Application of Digital Image Processing Techniques for Agriculture: A Review

Abstract

Agriculture plays a crucial role in human survival, necessitating the development of efficient methods for food production. This chapter reviews Digital Image Processing (DPI) methods that utilize various color models to segment elements like leaves, fruits, pests, and diseases, aiming to enhance agricultural crop production. Recent DPI research employs techniques such as image subtraction, binarization, color thresholding, statistics, and convolutional filtering to segment and identify crop elements with shared attributes. DPI algorithms have a broad impact on optimizing resources for increased food production through agriculture. This chapter provides an overview of DPI techniques and their applications in agricultural image segmentation, including methods for detecting fruit quality, pests, and plant nutritional status. The review’s contribution lies in the selection and analysis of highly cited articles, offering readers a current perspective on DPI’s application in agricultural processes.

1. Introduction

It is estimated that the human population will be 10 billion by 2050, this poses a challenge in the production of food for their livelihoods, particularly in agriculture. From 2010 to 2020, the cultivated area grew by 2,424,000 ha, and the use of fertilizers by 23,718,607 ha according to data from the Food and Agriculture Organization of the United Nations (FAO) [1].

The sense of sight is the principal source of information for humans. In agriculture, the observation of crops determines the state in which they are. Fatigue during a day’s work, inexperience, and a lack of knowledge of the crop can influence the farmer’s perception when making a judgment based on sight about the nutritional state of the crops, the presence of pests, and the quality of the fruit [2].

Science and technology support agricultural processes by increasing their efficiency [3, 4, 5, 6], and among them are the research areas of computer vision (CV) [7, 8, 9] and artificial intelligence [10, 11, 12] (AI). Computer vision systems are composed of three elements: light sources, image acquisition devices, and DPI algorithms [6, 13].

There are two types of illumination sources, natural and artificial, each with particular characteristics. The natural sources are usually taken with indoor or outdoor sunlight on cloudy or sunny days [14, 15]. Artificial light sources can be incandescent, fluorescent, and LEDS [16], among others. In this type of lighting, you can control more variants, such as the angle of incidence and light power.

There are a wide variety of options to choose from the capture devices, depending on the features they have and the objective to be achieved. There are cameras with particular specifications, such as hyperspectral [17], for professional photography, and the most commonly used peripherals are scanners [18], mobile devices, and wireless and security cameras [19, 20, 21].

In order to increase control over the variables on which they are affected, metallic structures are developed in which the illumination sources and image capture devices are fixed; this allows experiments in laboratory environments.

The purpose of DPI algorithms is to analyze and process the information shown in images [22]. It is important to mention that an image is represented as a two-dimensional function f(x,y), where x and y are spatial coordinates, and the amplitude of f at any pair of coordinates (x, y) is called the gray level of the image at that point. When x, y, and the intensity values are all finite, discrete quantities, we call the image a digital image [22]. There are several types of DPI algorithms, such as histogram operations, pixel neighborhood-base operations, morphological operations, binarization, edge detectors, Fourier methods, and other different classes of image transforms.

The DPI systems applied to agriculture were initially implemented in laboratory environments, food inspection lines, and real working conditions. The extension of their field of application is due to several factors, including the maturity reached by DPI algorithms, increased computational power, and the need for agriculture to become more efficient and sustainable from an environmental point of view.

In this review, an analysis of digital image processing methods applied to agriculture in recent years and reported in papers is made; these are emphasized in aspects such as the detection of pests and diseases (1), nutritional deficiencies (2), and inspection of fruit quality (3). The papers were selected based on the number of references that appear in Google Academic.

2. Digital image processing techniques applied to agriculture

2.1 Pest detection

Pests have always been present in nature. In modern agriculture, they are undesirable, as they affect the development and production of crops [23, 24, 25]. Because there is a wide variety of crops whose fruits are used for food, in the same way, the number of pests and diseases is equally diverse, which makes it difficult to detect, identify, and correct them in time [9, 26], especially if such detection is carried out exclusively by the farmer’s experience. Among the most common pests, there are different types of insects, such as white flies, aphids, snails, and others.

Table 1 lists the papers on pest identification using DPI, which aims at the timely, effective, and non-invasive detection of pests.

