

The next block of the classification process (see Fig. 11) is the sunflower node detection. Vegetal matter is segmented using a multi-spectral quadtree algorithm, to identify large square areas belonging to the same leaf (see Fig. 12(c)). The four bands and the derived NDVI layer are considered as inputs for this quadtree based segmentation. A large scale parameter (in an order of 25000 variations of colour) is required to obtain large squares that point to the most relevant leaves of the plants of interest. The scale parameter is an important element when performing segmentation in remote sensing, and its correct estimation is the subject of study in several articles (Drăguţ et al., 2014; Yang et al., 2019; El-naggar, 2018). The input images have a colour depth of 16 bits (65536 values). To obtain a multi-spectral quadtree segmentation result adjusted to the size of the sunflower leaves, an average variation of 7.5% (5000 colour variations on average in each of the 5 layers used) of the colour in each band has been estimated. Since there are 5000 possible variations on average in each of the 5 layers, the considered scale parameter has been 25000.

The next step consists of marking as sunflowers the previously obtained clusters because of their size, shape and luminosity level. The best cluster candidates to be sunflower leaves are large, almost perfectly shaped squares with a high level of luminosity in the NIR band. This step focuses on the NIR band because it is the brightest band and this allows to better identify the most developed and descriptive leaves to classify a sunflower plant.

The two factors for classifying these clusters as sunflowers have been luminosity level per size of cluster and compactness. Figure 12(d) shows the nodes marked as sunflowers.

The last block of the classification process (see Fig. 11) is based on a region growing algorithm. The region growing algorithm uses potential square-shaped clusters that can be part of the same leaf. To identify potential candidates to grow, a derived layer called Diff_RED_NIR is computed by adding the mean difference to neighbour clusters in RED and NIR bands. The objective of this new layer is to identify square-shaped clusters with similar values in the RED and NIR layers.

The region growing algorithm takes place in three stages. The first stage aims to mark as sunflower the neighbour vegetal matter clusters to the sunflower nodes. For this study case, the threshold is Diff_RED_NIR > −1000. The threshold has been calculated to have mean value differences between the two bands (RED and NIR) of less than 1% relative. This means that we are growing by adding very similar clusters (with a very small relative level of difference in those two bands).

To avoid growing into twigs that may be part of other plants, a compactness <1.5 is required for the neighbour clusters to be marked as sunflower. This stage is done ten times per iteration. Figure 12(e) shows the results of the first step of the region growing algorithm.

In the second stage of the iterative process, areas of vegetal matter enclosed by objects classified as sunflower are themselves marked as sunflowers. All sunflower regions are merged and grow to any neighbour cluster with a relative border greater than 30%. This means that growth occurs with clusters with which more than one side is shared, which would normally be 25%. The resulting sunflower objects are merged and all the other objects are classified as weed (see Fig. 12(f)). The final stage of this block is to compute a multi-spectral quadtree segmentation in the merged weed area, which considers as input the calculated layers, the bands and a scale parameter of 25000 (as it was explained in the sunflower node detection process). Obtained image of this step is depicted in Fig. 12(g). The iterative process is carried out three times since more iterations result in marginal growth. This iterative process results in clusters with a high potential to be the most representative leaves of sunflowers.

The last step of the region growing algorithm is to consider as sunflower any weed cluster with more than 5 % of the common border. These will be the darker parts and edges of the sunflower plant. It is possible at this point to accept small pieces of weed as false positives. Moreover, any small sunflower clusters are eliminated, since sunflower plants are expected to have a greater size than the weed. It is possible to falsely identify some nodes as sunflowers if the weeds form a big cluster of small leaves that can be misidentified as a large sunflower leaf. These false positives are removed by classifying any cluster of sunflower with an area inferior to 10% of the biggest sunflower cluster (3500 pixels as the threshold in our case as a reference) as weed. The output of this step is shown in Fig. 12(h). Figure 12(i) shows the output of the classification process, where yellow, red and black colours identify sunflower, weed and soil regions, respectively.

A general scheme of the developed classification method can be found in Algorithm 1.

