The Application of Deep Learning in the Whole Potato Production Chain: A Comprehensive Review

Abstract

The potato is a key crop in addressing global hunger, and deep learning is at the core of smart agriculture. Applying deep learning (e.g., YOLO series, ResNet, CNN, LSTM, etc.) in potato production can enhance both yield and economic efficiency. Therefore, researching efficient deep learning models for potato production is of great importance. Common application areas for deep learning in the potato production chain, aimed at improving yield, include pest and disease detection and diagnosis, plant health status monitoring, yield prediction and product quality detection, irrigation strategies, fertilization management, and price forecasting. The main objective of this review is to compile the research progress of deep learning in various processes of potato production and to provide direction for future research. Specifically, this paper categorizes the applications of deep learning in potato production into four types, thereby discussing and introducing the advantages and disadvantages of deep learning in the aforementioned fields, and it discusses future research directions. This paper provides an overview of deep learning and describes its current applications in various stages of the potato production chain.

1. Introduction

The potato is the world’s third-largest food crop [1], with significant nutritional and economic value, making it essential for ensuring food security globally [2]. Due to its characteristics of being easily digestible, high-yielding, and rich in carbohydrates and proteins [3], the potato is grown in over 125 countries. Many developing countries consider the potato to be a key crop for achieving food security strategies [4], with China and India being the largest producers. Figure 1 shows the statistical data of global potato production. The data are sourced from the FAO database.

However, throughout the entire potato production chain, various pathogens (such as blight and verticillium wilt) [5,6,7,8] and pests (such as the Colorado potato beetle, aphids, and nematodes) [9,10,11] frequently cause potato diseases. Additionally, improper irrigation strategies and soil nutrient/fertilization management [12,13,14] can negatively impact potato yields. Surface defects caused by pathogen stress or mechanical damage further affect the actual yields of potatoes. Moreover, predicting potato yields and timely adjusting planting strategies can effectively enhance tuber yields, thereby increasing economic benefits.

Deep learning is at the core of smart agriculture, thus playing a crucial role in executing complex tasks such as feature extraction, transformation, pattern analysis, and image classification due to its excellent capabilities [15]. Deep learning has been successfully applied in the production processes of various crops [16]. For instance, Guan et al. [17] proposed an improved DBi-YOLOv8 network for maize canopy organ detection. Liu et al. [18] developed a two-stage CNN model for assessing the severity of Alternaria leaf blotch in apple trees. Zhang et al. [19] proposed a real-time recognition and localization model for apple picking robots based on YOLOv5 and structured light. Furthermore, deep learning has been widely applied in agricultural fields such as sowing process estimation, crop yield prediction, irrigation strategy formulation, nutrient and fertilizer management, harvesting, pest and disease identification, and climate and soil mapping [20,21,22,23,24,25,26].

The high accuracy, scalability, excellent automatic feature extraction capabilities, and superior complex data processing abilities demonstrated by deep learning in various fields of other crop production chains make it possible to adopt new equipment, technologies, and algorithms in the potato production process. This greatly enhances the control over the potato production process and increases its overall yields.

Over the past few decades, the development of deep learning has progressed rapidly. The algorithms and architectures can be primarily categorized into the following types: Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Autoencoders, Generative Adversarial Networks (GANs), Reinforcement Learning (RL), Transfer Learning (TL), the Multilayer Perceptron (MLP), Deep Belief Networks (DBNs), and Restricted Boltzmann Machines (RBMs), among others [27,28]. To investigate the application of deep learning in the field of potato production, this study employs a qualitative research methodology. Using the Web of Science, we conducted searches with the keyword “Potato” combined with each of the terms “Deep Learning”, “CNN”, “Recurrent Neural Networks/RNN”, “LSTM”, “Autoencoder”, “Generative Adversarial Networks/GAN”, “Reinforcement Learning”, and “Transfer Learning”. The search results are illustrated in Figure 2.