Table of DPI algorithms applied to pest detection
DPI Topic	DPI Algorithm	Relevant aspects
Frequency Domain Processing and Image Transforms	Fourier [27, 28]; Wavelt [23, 29]	Nansen et al. [28] Automatic detection method of caterpillar pests in mustard and broad bean crops, achieved results greater than 90% in detection.; Asfarian et al. [27] Used Fourier transform spectra to carry out the detection of the most common pests in rice crops in Indonesia. The algorithms used show an accuracy of 83%.; Luo et al. [29] Estimated by spectroscopy and Wavelt transform the number of aphids. Mentioned that the model developed with 5 characteristics is better than with 6 in wheat crop.
Spatial Domain Processing	Histogram transformation [30, 31, 32]; Smoothening filter [7, 33, 34]; AdaBoost [35]; Saliency maps [36]; Sobel Operators [26, 31]	Thenmozhi and Reddy [26] Detect and identify the shapes of insects in sugarcane crops. Use of MATLAB API.; García et al. [31] Recognition of Lobesia botrana in vineyards by DPI and clustering, with 95% effectiveness.; Rupesh and Mundada [33] Detection of aphids and whiteflies in greenhouse crops.; Singh and Misra [7] Identifying and counting aphids on soybean plants grown in greenhouses.; Yao et al. [35] Develop an image capture device for counting planthoppers in rice fields. Reported average of detection of 85.2%; Nanni et al. [36] Automatic classifier insect pest in tea plants based on the fusion between Saliency methods. Uses a CNN as a classifier.
Image Segmentation	RDI algorithm [37]; Image subtraction [38]; Threshold [7, 17, 30, 32, 34, 35, 39, 40, 41]; K-Means classifier [9, 31]; K-nearest neighbor classifier [42]; Convolutional Neural Network [36, 43, 44]	Liu and Chahl [17] Detecting invertebrate insects on green leaves of several crops. It works in light spectra such as UV and NIR, allowing it to detect camouflaged pests.; Singh and Misra [7] Identifying and counting aphids on soybean plants. Results are presented with several classifiers.; Miranda et al. [34] Estimate the density of pest infestation in rice fields. Use of wireless technologies for image capture.; Huddar et al. [37] Segment and identify whitefly in unstructured environments in various crops. Reports 96% with metric accuracy.; Maharlooei et al. [38] Detect and count different-sized soybean aphids. Emphasizes the importance of lighting conditions, tries out different cameras.; Nanehkaran et al. [40] Detection of diseases on leaves of plants of cucumber corn and rice. Presents time comparisons with different segmentation methods, qualitative comparisons of results.; Barbedo [41] It uses classical DPI methods on images of soybean crops which makes it easy to replicate them in different contexts. Displays tables with the results obtained.; Hossain et al. [42] Proposes a method for the detection and classification of leaf diseases. Generates results to detect and recognize selected diseases with an accuracy metrics of 96.76%.; Sladojevic et al. [43] Developed convolutional neural networks for disease detection on leaves of different plants, obtaining a metric accuracy of 96% with the ImageNet and CaffeNet models.
Morphological Operators	Close operator [30]; Dilation operator [26, 37, 39, 45, 46]; Erosion operator [37, 39, 46]	Luo et al. [30] Automatically detect and classify weed in potato crops. Uses various color models, combines different AI techniques, and high efficiency.; Chang and Lin [39] Detect and identify weeds in lettuce crops. Using fuzzy logic as a classifier to classify watering. Identification of lettuce and herbs is performed in real-time.; Sabouri and Sajadi [45] Evaluated the relationship between leaf area and solar radiation in various crops.

Table 1.

Papers on pests detection applying digital image processing methods.

2.2 Nutrition deficiencies detection

Crops need a series of nutrients that they naturally take from the air, water, and soil for their development. The elements that plants feed on are classified into macronutrients and micronutrients, depending on the amount required by crops. Macronutrients Nitrogen (N), Phosphorus (P), and Potassium (K) are important for modern agriculture, as they are usually supplied to crops in different fertilization schemes [18, 47, 48].

Early detection of NPK excesses or deficiencies is important; DPI and AI algorithms represent an effective, non-invasive, and low-cost alternative for their detection.

