Digital Innovations in Agriculture

1. Introduction

Digital agriculture, defined as the analysis and collection of various farm data, is constantly evolving. Data collected on an ongoing basis from fields, machinery, weather stations, sensors, and systems are used to perform a wide range of tasks and to make optimal decisions in the running of agricultural production. Platforms or management software for agriculture have tremendous potential. Digital agriculture relies heavily on detailed image analysis, artificial neural networks, machine learning, the Internet of Things (IoT), and big data [1,2]. The aforementioned digital agriculture techniques can be successfully used for qualitative assessments of agricultural crops, diagnoses of plant diseases, yield predictions, classification issues, and intelligent weed controls [3,4,5,6,7]. With the rapid developments of precision farming and digital agriculture, more and more farms are turning to tools based on artificial intelligence. The purpose of this Special Issue was to publish high-quality research and review papers that cover issues related to digital innovations in agriculture. Below, we have included highlights from all of the articles published in this Special Issue.

2. Papers in This Special Issue

In the research described in the first article [8], the authors attempted to find correlations between several selected neural network models and statistical methods commonly used in agriculture. The comparison has a universal dimension—it applies to crop production, livestock production, and the quality of the natural environment. The authors emphasize that artificial neural networks are a convenient, fast, and accurate tool; therefore, they have become very popular. The authors recommend their use in preparing analyses for agriculture.

In the second article [9] , the authors performed a review of the current state of research on the use of artificial intelligence, the Internet of Things, and hyperspectral imaging for crop disease detection. In addition, they compared several different techniques for diagnosing disease symptoms in plants. Convolutional neural network models proved to be the most efficient and effective methods for locating visual patterns in images.

In the third article [10], the authors systematically analysed climate services for agriculture in Africa. They reviewed 50 literature items from 20 African countries to develop such services. It was found that the development of such services is still in the early stages, as innovations in mobile telephony and Internet services integrated with climate services are presently undergoing a trial stage. The article confirms the need to integrate indigenous and scientific knowledge systems in creating climate information in Africa.

In the fourth review article [11], the authors managed the discussion of Findable, Accessible, Interoperable, and Reusable (FAIR) data in sustainable agriculture. With digital agriculture generating more and more data, there is a need to create systems for collecting reliable information. In conclusion, the authors proposed a method for identifying the main problems in making FAIR data available.

The fifth comprehensive review article [12] is an interesting summary of the relationships between sustainable agriculture and the Internet of Things (IoT). The authors hope to involve robots, cloud computing, and artificial intelligence in agricultural production even more than has been done before. In addition, the paper evaluates selected tools and equipment used in wireless sensor applications in IoT agriculture. The challenges of combining the technology with conventional agricultural operations were pointed out.

In the sixth paper [13], the authors attempted to determine whether milk flow characteristics can be considered biomarkers of lameness incidence in cows. A study was conducted on more than 100 head of dairy cows, and both cortisol concentrations in the animals’ blood and milking characteristics were evaluated. The conducted experiment confirmed the thesis that milk flow characteristics can act as biomarkers of lameness in dairy cows.

The seventh article [14] deals with the simulation of fuel consumption depending on the load level of the tractor engine. A 95 kW engine with a partial power shift transmission was tested using the PTO dyno test method. It was observed that the engine load and fuel consumption were directly proportional to the engine load levels. Statistical analyses of the results indicated an exponential relationship between fuel consumption and engine load levels. The published research provides an opportunity to design an innovative agricultural tractor with a higher fuel efficiency.

In the eighth article [15], the authors made predictions regarding the reproductive success of multiparous dairy cows among a primiparous population, based on selected parameters generated by sensors. They tested an automatic milking system during the pregnancy of multiparous dairy cows and evaluated the system’s accuracy based on blood parameters. In the group of non-pregnant cows, more samples with elevated cortisol were observed. Other interesting correlations indicated the risk of mastitis or oxidative stress in cows, depending on the presence or absence of pregnancy.

