Original papersLocal descriptors for soybean disease recognition
Introduction
Soybean is one of the most important crops due to its beneficial effects on human health, to its role as a major nutrition source, and to its economic importance. It has been widely used in food and industrial applications because of its high protein and oil concentrations (Kumar et al., 2010). Soybean occupies very large crops in which the monocropping and conservation tillage are commonly used. Such cultivation systems, however, have favored the occurrence of a large number of diseases (Carmona et al., 2015) causing major economic losses. The solution is to apply preventive agrochemicals; but, because the identification of where the infestation took place is time-consuming, the usual practice is to use agrochemicals over the entire crop instead of only over specific subareas. This is an expensive practice that spreads unnecessary chemicals over terrain and air.
Accordingly, a more precise detection of the disease spots in the crop is an important step to decrease economic losses, to prevent the spread of diseases, and to reduce environmental pollution. Despite its importance, it is usually conducted visually by an expert (Moshou et al., 2004), an imprecise and time-consuming process, especially when carried out over large-scale farms. Alternatively, disease detection techniques based on chemical reactives are available, such as the ELISA (enzyme-linked immunosorbent assay) method and the PCR (polymerase chain reaction) method (Saponari et al., 2008, Yvon et al., 2009, Gutiérrez-Aguirre et al., 2009), however, they are expensive processes. Consequently, there is a demand for rapid and cheaper detection methods.
In this context, one active line of research is the use of image processing techniques. The idea is to have the computer analyze images of soybean leaves (and of other cultures) to detect diseases by means of pattern recognition methods. Gui et al. (2015), for example, proposed a method for soybean disease detection based on salient regions and k-means clustering. Shrivastava and Hooda (2014) proposed a method for detecting brown spot and frog eye, two common soybean diseases; they used shape features and k-nearest neighbors classification. Ma et al. (2014) proposed a technique for detecting insect-damaged vegetable soybean using hyperspectral imaging. A study to discriminate soybean leaflet shape using neural networks was proposed in the work of Oide and Ninomiya (2000). Yao et al. (2012) used hyperspectral images to study the damage caused by the herbicide glyphosate on soybean plants. Cui et al. (2010) reported image processing techniques for quantitatively detecting rust severity from soybean multi-spectral images.
Besides soybean, other cultures have been studied in the literature, such as the work performed by Rumpf et al. (2010), which presents an automatic system for classification of foliar sugar beet diseases based on Support Vector Machines and spectral vegetation indices. Moshou et al. (2004) investigated the automatic recognition of yellow rust in wheat using reflectance measurements and neural networks. Liu et al. (2010) applied techniques neural network and principal components analysis for classifying fungal infection levels in rice panicles. Imaging techniques are also applied in the recognition of plant species (Gonçalves and Bruno, 2013, Casanova et al., 2012, Backes et al., 2010). A review of techniques for detecting plant diseases can be found in the work of Sankaran et al. (2010); a survey on methods that use digital image processing techniques to detect plant diseases is presented in the work of Barbedo (2013).
This paper proposes a novel approach for soybean disease recognition based on techniques local descriptors and bag-of-visual words. We experiment with five local descriptors (SURF, HOG, DSIFT, SIFT, and PHOW, as detailed in Section 3) applied over a large set of digital images (gray scale and colored) acquired from a real-world soybean plantation in Brazil. The proposed approach is applied to scanned images (visible spectrum to the human eye), which does not require hyperspectral images and, therefore, can be used with commodity hardware such as smartphones. From the extracted features, we calculate a summary (lower-dimensional) feature vector using technique bag of visual words (BOVW).
The use of local descriptors is attractive because they are distinctive, robust to occlusion, and do not require segmentation. Due to these advantages, several local descriptors have been proposed in the literature. Scale-invariant feature transform (SIFT) (Lowe, 2004) is one of the most used and known local descriptors. Due to its good performance, SIFT was later applied at dense grids, known as Dense SIFT (Vedaldi and Fulkerson, 2010, Liu et al., 2011), and at multiscale dense grids, known as Pyramid histograms of visual words (PHOW) (Bosch et al., 2007). SIFT also inspired other local descriptors such as Speeded-up Robust Features (SURF) (Bay et al., 2008), which uses some approximations to achieve better performance. Histogram of oriented gradients (HOG) (Dalal and Triggs, 2005) has also been widely used and has shown interesting results, especially in pedestrian recognition. The reader may refer to the work of Mikolajczyk and Schmid (2005) for a review of local descriptors applied over images in general.
For classification purposes – considering classes disease and no disease, we use the supervised machine learning technique Support Vector Machine (SVM) having as input the BOVW vectors. We evaluate our classification using classic ten-fold cross-validation and the metric Correct Classification Rate (CCR). In our experiments, descriptor PHOW, over colored images, achieved the highest performance. Therefore, we contribute by (i) introducing a systematic method for computational identification of diseases in soybean leaves; (ii) conducting an experiment over soybean that is unprecedented in its control, methodology, and scale; (iii) empirically comparing the main local descriptors found in the literature, providing guidance for future works on image-based classification.
The rest of this paper is organized as follows. Section 2 describes the fundamentals of bag-of-visual-words. The five local descriptors evaluated in this work are described in Section 3. Section 4 details the experimental design and the image dataset of soybean leaves, while Section 5 describes the results of the proposed approach. Finally, Section 6 concludes the paper and suggests future works.