Table 2 lists a selection of papers that make use of DPI algorithms to detect macronutrient deficiencies in different crops.

Table of DPI algorithms applied to nutrition deficit detection
DPI Topic	DPI Algorithm	Relevant aspects
Frequency Domain Processing and Image Transforms	Fourier [49, 50, 51]; Wavelt [52]	Nejati et al. [49] Proposed a method that uses the fast Fourier transform and leaf edges to classify between crop leaves and weed leaves in cornfields, obtained 92% accuracy metrics.; da Silva Leite et al. [50] Characterized nitrogen uptake by fractions of plants, with different fertilization schedules, by means of Fourier Transform Attenuated Total Reflectance Analysis on leaves of Physalis peruviana and Physalis angulata.; da Silva et al. [52] Measure the amount of Magnesium in corn leaves by transforms using the walvet transform in combination with other DPI algorithms. Reports an efficiency of 75%.; Ray [51] Using the Fourier Transform to identify cell walls in leaves.
Spatial Domain Processing	Histogram transformation [51, 53, 54, 55, 56]; Smoothening filter [57, 58, 59]; Image subtraction [55]; Color Index [53, 60, 61, 62, 63, 64, 65, 66, 67, 68]; Canny Operator [58]; SIFT [68]; Convolutional Neural Network [44, 69]	Qi et al. [53] Relating foliar Nitrogen content to wheat crop development. Uses neural networks as classifier and displays statistical performance data.; Larijani et al. [54] The timely and rapid diagnosis of rice blast based on the image processing technique in field conditions. The overall accuracy metrics was 94; Dhingra et al. [58] Determined the health status of the leaves by means of a neutrosophic technique in basil leaves, 3 groups are formed: healthy, unhealthy, and undetermined, reaching a classification accuracy metric of 98.4%.; Wang et al. [60] Relate the intensity of the green color of the rice leaves with the amount of Nitrogen used to fertilize. Presents an analysis of the biomass of rice in relation to its green color with the use of Nitrogen.; Wang et al. [64] Relationship between image color indexes and Nitrogen status in rice crops under natural light. Presents graphs contrasting green color versus Nitrogen and chlorophyll content. Lee and Lee [65] Construct a non-destructive method for monitoring rice crop growth and Nitrogen nutrition status with digital camera image analysis. Presents the relationship between Canopy cover and green color to determine the nutritional status. Vesali et al. [63] Develop an application for mobile devices for the estimation of chlorophyll in corn leaves using DPI algorithms. Use a commercial device to measure the amount of chlorophyll present in the leaves and validate the results obtained. Buchaillot et al. [66] Evaluate the adaptation of maize crops to limiting abiotic and biotic conditions using RGB images. Shows tables with data on the growth of corn varieties, showing their adaptation to growing conditions.; Jia et al. [68] Detection of diseases and pests, monitoring of rice crop growth. Use the SIFT method to detect points of interest in the image mosaic.; Kamelia et al. [58] Detecting plant nutrient deficits using a combination of DPI algorithms.; Ponce et al. [69] Detected low nutrition in tomato crops by implementing the CNN + AHN model testing various configurations, delivering an accuracy metrics of 95; Fuentes et al. [47] Determined the most suitable architecture for pest and disease detection in tomato crops. Faster R-CNN, Region-based Fully Convolutional Network models were tested with average accuracy metrics of 82% and 85%, respectively.
Image Segmentation	Threshold [18, 19, 57, 58, 60, 61, 62, 64, 65, 70, 71, 72]; Binarization [62, 73, 74]; Otsu [54]; K-Means [54, 57]	Sun et al. [18] Determine the changes in the characteristics of rice leaves under nutritional stress. Easy segmentation due to capture conditions.; Story et al. [19] Detecting calcium deficiency and growth in greenhouse lettuces. Presents a simple segmentation, it is based on methods proven by other authors.; Li et al. [70] Establishes the relationship between reflectivity and Nitrogen content in cereals. Report a mean squared error of less than 16% compared to other methods.; Rorie et al. [71] Evaluating the Nitrogen content of maize in greenhouses by DGCI, used a simple segmentation. Makes use of color indexes to associate them with the amount of Nitrogen in the leaves.; Chen et al. [73] Identifying NPK deficiencies in rice based on static scanning technology and hierarchical identification.; Mao et al. [61] Develop a non-destructive method for Nitrogen estimation in lettuce using spectroscopy and DPI. Adopted different classifiers.; Sun et al. [62] To analyze the temporal dynamics of the morphology and color of rice to evaluate their efficiency in identifying deficiencies of NPK. Makes use of morphological indexes for the identification of deficiencies in partially and fully developed leaves.; Chen et al. [74] Identify Phosphorus stress in four stages of rice development. Using ANOVA to find differences in fertilization schedules.; Aditini and Gupta [57] Generate a method for the identification of macronutrient deficits by means of DPI algorithms.
Morphological Operators	Close operator [59]; Open operator [59]; Erosion operator [42, 59]; Dilatation operator [42]	Su et al. [59] Implemented a set of DPI algorithms for automatic location of soybean plants. Obtain results of 97% accuracy metrics.