The ninth article [16] dealt with image analysis techniques for the optimization of a peanut planting space in Hainan. The study’s authors used PlanetScope images with a spatial resolution of 3 m. The research was based on the construction of three models for peanut planting area extraction, based on the support vector machine (SVM), BP neural network (BPNN) and random forest (RF) classification algorithms. The results confirmed the effectiveness of PlanetScope imagery in solving agricultural problems. The RF model successfully optimized the planting of peanuts.

In the tenth article [17], the authors used deep learning neural networks, specifically the LC-DenseFCN point surveillance algorithm, to count chickens in a closed-circuit camera environment. Compared to conventional counting and other popular techniques, the proposed optimized method offers the potential for the fast and accurate counting of chickens residing at high densities.

The eleventh article [18] concerned on the identification of risk factors for detecting lameness in cows using natural biosensors. Before the study, the authors hypothesized that the formation of inline biomarkers would depend on the lameness of dairy cows in early lactations. Among other things, the study indicated that low milk fat content was maintained from before the onset of the disease until the very day of its confirmation.

In the twelfth article [19], the authors presented a study to assess the impact of lameness on the feeding attributes of dairy cows. The study used a sensor-nose band that a dairy cow wore just after calving. They found that lameness affected cows’ feeding preferences and changes in biomarkers. For example, the lowest eating time was found on the day of diagnosis, and the highest on the ninth day before lameness was detected.

The thirteenth article [20] discusses developing a lettuce growth model using U-Net. Arabidopsis plants were used as the model-training material. The DL model developed for Arabidopsis works well for modelling the growth of other target species.

In the fourteenth article [21], the authors developed a predictive yield model for rice. Phenological data and hyperspectral imaging data were used to build the model. The final model was based on the values of vegetation indices: NDVI, EVI, SAVI, and REP. The generated model was characterized by a high accuracy (R² = 0.84). In addition, the optimal time intervals for predicting rice yields were identified. The results show the high usefulness of data from late vegetative and flowering stages.

In the fifteenth article [22], the authors dealt with the individual identification of cattle faces using neural networks, a deep learning method. The activity involved an image analysis technique and modern RetinaNet technology. The selected algorithm allowed for a very high identification precision of up to 99.8%, and a very short average processing time of 0.0438 s per image.

In the sixteenth article [23], the authors determined the management zones within a field using remote sensing for variable rates of nitrogen fertilization of wheat. Simplified and hybrid models were used, and machine learning was also involved. The methodology for creating the models was enriched with information on phenological phases and the occurrence of agricultural droughts. The results showed that agronomic and climatic information allow for improving and optimizing the designation of management zones.

The seventeenth article [24] addresses an innovative method for monitoring soil nutrients using hyperspectral remote sensing. Linear and nonlinear algorithms were used to perform the task, including machine learning, LASSO, and GBDT algorithms. The results showed that LASSO and GBDT algorithms can improve the quality of TN, TP, and TK soil content estimation. This has great application significance in agriculture.

In the eighteenth article [25], the authors used the method of multiple regression and artificial neural networks to model the essential oil content and yield of trans-anethole obtained from fennel. The study aimed to identify the most accurate predictive tool. The results showed that artificial neural networks made more accurate predictions than regression methods. The published research may be useful for breeders of plants in the Apiaceae family to model the various complex polygenic traits of crop plants, which is important from the points of view of industry, herbal uses, pharmaceuticals, etc.

The nineteenth article [26] uses a four-stage image processing algorithm to identify and count Metis plana Walker, an oil palm pest from Malaysia. The solution described by the authors was designed to distinguish between live and dead bagworm larvae using motion detection. Highly accurate results were obtained, i.e., 73–100% accuracy was achieved at a camera distance of 30 cm in close conditions. Using deep learning with Faster R-CNN in the methodology is a feasible, practical, and reliable method for bagworm detection. The above research is of great practical importance.