This review is divided into five sections: Section 2 provides an overview of deep learning, thus elucidating its fundamental concepts and advantages; Section 3 categorizes the applications of deep learning in the potato production process into crop monitoring and health management, crop yield prediction, and intelligent irrigation and fertilization management. It offers a detailed review of the applications of deep learning in each potato production process and summarizes the existing issues; Section 4 discusses the current advantages of deep learning in various processes of the potato production chain, the challenges it faces, and future research directions; Section 5 provides a review and summary of the entire paper.

2. Overview of Deep Learning

The concept of Deep Learning (DL) originated from research on Artificial Neural Networks and was first introduced by Hinton et al. in 2006 in the journal “Science” [29]. The deep belief network proposed by Hinton et al. [30] marked the advent of deep learning. The fundamental idea of deep learning can be summarized as follows: using neural networks for data analysis and feature learning, where data features are extracted through multiple hidden layers. Each hidden layer can be considered to be a perceptron, which extracts low-level features and then combines these low-level features to derive abstract high-level features, thus significantly alleviating the problem of local minima [31].

Deep learning has garnered increasing attention from researchers in recent years due to its ability to overcome the limitations of traditional algorithms that heavily rely on manually designed features, thereby becoming a research hotspot. It has demonstrated excellent potential in the field of big data analysis, thus partially addressing the issue of overfitting in training data. Moreover, the pretraining procedure of self-supervised learning assigns nonrandom initial values to the network, thus leading to better local minima after training and achieving faster convergence rates. Consequently, it has been successfully applied in computer vision, pattern recognition, and natural language processing [32].

Traditional image classification and recognition methods based on manually designed features can only extract low-level information, thus making it difficult for these techniques to capture deep and complex image feature information. In contrast, deep learning can directly perform self-supervised learning from raw images, thus obtaining multi-level image feature information, including low-level features, intermediate features, and high-level semantic features. This capability addresses the bottleneck in traditional manually designed feature-based image classification and recognition methods, which struggle to extract deep and complex image features. Given the characteristics of the potato production process—such as an unstructured environment, high biodiversity, and variable crop conditions [33]—traditional methods for applications in potato crop monitoring and health management, crop yield prediction, and intelligent irrigation and fertilization management primarily employ manually designed feature-based image recognition methods. These methods are challenging, thereby being highly reliant on experience and unable to automatically learn and extract features from raw images. Deep learning, on the other hand, can automatically learn features from big data without manual intervention. Additionally, deep learning models typically consist of multiple layers, possess good autonomous learning and feature representation capabilities, and are able to automatically extract image features for tasks such as image classification and recognition, segmentation, and counting. Therefore, deep learning plays a crucial role in various processes within potato production.

Over the past two decades, DL has developed numerous well-known models [28]. These models can achieve automated feature extraction in high-dimensional feature spaces, thus providing significant advantages in the field of image recognition within potato production compared to other methods. Moreover, with the increase in computational power and the growing number of training samples for potato agricultural production tasks, the representational capabilities of deep learning are continually improving.

Common open-source tools used for deep learning models include Torch/PyTorch [34], Caffe [35], TensorFlow [36], and Theano [37]. All of these widely used deep learning open-source tools support crossplatform operations and can run on operating systems such as Windows, Android, iOS, and Linux. Training on GPUs with these tools can achieve faster training speeds. The characteristics of these tools are summarized in

Table 1. Features of four common deep learning open-source tools.

Tools	Supporting Hardware	Applicable Interface	Usability
Torch/PyTorch	GPU, CPU, FPGA	C, Python, Lua	Easy to develop and debug, modular, easily extensible, and easy to learn.
Caffe	GPU, CPU	Python, MATLAB	Fast, highly extensible, and have strong readability.
TensorFlow	GPU, CPU, Mobile	C, Python	High flexibility, support for distributed applications
Theano	GPU, CPU	Python	Highly flexible and effective

.