Table 2.

Papers on nutrition deficiencies detection applying digital image processing methods.

2.3 Inspection of quality detection

The final consumer of agricultural products seeks the highest quality in the items they will consume. It is important to mention that in order for farm products to reach their final destination, they have already passed through different filters for their classification, which is carried out by experts trained to detect characteristics such as color, shape, size, and weight, among others [75, 76]. The classification made by human beings is subject to conditions such as their state of health, mood, and fatigue, which can affect their judgment, and this can generate errors in the classification of the inspected products [77]. Researchers around the world have presented methods that support the task of determining attributes related to the quality of field products.

Table 3 shows a series of papers that make proposals based on DPI techniques for determining the quality of inspected field products.

Table of DPI algorithms applied to quality detection
DPI Topic	DPI Algorithm	Relevant aspects
Frequency Domain Processing and Image Transforms	Fourier [78, 79, 80]; Fast radial symmetry [81]; Wavelet [82]	Ambrose et al. [78] Fourier transform near infrared spectroscopy and Raman spectroscopy techniques were used to evaluate the viability of corn seeds. It classifies seeds into viable and non-viable with an accuracy metrics of 100% and prediction of 95%, respectively.; Bansal et al. [79] Proposed a technique for immature citrus from an outdoor color image and counting the number of fruits. It obtained results in correct detection of 82.2%.; Xu and Katchova [80] Predicted soybean production using FFT and NDVI.; Pérez-Zavala et al. [81] Recognition and detection of grape bunches. Shown a 96% accuracy metric.; Dhakshina Kumar et al. [82] Proposed system performs classifications of tomatoes in 3 stages. Reports greater than 90% accuracy metrics.
Spatial Domain Processing	Smoothening filter [20, 77, 82, 83, 84, 85, 86]; White Path [76]; Decorrelation Stretch [87]; Histogram equalization [81, 84]; Image Subtraction [20]; Sobel Operators [88]; Color Index [80]; SIFT [87, 89]	Nyalala et al. [20] Calculate tomato mass and volume using two and three-dimensional imaging, use a Kinect for image acquisition. Reports 97% in accuracy metrics.; Sadeghi-Tehran et al. [87] Detect 3 stages of flowering and ear growth of wheat crops. Reports 95.21%, 97.29%, and 99.59% accuracy metrics of the 3 spike stages and 85.45% in the flowering.; Ireri et al. [88] System for tomato grading based on RGB images. Shows an efficiency higher than 97% in tomato grading. Use ANN as classifier.; Castro et al. [83] The combinations of machine learning techniques and three color spaces were evaluated to classify Cape gooseberry. Performs tests in controlled environments. The classifier with the best result was SVM with 70%.; Zhu et al. [89] Identifying the ear in wheat plants during its development. Does not report a quantitative metric, defines its efficiency as acceptable.; Wan et al. [86] Estimate tomato maturity at 3 levels, green, yellow, and red in roma and pear varieties. Reports 99.31% in accuracy metrics of the 3 maturity levels. Uses an ANN.
Image Segmentation	Binarization [77, 82, 90]; Otsu [20, 76, 88]; Threshold [15, 20, 83, 86, 88, 89, 90, 91]; Contour tracking [86]; K-Means [14, 92]; Convolutional Neural Network [93, 94, 95]	Fan et al. [14] Segmenting apples in unstructured light conditions. Shown efficiency of 99.26% of accuracy metrics with various types of classifiers.; Rico-Fernández et al. [15] Segmenting foliage of different crops. Reports 89% efficiency in 3 different crops.; Arakeri and Lakshmana [77] Estimate tomato quality by detecting maturity and defects present in the fruit. Reports an 96.47% of accuracy metricsin identifying fruit maturity. Use a multilayer neural network for recognition.; Villaseñor-Aguilar et al. [90] Estimating the maturity of the morron pepper. Report efficiency of 100% and 88% with two different classifiers. Uses fuzzy system and ANN as classifiers.; Garcia et al. [91] Development of an automatic tomato maturity identification system. Obtains an accuracy metrics of 84%. Uses simple segregation by thresholds against a controlled background.; Roy et al. [93] Work in classification of apples by means of the UNet and a proprietary model called En-UNet, obtaining in the IoU metric values above 90%. Majeed et al. [95] Segmented grape crops to determine the position and orientation of grape plant cordons using several CNN models such as SegNet-VGG16 obtaining up to 80% efficiency in determining the trajectory of the cordons.
Morphological Operators	Dilation operator [76, 84, 86] Erosion operator [76, 84] Closing operator [89]	Portugal-Zambrano et al. [76] Evaluating physical defects in green coffee beans. Reports an average of 98.48% accuracy metrics identification.