In the twentieth article [27], the authors detected and identified sows using image analysis. Two accurate models based on deep learning, Mask-RCNN and UNet-Attention, were developed. A very high recognition rate of 96.8% for a specific individual was obtained by using the Mask-RCNN model. The proposed system allowed for recognition of various behaviours, such as eating, drinking, etc.

The twenty-first article [28] deals with modifying and improving classification methods that can be applied to agriculture. The authors applied multi-temporal data fusion to specific types of images (i.e., MS and SAR) using a dynamic time-warping method in paddy rice classification. Three different types of SPOT6 satellite images and nine Sentinel-1A synthetic aperture radar images were used. Using neural networks made it possible to significantly improve the overall accuracy of the collected images. In turn, the best performance of the classification results was obtained from the decision tree.

In the twenty-second article [29], the authors presented the concept of modelling the mechanical properties of fresh and stored large cranberry fruit using multiple linear regression and artificial neural network models. The analyses aimed to determine the apparent elasticity index of large cranberry fruit variables relating to harvest time, water content, storage time and conditions. Using neural modelling techniques allowed for more accurate prediction of the elasticity of the tested material compared to classical regression techniques.

The twenty-third article [30] presents an innovative approach for fast identification and counting of pig herds in complex image segmentation. The proposed DeepLab V3+ network model was created based on the neural networks, a deep learning method. The approach used made it possible to obtain the values of comprehensive evaluation indicators at a very high level of 86%.

In the twenty-fourth article [31], artificial neural networks were used to identify the key meteorological factors affecting the harvest date and yield of soybeans (Glycine max L. Merrill) variety Augusta. To perform the task, the most important dates of successive development stages, meteorological data, and yield data were collected. The ranking of factors shaping yield and harvest date was obtained using the sensitivity analysis of a neural network. The study results are highly practical in nature, indicating “difficult for soybeans” periods, during which special care should be taken to ensure they remain in good condition.

The twenty-fifth article [32] evaluates the performance of various communication systems used through the Internet of Things in agricultural production. The authors conducted a detailed analysis of the specific protocols and simulation tools used to improve connectivity and connection quality. Using two gateways with Adaptive Data technology can increase the network delivery without changing energy consumption. The results presented are promising for broader data optimization.

In the twenty-sixth article [33], the authors used an improved YOLOX-S algorithm for robots picking kiwi fruit based on image analyses. The analysis results proved that the authors’ improved model helped improve the precision of kiwi fruit detection, reduced the number of model parameters, and improved the system’s speed. This approach can also be applied to other fruits or objects in general, so-called small, targets.

The twenty-seventh article [34] proposed a tea recognition method based on a light convolutional neural network and support vector machine (L-CNN-SVM). The purpose of this method was to recognize tea using wavelet numerical data generated by wavelet decomposition and reconstruction of the time–frequency signal. First, a redundant discrete wavelet transform was used to decompose the wavelet components of hyperspectral images of three teas (black, green, and yellow), which were used to construct the datasets. Second, a lightweight CNN model was improved to generate a tea recognition model. The results showed that the tea recognition results based on the L-CNN-SVM method outperformed MobileNet v2+RF, MobileNet v2+KNN, MobileNet v2+AdaBoost, AlexNet, and MobileNet v2. For the recognition results of the three teas using a LL plus HL plus LH wavelet component reconstruction, the overall accuracy rate reached 98.7%.

The twenty-eighth article [35] uses a fractional differential order for the soil hyperspectral inversion of iron oxide. Among other things, the content and movement of iron oxide in soil informs numerous degradation processes. Analysing the content of this component in the soil is difficult, because its analytical spectra overlap with the infrared components of organic matter. In addition to a specific spectral transformation, a soil iron oxide content prediction was made using artificial intelligence. It was shown that using a fractional-order differential transformation can significantly improve the results for this type of analysis.

In the twenty-ninth article [36], the authors used supervised classifiers for feature analyses, in order to evaluate the accuracy of a maturity analysis of fruit palm fruit. For this purpose, unconstrained remote sensing, advanced multivariate techniques, and artificial neural networks were used. Measurements were made in real-time. A very high image processing efficiency of more than 93% was achieved.