OPEN IN VIEWER

Algorithm 1

Proposed classic computer vision algorithm

The comparison of the proposed classificator (based on classical techniques) with other state-of-the-art classification models is of great interest. Widely used Deep Learning-based segmentation methods are U-Net and Feature Pyramid Network (FPN) and we have implemented both is this paper.

U-Net is an encoder-decoder based model, these kind of models are the most popular segmentation models based on Deep Learning (DL). U-Net was proposed in Ronneberger et al. (2015) for image segmentation in the medical and biomedical field. The U-Net strategy is to supplement a usual contracting network by successive layers, where pooling operations are replaced by upsampling operators for achieving higher resolution outputs on the input images. On the other hand, FPN is a feature extractor that takes a single-scale image of an arbitrary size as input, and outputs proportionally sized feature maps at multiple levels, in a fully convolutional fashion (Lin et al., 2017). This approach combines low-resolution, semantically strong features with high-resolution, semantically weak features via a top-down pathway and lateral connections.

Deep Learning models need to use pre and post-processing techniques for improving the quality of the obtained results in complex problems. Some of the most important methods are: transfer learning, fine-tuning and data augmentation methods. We have considered these three techniques in both U-Net and FPN implementations.

The transfer learning technique considers the use of pre-trained models on a database of millions of instances (such as ImageNet,3 Russakovsky et al., 2014) as a starting point to initialise the weights of the network, thus taking advantage of previously learned knowledge. With the fine-tuning technique, the new model only has to train the last few layers, thus taking less time to obtain favourable results. Finally, the Data Augmentation (DA) technique is used to increase the number of input images by applying variations to the starting images. In this way, the model will have a larger database for training and thus make the model more generalisable. Taking into account such techniques, we have prepared three training datasets as follows.

•

I_Jes: It considers Imagenet for transfer learning and the Jesi_05_18 dataset of sunflower crops for fine-tuning (146 RGB images). The latter dataset was introduced in Fawakherji et al. (2021) and can be downloaded from.4 Every image of the Jesi_05_18 dataset has been augmented with 1000 images using basic image operations (rotation, shifting, flipping, zooming, and cropping). Moreover, authors supply mask images required for training. They consist of segmentation masks where crop, floor and weed are identified. I_Jes dataset contains a total of 146000 images. The percentage of these images for training and testing was 90% and 10%, respectively.

•

Multi: Training is carried out from scratch using RGB representations of 10 multi-spectral images (generated in Section 2.6) together with the corresponding masks and data augmentation. In this case, we have manually created the masks. Of the 10 multi-spectral images, 9 were considered for training and increased 1000 times (obtaining a total of 9000 images), and one multi-spectral image was used for testing, with an increase of 1000 times.

•

I_Jes_Multi: It is an extension of the I_Jes dataset adding the 10000 images of Multi dataset. The total number of images is 156000. The percentage of these images for training and testing was 90% and 10%, respectively.

2.9 Assessment of Classification Process

To quantify the accuracy of the proposed classification algorithm and other classification approaches based on DL, a comparison between predictions for each segmented object with the corresponding ground truth information has been carried out. To obtain the ground truth, we manually labelled all images used for the evaluation. The model performance has been measured in terms of Intersection-over-Union (IoU) which is described as: I o U = prediction ∩ ground truth prediction ∪ ground truth = TP FP + TP + FN , where TP, FP and FN define the number of true positive, false positive and false negative samples, respectively. IoU is measured for every class (sunflower, weed and soil) of the dataset resulting to the mean IoU (mIoU) as the average IoU measured over each class. Results will be discussed in Section 3.2.

3.1 Lens Calibration Discussion

Table 1 shows the results of the calibration process using several checkerboard images (sample images) per sensor. This table points out the number of calibration and test set images and also the average error (in pixels) after finishing the calibration test. As can be drawn for this table, an average error of fewer than 0.3 pixels has been obtained in all bands of interest. The NIR band is the one that yields the best quality in the camera, as most of the checkerboards in the images were recognizable.

Three different approaches have been implemented to classify soil, sunflower and weed: a segmentation based on classical techniques and two Deep Learning models (U-Net and FPN). The final goal is to automatically determine the effectiveness of a herbicide in weed.