To systematically review the application of deep learning in various processes of potato production, this paper categorizes the application of deep learning in potato production into the following categories:

• Crop monitoring and health management;
• Crop yield prediction;
• Intelligent irrigation and fertilization management;
• Deep learning models for potato price forecasting. These are further subdivided into the following:

1.a.

Pest and disease detection and diagnosis;

1.b.

Plant health status monitoring;

2.a.

Yield prediction models;

2.b.

Product quality inspection;

3.a.

Irrigation strategy formulation;

3.b.

Precision fertilization management.

These categories are shown in Figure 3.

3. The Application of Deep Learning in Potato Production

3.1. Crop Monitoring and Health Management

3.1.1. Detection and Diagnosis of Pests and Diseases

Potatoes are often infested with diseases (e.g., early blight, late blight, black scurf, and bacterial wilt) and pests (e.g., Colorado potato beetle (CPB), Mexican bean beetle (MBB), and aphids) [38,39,40], which severely affects potato yields and product quality. However, traditional methods for detecting and diagnosing potato diseases and pests are labor-intensive and expensive, thus requiring significant time even for experienced experts [41]. With the advancement of deep learning, the early detection of potato diseases and pests has become easier, faster, and more cost-effective. In recent years, the development of artificial intelligence (AI) and computer vision (CV) technologies has paved the way for improving crop disease classification [15,42], thus leading to extensive research on the localization and classification of diseases and pests based on deep learning models [43,44,45,46,47].

• Detection and Diagnosis of Diseases

As shown in Figure 4, early blight and late blight are the most common leaf diseases of the potato. Based on [48], Singh et al. [49] proposed an efficient CNN for potato leaf disease detection using TL and Base Learning. Through comparative experiments with VGG16, InceptionNetV3, ResNet, NasNetV2, MobileNetV2, and VGG19, the proposed model demonstrated higher detection accuracy than existing technologies. Additionally, Kumar et al. [50] applied 10 deep learning models, including DenseNet121, ResNet152V2, and VGG19, and they used preprocessing techniques such as dilation and erosion to adjust image sizes. They developed a method for detecting and classifying crop diseases, including early blight and late blight in potatoes. Lizarazo et al. [51] proposed an end-to-end detection method for potato late blight based on UAV multispectral images.

Rashid et al. [52] used YOLOv5 image segmentation technology to extract potato leaves and detected early blight and late blight using a CNN. They proposed a multilevel deep learning model for potato leaf diseases, thereby demonstrating excellent recognition ability. Tiwari et al. [53] proposed a method for diagnosing blight in potatoes using VGG19 and other pretrained models through fine-tuning (TF), thus achieving an accuracy of 97.8%. Al-Adhaileh et al. [54] proposed a model using a CNN for fine-tuning to detect potato blight, thus achieving 99% accuracy. Additionally, researchers have used VGG19 [55], CNNs [56], and transfer learning [57] to identify diseases such as potato blight, leaf roll disease, and verticillium wilt, thus demonstrating excellent results on several datasets. Notably, Paul et al.’s [55] research is the first to successfully implement and report the detection and classification of four potato diseases. Shukla et al. [56] took the process a step further by developing an app, as shown in Figure 5a, thus building on theoretical research.

Bengamra et al. [58] used reinforcement learning and explainable AI (XAI) to perform a quantitative and qualitative analysis of potato leaf diseases on the PlantDoc dataset. XAI employed deletion and insertion performance metrics for quantitative evaluation and used saliency maps for qualitative evaluation, thus achieving good results in both quantitative and qualitative analyses (as shown in Figure 5b,c).