Table 3.

Papers on Inspection of quality detection applying digital image processing methods.

3. DPI techniques analysis

The literature documents different DPI techniques used for different purposes in fields ranging from agriculture to industrial applications [9, 96, 97, 98]. Tables 1–3 are grouped papers according to the DPI techniques used; such classification is based on the following categories: filtering in the frequency domain, intensity and transformations spatial filtering, convolutions, image segmentation, morphological operators, and color image processing [22].

3.1 Frequency domain processing and image transforms

The concept of a transform introduces an alternative mathematical representation of an image, deviating from the conventional depiction based on two spatial variables. These tools wield substantial power, finding broad applications beyond digital image processing.

The integration of the Fourier transform has revolutionized digital image processing, offering an alternative mathematical representation that facilitates in-depth analysis in the frequency domain. This technique decomposes images into frequency components, revealing intricate patterns that are not always evident in the spatial domain. Fourier transform enhances image quality through frequency domain filtering, eliminating noise, improving contrast, and highlighting specific features. Its efficiency in compression ensures streamlined transmission and storage.

Asfarian [27] and Nasen [28] used the Fourier transformation for the purpose of detecting pests or diseases. The first one employed RGB images for the detection of caterpillars, and the second one used hyperspectral images for the detection of pesticide residues in corn. The application of frequency domain processing for the detection of nutrient deficiencies in crops, for example, in the identification of magnesium deficiency [52] or nitrogen utilized some variants, such as the application of filters [50]. An application and variant of this type of algorithm is the Fast Fourier transform, which was implemented to determine the viability of corn seeds [78] or for the detection of citrus fruits [79]. It can also be combined with the color index to estimate soybean yields [80].

Wavelet transforms and other image transforms excel in simultaneous frequency and spatial analysis. Unlike Fourier transforms, they offer localized assessments, proving effective for diverse image structures. Wavelets ensure efficient compression, preserving essential features while reducing the size of data. Their adaptability extends to applications like denoising and compression, with variants like discrete wavelet transform (DWT) and continuous wavelet transform (CWT) tailoring analytical approaches. Nagar [23] used the Wavelt transformation for the detection of caterpillars on crops in India, highlighting features such as ORB key points that can be extracted or utilized to estimate the density of aphids using spectroscopy and continuous wavelet analysis, as did Luo [29].

In agriculture, frequency domain processing and image transforms play an essential role. Fourier transforms reveal frequency patterns indicating variations in crop health or the presence of pests. Wavelet transforms offer advantages in capturing spatial and frequency information. These tools are fundamental for early disease detection, crop management, and optimizing agricultural practices, driving precise and sustainable decision-making.