The thirtieth article [37] deals with improved optimization systems using the NSGA-III algorithm to improve the precision of two-way fertilizer applications. A test of the rotational speed and opening length of a bidirectional granular fertilizer applicator was conducted. A fertilizer rate prediction model based on machine learning was developed. Applying the new method led to a halving of the absolute error. The presented research significantly improves the quality of fertilizer spreading.

In the thirty-first article [38], the authors proposed PLSR models for predicting of the content of functional components in Brassica juncea based on hyperspectral imaging. The contents of chlorophyll, carotenoids, phenols, glucosinolates, and anthocyanins were studied. Incorporating SNV and first-order derivative pre-processing with spectral data into the study methodology yielded models with low prediction errors.

In the thirty-second article [39], the authors collected a dataset of whitefly-infested cotton leaves containing 5135 images divided into two main classes, namely healthy and unhealthy. They then used a compact convolutional transform (CCT) approach to classify the image dataset. Experimental results showed the effectiveness of the proposed CCT-based approach compared to other state-of-the-art approaches. The produced model achieved an accuracy of 97.2%.

The thirty-third article [40] presents a fuzzy certification of wheat quality. This analysis is based on a fuzzy model for wheat analysis. The authors developed a MATLAB application, with which they modelled perceptions concerning wheat’s main physical and chemical characteristics, obtaining a wheat batch quality index. The generated algorithm makes it possible to obtain and use a global quality index, which is applicable not only in the commercial sphere as a quality reference and for pricing, but also as a measure for evaluating processing capabilities.

In the thirty-fourth article [41], a multi-objective logistic distribution path optimization model with time constraints was constructed, and a genetic algorithm was used to optimize the commercial distribution path for fresh agricultural products. Combining the genetic algorithm with a real case to be studied, the study aimed to solve enterprises’ narrow distribution paths and promote the model’s application to similar enterprises with similar characteristics. The results show that (1) the commercial distribution path scheme optimized by the genetic algorithm can reduce distribution centre distribution costs and improve customer satisfaction, and (2) the genetic algorithm can bring economic benefits and reduce transportation losses in trade for trade distribution centres with the same spatial and quality characteristics as distribution centres for fresh agricultural products.

The thirty-fifth article [42] presents a data science approach that agglomerates the soil parameter space into a limited number of soil process functional units (SPUs) that can run agricultural process models. In reality, two unsupervised classification methods were developed to generate a 3D multidimensional data product consisting of SPUs, each defined by a multidimensional parameter distribution along a depth profile from 0 to 100 cm.

In the thirty-sixth article [43], the authors hypothesised that an automated body condition scoring system could indicate health and pregnancy success in cows. Therefore, this study aimed to determine the association of automated recorded body condition score (BCS) with pregnancy and inline biomarkers, such as milk beta-hydroxybutyrate (BHB), milk lactate dehydrogenase (LDH), milk progesterone (mP4), and milk yield (MY) in dairy cows.

In the thirty-seventh article [44], an intelligent control system for temperature and humidity in a piggery was proposed, based on machine learning and a fuzzy control algorithm. Sensors were used to collect data on temperature and humidity values and store these data in chronological order. These data formed a time series to train the GRU model, which was used to predict the curve of temperature and humidity changes in the piggery over the next 24 h.

In the thirty-eighth article [45], the authors proposed a lightweight maize disease identification model called the Double Fusion block with Coordinate Attention Network (DFCANet). DFCANet mainly consists of two components: Double Feature Fusion with Coordinate Attention and Down-Sampling (DS) modules. The results show that DFCANet has an average recognition accuracy of 98.47%.

The thirty-ninth article [46] is about a high-accuracy model for predicting blueberry yields, trained using structurally innovative datasets. Data were collected between 2016 and 2021, and included agronomic, climatic, soil data, and satellite data on vegetation. In addition, vegetation periods by BBCH scale and aggregates were considered. Of the 11 models, the Extreme Gradient Boosting algorithm performed best, with a MAPE prediction error of 12.48%.