In this subsection, we comprehensively reviewed the research on deep learning in the field of potato disease detection and diagnosis. Due to its high accuracy, low cost, rapid detection, and nondestructive nature [59,60,61], deep learning technology has achieved significant success in this area. However, in practical production applications, current deep learning technology still faces several limitations and challenges:

• The lack of datasets specifically tailored for potato diseases limits the development of related research. Additionally, many existing studies lack self-constructed datasets, thus resulting in the need for further verification of the models’ practicality.
• Current datasets generally suffer from insufficient sample sizes and a lack of sample diversity, which may lead to model overfitting, as well as poor generalization and robustness of the models.
• Furthermore, most current research focuses primarily on algorithm development and improvement. Future efforts could consider developing apps to enhance the practicality of the research and apply the findings to actual production.

B.

Detection and Diagnosis of Pests

Talukder et al. [62] developed a neural network based on CT Inception V3-RS for identifying potato pests. The proposed PotatoPestNet model utilized five transfer learning models, including CMobileNetV2, CNASLargeNet, and CXception for pretraining. Notably, as shown in Figure 6A, they constructed a dataset containing eight common potato pests and validated the effectiveness of the PotatoPestNet model on this dataset.

Roldán-Serrato et al. [63] proposed an automated pest detection system using artificial neural networks—specifically targeting the MBB and the CPB—and utilizing RSC and LIRA neural classifiers. The recognition rates for CPB and MBB were 88% and 89%, respectively, using the LIRA and RSC. Thenmozhi et al. [64] proposed a model based on deep convolutional neural networks and transfer learning for crop pest classification. They applied their algorithm to the National Bureau of Agricultural Insect Resources (NBAIR) dataset (40 classes), Xie1 dataset (24 classes), and Xie2 dataset (40 classes), thus achieving detection accuracies of 96.75%, 97.47%, and 95.97%, respectively. A visualization of some dataset examples is shown in Figure 6B. Additionally, Zhang et al. [65] improved the YOLOv3 network structure by combining upsampling and convolution operations to achieve deconvolution, thus effectively detecting smaller crop pest samples in images.

Through a systematic and comprehensive review, we have observed significant progress in pest detection and diagnosis tasks in potatoes using deep learning techniques such as CNN and transfer learning. However, due to some limitations in existing potato pest datasets, deep learning still faces some constraints and challenges in this task:

• High-accuracy models require a large amount of data. Potatoes are affected by a wide variety of pests, which have significant impacts on their growth and yield. However, existing datasets commonly suffer from insufficient samples, limited pest species, and low data quality. Additionally, the data for many potato pests are still lacking.
• Images in existing datasets lack multiple angles and images of potatoes affected at different growth stages, thus limiting the model’s generalization ability in applications.
• Potato pests exhibit morphological differences at different growth stages, and these differences may be subtle. The features of different pests may also interfere with each other, thus leading to reduced identification accuracy. Therefore, future research needs more images for different pests at different growth stages to improve dataset completeness.

3.1.2. Plant Health Status Monitoring

The health condition of plants has a significant impact on their yield. The timely and effective detection of the health status of potato plants helps in adjusting planting strategies in real time and taking appropriate measures to enhance both the overall yield and yield per unit area of potatoes. Previously, traditional machine learning methods have been attempted to monitor the health status of potato plants. For example, Hou et al. [66] developed a method for predicting the severity and epidemic period of late blight in potatoes using visible/near-infrared spectroscopy based on machine learning techniques. Ma et al. [67] combined radiative transfer modeling with active learning for estimating chlorophyll content in potato leaves. However, traditional machine learning models struggle to address the limitations of methods such as radiative transfer modeling for characterizing plant status [67]. Therefore, deep learning models have gradually become a research focus in monitoring the health status of potato plants.

The detection of potato buds is a primary concern in achieving the intelligent cutting of potato seeds [68], as it determines their health status and sprouting ability, thus ensuring seed potato quality. Xi et al. [69] proposed an improved Faster R-CNN two-stage potato bud detection method based on ResNet-50 as the backbone network, thus achieving a detection accuracy of 96.32% for sprouted potato buds. Wang et al. [70] introduced an in situ potato leaf detection method by incorporating a convolutional attention module into the Faster R-CNN network, thus achieving an average precision of 95.7%. Xu et al. [71] developed two lightweight two-stage models based on the ConvNeXt V2 model for monitoring the severity of early and late blight diseases in potatoes. These studies provide crucial technologies for the applications of smart agriculture and ecological monitoring.