3.2 Spatial domain processing

Spatial domain processing is a foundational aspect of image processing, distinct from transform domain processing. It involves directly manipulating pixel values in an image, allowing a nuanced understanding of intensity transformations. Essential techniques in spatial domain processing include intensity transformations and the manipulation of image histograms for enhancement. The mechanics of spatial filtering, encompassing the formation of filters, spatial convolution, and correlation, play an important role. Understanding various spatial filters, their applications, and the relationships between them, especially the fundamental role of lowpass filters, is integral. Additionally, awareness of combining enhancement methods is vital for addressing cases where a singular approach proves insufficient, highlighting the practical importance of spatial domain processing in refining and optimizing digital images. Spatial-domain image processing in color extends its significance to detailed imagery. This approach involves directly manipulating pixel values within the image space, offering nuanced adjustments for each color channel. From improving color contrast to refining object boundaries, spatial domain processing plays a crucial role in optimizing the visual appeal of color images. Techniques like spatial filtering and histogram manipulation have become powerful tools for tasks like color correction and feature enhancement. In the realm of color image processing, spatial domain techniques provide essential means for tailoring and refining the visual characteristics of intricate images. Spatial domain processing is pivotal in agriculture, driving precision and efficiency in crop management. By directly manipulating pixel values, it enables nuanced analysis of agricultural images. Understanding edge characteristics, employing spatial filtering for edge detection, and utilizing region-based segmentation refine identifications of crops, weeds, and soil conditions. Threshold and spatial filtering synergies enhance segmentation accuracy, while techniques like graph cuts and morphological watersheds further aid in analyzing complex agricultural landscapes.

Spatial domain processing algorithms are among the most widely used in agriculture. Image histogram manipulation is performed to improve brightness and contrast conditions mainly [30, 32, 76], and an adaptation of the histogram equalization is carried out to calculate a histogram of a given region by means of the K-means algorithm [31] for Lobesia botrana detection or in the grouping of bunches of grapes [81]. Other applications are used to adapt histograms to a certain pattern that is considered good for a certain purpose [84]. In the detection of nutrients such as nitrogen present in the leaves of different crops, histogram processing is applied, for example, by Qi [53]. Larijani [54] studied the use of histograms for the determination of thresholds, with the objective of removing background from rice crops. The calculation of histograms in images is a relatively simple process, and its contribution to DPI applications should not be underestimated.

The application of filters in the spatial domain is classified into low-pass and high-pass filters. Among the first ones, some of the most used are the mean, median, mode, and Gaussian filters, aimed at eliminating noise and smoothing the images. On the other hand, high-pass filters are used to detect sharp changes in contrasts, for example, edges. Among these types of filters are the Roberts, Sobel, and Prewitt operators. Arakeri [77] employed a median filter for noise removal following a segmentation process. Conversely, Rupesh [33] and Miranda [34] implemented low-pass filtering using mean, median, and mode with 3x3 and 9x9 kernels. This is done to eliminate “salt and pepper” noise, enabling the subsequent processes of pest detection and classification. Singh [7] utilized the media filter to remove unwanted elements in crops of beans, lemons, roses, and bananas with the intention of detecting pests and diseases. Edge detection is the counterpart of smoothing filters, which are used in order to highlight the borders of objects in the image. Thenmozhi [26] applied the Sobel filter for sugarcane pest detection, utilizing the absolute gradient. Garcia [31] uses the same filter to detect Lobesia botrana outline in Lobesia botrana trap images. Ireri [88] searched to extract the characteristics that describe the quality of a tomato by applying a Sobel operator. The detection of nutrient deficits is fundamental for modern agriculture. Kamelia [58] developed a method for the detection of macronutrients using a Canny detector for this purpose. Xu [56] applied a Roberts operator to extract characteristics; in his study, it showed better results than other operators used to detect nitrogen and potassium deficiencies.

The normalization of the color bands of the different models, mainly in RGB, allows to obtain a proportion of each band in a given pixel, in agriculture, this is used to relate a color to the nutritional status of the plant.