The fortieth paper [47] proposed a novel scheme to optimise the decision-making capability of a combination of control and PID controller parameters, in order to improve the feasibility and practicality of variable fertiliser applicators. Firstly, an EDEM was applied to obtain the minimum acceptable bore length and the appropriate gap between the spiral vanes and the discharge cavity wall, followed by calibration experiments to establish a fertiliser rate-fitting model using polynomial fitting. Secondly, a modified sparrow search algorithm (SSA) with a chaotic operator and a mutation section of the DE algorithm was used to optimise the control combination using accuracy, homogeneity, and control time as evaluation criteria.

In the forty-first article [48], the research aimed to develop linear and non-linear models for predicting the protein content percentage of pea seeds, and to conduct a comparative analysis of the performance of these models. The analyses also focused on identifying the variables with the greatest influence on protein content. The research involved machine learning (specifically, artificial neural networks) and multiple linear regression (MLR) methods. The input parameters of the models were weather, agronomic, and phyto-phenological data from 2016 to 2020. The neural model (N1) performed better than the multiple regression (RS) model. The RMS error of the N1 model was 0.838, while the RS model obtained a mean error value of 2.696. The MAPE error for the N1 and RS models was 2.721 and 8.852, respectively. The sensitivity analysis conducted for the best neural network showed that the independent variables most influencing the protein content of pea seeds were soil abundance of magnesium, potassium, and phosphorus.

The forty-second article [49] aimed to investigate the automatic detection of Basal Stem Rot (BSR) at the seedling stage, using a pre-trained deep-learning model and hyperspectral images. The oil palm seedling image was divided into three regions to determine whether there were significant spectral changes at different leaf positions. How background images affect detection performance was also investigated. The segmented images of the plant sapling were automatically generated using a convolutional neural network (RCNN) based on the mask region. Three models were used for BSR detection: a convolutional neural network with a depth of 16 layers (VGG16) trained on the segmented image, and the VGG16 and Mask RCNN models trained on the original images. The results show that the VGG16 model performed best in terms of accuracy (91.93%), precision (94.32%), recall (89.26%), and F1 score (91.72%).

In the forty-third article [50], an automatic classification of the larval stage of the cutworm, starting from the second (S2) to the fifth (S5) stage, was proposed using a structure based on transfer learning. Five different CNN architectures, namely VGG16, ResNet50, ResNet152, DenseNet121, and DenseNet201, were used to categorise the larval stages. Of the five models used, the DenseNet121 model had the best classification accuracy of 96.18%. In addition, all developmental stages from S2 to S5 could be identified with high accuracy (94.52–97.57%), precision (89.71–95.87%), sensitivity (87.67–96.65%), specificity (96.51–98.61%), and F1 score (88.89–96.18%).

In the forty-fourth article [51], a study was carried out that created a dataset containing images of the leaves of cash crops, which were divided into two basic categories, namely healthy and unhealthy. The next step was to train a deep model to identify healthy and unhealthy leaves. The trained YOLOv5 model was used to identify stains in exclusive and public datasets. This study quickly and accurately identified even a small disease patch using YOLOv5. This research aimed to provide the best hyper-parameters for classifying and detecting healthy and unhealthy parts of leaves in exclusive and public datasets. The trained YOLOv5 model achieves a 93% accuracy on the test set.

The forty-fifth article [52] concerned research on the trajectory of agricultural machinery. The paper proposed a multi-node path planning algorithm based on the Improved Whale Optimised ACO (ACO) algorithm, called IWOA-ACO. The algorithm first introduces an inverse learning strategy, a non-linear convergence rate, and an adaptive inertia factor to improve the global and local convergence ability. The simulation results show that, in a flat environment, the length and energy consumption of the planned IWOA-ACO path are the same as those of Particle Swarm Optimisation ACO (PSO-ACO), and are 0.61% less than those of WOA-ACO. Furthermore, in a bumpy environment, the length and energy consumption of the IWOA-ACO planned path is 1.91% and 4.32% less than for PSO-ACO, and 1.95% and 1.25% less than for WOA-ACO.