Section snippets
Bag-of-visual-words – BOVW
The bag-of-visual-words (BOVW) (Csurka et al., 2004) is a popular model for image recognition inspired by the bag-of-words (BOW) used in natural language processing. According to BOVW, descriptors are extracted from images in order to build a vocabulary of visual words. Given the vocabulary, each descriptor of an image is assigned to one visual word and then a histogram of visual word occurrences is obtained to represent the image. Basically, this model can be described in the following steps (
Local descriptors
This section briefly describes the local descriptors used by our method for soybean disease recognition. We use gray-scale images and the local descriptors are applied in sub-regions of the leaf images, suggesting the use of local methods. Several methods for generating local descriptors have been reported in the literature and can be used as a previous step, such as selective search (Uijlings et al., 2013), objectness (Alexe et al., 2012), category-independent object proposals (Endres and
Experimental design
The plant experiment was done in four fields of the Phytopathology Department of the Federal University of Grande Dourados (UFGD), Brazil. The crop evaluated was soybean [Glycine max (L.) Merr.] of BMX Potencia RR (BRASMAX).
The density of the soybean fields was of about 300,000 plants ha−1. For all fields, 320 kg ha−1 of N-P-K (02-23-23) were applied in-furrow immediately before sowing. No N-fertilizer was applied in any field. The experimental design was a completely randomized block with four
Experiments and discussion
In this section, we describe experiments and results obtained with the use of local descriptors and BOVW. In the classification step, we have used the Support Vector Machine – SVM classifier using stratified 10-fold cross-validation. This methodology is well-accepted in machine learning and artificial intelligence. In the stratified 10-fold cross-validation, the images of the dataset are partitioned into 10 folds ensuring that each fold has the same proportion of each class. Then, one fold is
Conclusion
Soybean disease detection is an important process to decrease economic losses in agriculture, and to reduce environmental pollution due to the excessive use of agrochemicals. Answering to this demand, we proposed a new approach for automated detection of soybean diseases. We used image local descriptors and Bag of Visual Words (BOVW) to define a methodology able to computationally represent images of soybean leaves, while maintaining visual information regarding potential diseases. Our
Acknowledgment
The authors acknowledge support from PET-Fronteira, CAPES, CNPq and FUNDECT.
References (37)
- et al.
Texture analysis and classification using deterministic tourist walk
Pattern Recogn.
(2010) - et al.
Speeded-up robust features (surf)
Comput. Vis. Image Underst.
(2008) - et al.
Development and validation of a fungicide scoring system for management of late season soybean diseases in argentina
Crop Prot.
(2015) - et al.
Image processing methods for quantitatively detecting soybean rust from multispectral images
Biosyst. Eng.
(2010) - et al.
Combining fractal and deterministic walkers for texture analysis and classification
Pattern Recogn.
(2013) - et al.
Real-time quantitative {PCR} based sensitive detection and genotype discrimination of pepino mosaic virus
J. Virol. Methods
(2009) - et al.
A comparative assessment of total phenolic content, ferric reducing-anti-oxidative power, free radical-scavenging activity, vitamin c and isoflavones content in soybean with varying seed coat colour
Food Res. Int.
(2010) - et al.
Application of neural networks to discriminate fungal infection levels in rice panicles using hyperspectral reflectance and principal components analysis
Comput. Electron. Agric.
(2010) - et al.
Automatic threshold method and optimal wavelength selection for insect-damaged vegetable soybean detection using hyperspectral images
Comput. Electron. Agric.
(2014) - et al.
Automatic detection of ‘yellow rust’ in wheat using reflectance measurements and neural networks
Comput. Electron. Agric.
(2004)
Discrimination of soybean leaflet shape by neural networks with image input
Comput. Electron. Agric.
Early detection and classification of plant diseases with support vector machines based on hyperspectral reflectance
Comput. Electron. Agric.
A review of advanced techniques for detecting plant diseases
Comput. Electron. Agric.
Quantitative detection of citrus tristeza virus in citrus and aphids by real-time reverse transcription-pcr (taqman)
J. Virol. Methods
Using vegetation index and modified derivative for early detection of soybean plant injury from glyphosate
Comput. Electron. Agric.
Specific detection and quantification of the phytopathogenic agent ‘candidatus phytoplasma prunorum’
Mol. Cell. Probes
Measuring the objectness of image windows
IEEE Trans. Pattern Anal. Mach. Intell.
Multiscale combinatorial grouping
Cited by (74)
Pseudo-label semi-supervised learning for soybean monitoring
2023, Smart Agricultural TechnologyTrends in vision-based machine learning techniques for plant disease identification: A systematic review
2022, Expert Systems with ApplicationsCitation Excerpt :Their model attained an accuracy of 94.21% on 500 rice images in 4 categories collected from the field and laboratory. Pires et al. (2016) worked on soybean disease identification using local image descriptors and a bag-of-visual-words (BoVW). Their model achieved a 99.83% correct classification rate (CCR) using Pyramid Histograms of Visual Words (PHOW-HSV) with 2000 visual words on 1200 soybean leaf images from controlled conditions in four classes.
Optimized deep residual network system for diagnosing tomato pests
2022, Computers and Electronics in AgricultureCitation Excerpt :Kumar et al. used KNN clustering and an SVM classifier to diagnose sugarcane infection diseases, and the model achieved 96% accuracy (Kumar et al., 2017). Pires et al. proposed using image local descriptors and visual bag-of-words techniques to diagnose three common soybean diseases (Pires et al., 2016). All the above studies extracted pest and disease features by manual design.
Deep Learning-Based Classification of Image Data Sets Containing 111 Different Seeds
2023, Advanced Theory and Simulations