Wang [60] implemented NRI, NGI, and GMR to estimate the amount of nitrogen in rice crops, as well as complementing a separate threshold for the foliage background. Sun [62] provided and applied formulas to calculate different indexes such as dark green, excess green vegetation, excess green minus excess red, green-red vegetation, and Kawashima, which are basic elements to monitor NPK absorption. Another application of color indexes is to measure chlorophyll in crops. Vesali [63] utilized these indexes in a mobile application, where images are captured by coloring the corn leaf directly on the camera lens. Indexes such as GMR, GDR, and DGCI are among those employed by Vesali. Additionally, Lee [65] used color indexes to estimate rice crop growth by measuring leaf development. While color indexes are widely used for detecting nutritional deficiencies in crops, they can be used for other purposes as well. Xu [80] applied NDVI in conjunction with the Fourier transform to predict soybean yield in the United States.

Spatial domain processing, fundamental for decision-making, optimizes resource allocation, early disease detection, and sustainable agricultural practices, underscoring its significance in modern precision agriculture.

3.3 Image segmentation

Image segmentation is a foundational process in diverse applications, involving the partitioning of images for meaningful interpretation. It encompasses edge detection through spatial filtering, diverse thresholding approaches, and region-based segmentation. In agriculture, segmentation is pivotal for precision farming, distinguishing crops, weeds, and soil conditions. Spatial filtering and advanced edge detection refine these identifications. Thresholding becomes vital for discerning agricultural features, and the fusion of thresholding and spatial filtering enhances segmentation precision. Techniques like region-based segmentation, graph cuts, morphological watersheds, and motion-based segmentation prove invaluable for detailed crop and soil analysis.

One of the most used algorithms in image segmentation is bianrization, which can be carried out in several ways: the simplest ones, like thresholding, and others with more statistical bases, like OTSU. Within the segmentation by means of thresholds, the simplest way is to control the main variables such as illumination and background [19, 38, 86], which allows to define the threshold easily. Nanehkaran [40] implemented the use of thresholds in Lab color spaces; HSI in the first one used OTSU to divide the image into two classes, and HSI used a threshold to segment the green area in rice, corn, and cucumber crops. Larijani [54] used OTSU to perform an automatic threshold histogram of images of rice based on shape or to reduce the gray level in binary images. Finally, the binarization process can be applied to multispectral images. Liu [17] did it for invertebrate detection by manually setting the threshold for gamma. Barbedo [41] established trial-and-error thresholds in CMYK color space to achieve white fly segmentation and subsequent counting. Zhu [89] developed adaptive thresholds with RGB images to minimize noise and observe the wheat-gleaning phase.

Clustering algorithms are of importance for DPI, as they have the ability to label pixels into more than two groups. The k-means algorithm adds an element to the set to which its representative is closest. Fan [14] presented an apple segmentation algorithm in natural environments based on the K-means algorithm that uses blocks of pixels to determine the regions of the apple from the rest of the image. Garcia [31] performed a K-means algorithm for segmenting Lobesia botrana using decryptors in clustering. Sometimes a way to segment images is to have a base image and subsequent images. A pixel-by-pixel subtraction is carried out, and with this, it is possible to identify elements that are not in the base image, Maharlooei [38] used this algorithm with Sobeyian cultures for the detection of aphids. Perez-Zavala [81] used Density-based spatial clustering of applications with noise filters (DBSCAN) to reduce noise in images of grapes.

One of the algorithms of relatively recent emergence is deep learning, in particular, what is referred to as convolutional neural networks, which are based on an operation called convolution that uses a kernel as a base. Sladojevic [43] implemented a CNN model called CaffeNet that allows for the classification of diseases in various types of crops. Fuentes [44] explored the efficacy of various models of CNN Faster, R-CNN, R-FCN, VGG Net, and ResNet proposed by other authors in the detection of diseases and pests in tomato crops. The CNN has been tested to estimate the quality of fruits of various crops. Roy [93] proposed to perform a semantic segmentation and classification of apples by means of the UNet and a proprietary model called En-UNet, obtaining IoU metric values above 90%. Majeeda [95] processed images with grape crops to determine the position and orientation of grape plant cordons using the SegNet-VGG16, SegNet-VGG19, and FCN-AlexNet models, obtaining up to 80% efficiency in determining the trajectory of the cordons.

Image segmentation is integral for optimizing decision-making in agriculture, and fostering sustainability and efficiency in farming practices.