The forty-sixth article [53] dealt with the development and evaluation of the Maize Yield Prediction System (MYPS), which uses a Short Message Service (SMS) and the Internet to enable rural farmers and government officials to predict end-of-season maize yields in Tanzania. The system uses Long Short-Term Memory (LSTM) deep learning models to predict end-of-season maize yields at the district level, based on remote sensing (NDVI) and climate data. The deep learning models are very effective in yield prediction, achieving a mean absolute percentage error (MAPE) of 3.656% and 6.648%, respectively, on test data.

The forty-seventh article [54] analysed the performance of pea (Pisum sativum L.) seed yield prediction using a linear (MLR) and a non-linear (ANN) model. The study used meteorological, agronomic, and phytophysical data from 2016 to 2020. The neural model (N2) generated highly accurate predictions of pea seed yield, with a correlation coefficient of 0.936 and RMS and MAPE errors of 0.443 and 7.976, respectively. The model significantly outperformed the multiple linear regression model (RS2), which had an RMS error of 6.401 and a MAPE error of 148.585. The sensitivity analysis conducted for the neural network showed that the traits with the greatest impact on pea seed yield were the onset of maturity, harvest date, total rainfall, and mean air temperature.

The forty-eighth paper [55] trained and compared the performance of two machine learning methods, a multivariate regression network and a ResNet-50-based neural network, for predicting plant biomass and determining plants’ relative growth rates in aeroponic cultivation. The training dataset consisted of images of 57 plants taken at two different angles every hour over five days. The results show that images taken from a top-down perspective give better results for the multivariate regression network. In contrast, images taken from the side are better for the ResNet-50 neural network. The best biomass estimates are obtained from the multivariate regression model trained on the top camera images using a moving average filter, giving a mean square error of 0.0466 g. The best estimates of relative growth rate are obtained from the ResNet-50 trained on images from both cameras, giving a mean square error of 0.1767 g/(g-day).

The forty-ninth article [56] aimed to evaluate different vegetation indices to predict the growth rates and harvest points of lettuce. Twenty-five genotypes of biofortified green lettuce were evaluated. Green leaf index (GLI), normalised green–red difference index (NGRDI), spectral slope saturation index (SI), and total colour index (HUE) were calculated from images taken 1, 8, 18, 24 and 36 days after transplanting (vegetative state). Averages were compared using the Scott-Knott test (p ≤ 0.05), and simple linear regression models were generated to monitor growth rates, yielding R² values ranging from 62% to 99%. Multivariate analysis confirmed genetic dissimilarity, with a correlation coefficient of 88.49%.

Finally, in the fiftieth article [57], a machine vision system was developed using a dataset of 7328 high-density images (1229 pixels per centimetre) of planthoppers collected in the field using sticky light traps. The dataset included four planthopper classes: brown planthopper (BPH), green leafhopper (GLH), white-backed planthopper (WBPH), and zigzag leafhopper (ZIGZAG). Five deep CNN models—ResNet-50, ResNet-101, ResNet-152, VGG-16, and VGG-19—were applied and tuned to classify the planthopper species. The experimental results indicated that the ResNet-50 model performed the best overall, achieving average values of 97.28% for accuracy, 92.05% for precision, 94.47% for recall, and 93.07% for the F1 score.

3. Conclusions

The Special Issue “Digital Innovations in Agriculture” brings fascinating insights into the agricultural sector’s future. The use of advanced ICT, data analytics, and artificial intelligence makes it possible to achieve sustainability, increase production efficiency, and improve animal husbandry. Digital innovations have the potential to revolutionise agriculture, and their implementation can contribute to a more sustainable, productive, and knowledge-based agricultural space. We hope that these papers will stimulate further research into this domain.