Regarding U-Net and FPN, a preliminary study has been carried out to determine the best models and datasets for the classification of crop and weed. U-Net and FPN models have been trained using the three datasets detailed in Section 2.8. The six resulting neural networks have been validated using a test dataset of 16 RGB representations of multi-spectral images (generated as explained in Section 2.6), as it shown in Table 2. Moreover, the batch size was 16. In such table “Epochs for best IoU” column identifies the number of epochs that the neural network has needed to obtain the best value of the loss function in the validation dataset. Notice that in this problem the considered loss functions have been Dice and Local Loss Functions. Both functions have demonstrated better results than Jaccard Loss Function for detecting the classes with less representation (weed) in the dataset images. Furthermore, “Total epochs” column identifies the total number of epochs of each experiment considering the early stopping criteria with a value of 30. The best models in terms of mIoU for sunflower and weed will be compared to the proposed classical segmentation in this paper.

Table 4 reports, in terms of mIoU, the performance to classify vegetal matter, sunflower and weed using the classical and the two DL approaches. The same 16 sample images of Table 3 have been considered for the average. As can be seen from Table 4, the proposed methodology is able to classify vegetal matter with an accuracy of 91%, the plant of interest with an accuracy of 89%, and weed with an accuracy of 72%. If we compare these values with respect to the DL-based models studied, we obtain that our model is the best at classifying vegetal matter in general and weed in particular. U-Net_I_Multi approach is the best at classifying sunflower, although our approach is quite competitive, as it obtains a very close mIoU. Therefore, we can conclude that when we do not have a high number of images of the study field and their corresponding segmentation masks, the traditional classification proposed in this work is very appropriate.

Furthermore, a comparison among the obtained results in this paper and the obtained mIoU obtained in the paper (Fawakherji et al., 2021) has been carried out. In this state-of-the-art paper, authors performed a study based on DL approaches (incorporating a modification of the common data augmentation methods) on a sunflower crop using multi-spectral images. The considered dataset to train and test the model was an extension of the I_Jes dataset (by using sunflowers in different stages of growth). The best IoU value for weed was 0.69 using Bonnet architecture. Therefore, we can establish that the classical method that we have proposed in the present paper is slightly better for weed classification. Moreover, our approach does not require the generation of datasets and the model training process which are time-consuming tasks.

From Table 2 we draw the following conclusions: (1) using a dataset of images (I_Jes) similar to our problem for fine-tuning do not show sufficient performance for the studied problem of weed classification; (2) using ad-hoc images of our problem to train the neural network (I_Jes_Multi and Multi) results in a large improvement in quality, despite using a smaller training dataset; (3) the use of I_Jes_Multi dataset results in a considerable increase in the number of epochs to obtain the best IoU value. This latter option is slightly better for the FPN model, and slightly worse for the U-Net model. Although the obtained results of U-Net and FPN models using I_Jes did not obtain sufficient performance for our problem, we have validated them over the authors’ original problem obtaining similar results to our best results from the Table 2. The U-Net_Multi and the FPN_I_Jes_Multi obtained the best mIoU for classifying crop and weed, therefore they have been considered to be compared with our approach based on classical segmentation.

Table 3 shows the IoU values for sunflower, weed and vegetal matter (sunflower ∪ weed) using the methodology proposed throughout this paper for the three studied approaches (Proposed, U-Net_Multi and FPN_I_Jes_Multi) and 16 multi-spectral sample images. Furthermore, in such table the percentage of vegetal matter, sunflower and weed of each image sample is shown. From the table it can be seen that: (1) the IoU of sunflower is slightly better using DL than our proposal; (2) the IoU of weed is slightly better in the classical segmentation-based approach; and (3) in the case of vegetal matter, our proposal improves on the DL-based segmentation to a greater extent.

As an example, Fig. 13 shows the RGB representation of three bands (RED, NIR and GRE) for samples 34, 25 and 15, its corresponding segmented images after applying the proposed classic algorithm; and, finally, its ground truth.

OPEN IN VIEWER

Fig. 13

Three sample images (34, 25 and 15) segmented by the proposed classical algorithm (one per row). (a), (d) and (g) show a RGB representation of the multi-spectral image using RED, NIR and GRE bands; (b), (e) and (h) show the output of the algorithm; and (c), (f) and (i) show the ground truth determined by humans.