3.4 Morphological operators

Morphological operations in image processing are essential techniques for shape manipulation. By applying a structuring element to an input image, these operations generate an output image with identical dimensions. Initially designed for binary images, they have been adapted for grayscale applications. Key operations include dilation, expanding object boundaries, and erosion, contracting them. Closing, a two-step process of dilation followed by erosion, effectively closes gaps in objects. Opening, the complementary operation, involves initial erosion followed by dilation. These operations, fundamental for tasks like object detection and image enhancement, maintain original dimensions and contribute significantly to diverse fields, from medical imaging to computer vision, showcasing their importance as foundational tools in image processing. Morphological operations in image processing are pivotal for advancing precision agriculture. These techniques, such as dilation and erosion, refine crop analysis, weed detection, and soil assessment. The morphological operators play a crucial role in closing gaps, enhancing the identification of crops, and contributing to more accurate agricultural land evaluations. Portugal-Zambrano [76] carries out a dilation algorithm on images of trays with coffee beans to fill in any gaps that may exist in the contour, and then uses a dilation operator to reduce the thickness of these gaps in order to evaluate the quality of the coffee beans. Hamuda [84] proposed an automatic cauliflower crop detector using an erosion operator with a 3 x 3 structuring element to remove the remaining isolated white pixels from the binary image. Dilation with a 7 x 7 structuring element to recover the shape of the remaining objects. To estimate the freshness of tomato crops, Wan [86] applied erosion and expansion operators to improve fruit edges. In the detection of pests, the morphological operators have importance in the fact that it is necessary to identify perfectly where one insect or pest starts from another. Rajan [46] employed an erosion operator to eliminate elements that were not of interest and then implemented a dilation operator to close the outlines of whiteflies. Chang [39] used morphological operators applied to an image from a thresholding process to enclose the shape of the crop in a mask that shows its location within the image.

By optimizing decision-making, these operations contribute to sustainable farming practices, offering valuable tools for improved resource allocation and enhanced crop management in agriculture.

4. Conclusions

The algorithms discussed in the analyzed papers offer a glimpse into the future of the agricultural industry, indicating the growing presence and impact of various technologies. Computational algorithms, particularly those involving DPI and AI, have shown remarkable advancements in influencing the cultivation of diverse food types. Notably, prevalent DPI methods encompass color space transformation, binarization processes, morphological operators, edge detection, threshold segmentation, low-pass filters, and other techniques operating in the frequency domain.

In the analyzed papers, one notable aspect is the efficiency reflected in the results of the presented methods and algorithms. The most frequently employed metric was accuracy, consistently surpassing 90% and, in certain instances, achieving 100%. The high accuracy values underscore the significance and effectiveness of employing algorithms in agricultural processes, thereby supporting their development and the attainment of objectives.

Three elements influence the efficiency mentioned in the objectives of the different papers, such as the image acquisition conditions to work with, the features that can be extracted, and the classification algorithms used.

Image acquisition marks the inception of any DPI system. In the reviewed studies, images were captured under controlled conditions, encompassing controlled illumination sources, carefully chosen backgrounds to enhance contrast, and a specific arrangement of leaves and fruits—typically within a laboratory setting. Conversely, another category of image acquisition involved environments with less stringent controls, particularly concerning light and background; these images were directly captured in crop fields.

In addition to these features that are extracted from the images by applying the DPI algorithms, which were mentioned in the reviewed papers, there are those related to the color, shape of objects, and texture of these.

Once the images are processed and the features are defined and extracted, the subsequent step involves employing a classifier to ascertain the class to which the analyzed object belongs. Various classifier types exist; those outlined in the analyzed literature include minimum distances, SVM, simple and convolutional RNN, as well as linear regressions.

A challenge confronting the proposed methods lies in transitioning from the laboratory environment to the field, where the diverse algorithms need to contribute value to the agricultural production chain. Researchers are banking on technological advancements, particularly in terms of the mobility and accessibility of sensing devices, along with the natural increase in computing power, to effect changes in the field environment.

In conclusion, the utilization of DPI algorithms, is gaining prominence across diverse agricultural tasks and crops. The objective is to enhance productivity while concurrently reducing environmental impact, thereby ensuring that agriculture continues to serve as a fundamental pillar in the development of modern society.