However, two-stage object detection algorithms require the generation of a series of candidate boxes first, followed by classification and regression based on the candidate box regions. Although these methods achieve high detection accuracy, they are time-consuming and not suitable for real-time detection needs in agricultural operations. One-stage methods perform candidate box generation, classification, and regression simultaneously, thereby balancing detection speed and accuracy and making them more suitable for the real-time requirements of agricultural production. The YOLO series of algorithms [72,73,74] represent the most prominent one-stage object detection methods.

Xiu et al. [75] proposed a potato automatic seedling bud detection method based on the YOLOv4 network. Additionally, Zhang et al. [68] presented an improved potato bud detection algorithm based on YOLOv5s. Experimental results demonstrated that the model outperformed the original YOLOv5s model, Faster R-CNN, YOLOv6, YOLOX, and YOLOv7 models in potato bud detection tasks. Geng et al. [76] embedded DAS variability attention in the feature extraction network and the feature fusion network, thereby proposing a DAS-YOLOv8 model. This model achieved an mAP of 94.25% on the task of potato eye detection. Additionally, Liu et al. [77] proposed an improved Bud-YOLOv8s model. Experimental results indicated its excellent performance in the task of potato eye detection, thus significantly outperforming other models of YOLO series. Wang et al. [78] replaced the feature pyramid in the YOLOv3 algorithm with an attention feature pyramid, thus overcoming interference during feature fusion, and they proposed an efficient potato deformity recognition algorithm with an mAP of 94.46%. Accurately obtaining information on potato seedlings is crucial for high-quality potato production. As shown in Figure 7A, Wang et al. [79] employed VanillaNet as the backbone network and proposed the VBGS-YOLOv8n model for potato seedling detection. Additionally, they developed a dataset of potato seedling images captured by drones, thus providing technical support for potato health status monitoring.

Late blight is one of the diseases significantly affecting potato yields. Timely and effective monitoring of the severity and spatial distribution of late blight in potatoes helps adjust production strategies promptly and improve potato yields. Sun et al. [80] compared and constructed severity monitoring models for late blight using SVM, Random Forest (RF), and KNN. They proposed a relief-mRmR algorithm based on dual-drone collaboration. Experimental results showed that the RF model achieved the highest accuracy, with an overall accuracy and kappa coefficient of 92.50% and 0.90, respectively, thus demonstrating excellent performance in this task. Duarte-Carvajalino et al. [81] conducted similar research, thereby evaluating the severity of late blight in potato using an MLP, a Deep Learning CNN, a Support Vector Regression, and an RF. Additionally, based on hyperspectral images acquired by drones, Shi et al. [82] proposed a novel end-to-end deep learning model named CropdocNet. This model can extract the spectral–spatial hierarchical structure of late blight from UAV hyperspectral image data, thus enabling the monitoring of potato late blight and its severity. The model integrates spectral–spatial scalar features into hierarchical vector features to represent the rotational invariance of potato late blight under complex field conditions. Experimental results showed that the CropdocNet model generated mAP values of 98.09% and 95.75% on the training and test datasets, respectively. This result provides a solid scientific basis for the visual monitoring of potato and other crop diseases. The visualization of the monitoring results can be seen in Figure 7B.

Unlike the application of traditional machine learning or hyperspectral imaging techniques in monitoring the health status of potato plants, Kool et al. [83] combined hyperspectral imaging technology, local comparison, and a CNN to propose an early detection method for potato plant diseases. This method has the potential to be integrated with remote sensing technology, thus advancing research in potato health status monitoring. Notably, in addition to remote sensing-based applications for monitoring potato plant health, DL has also been successfully applied to portable ground monitoring equipment. Based on the explainable deep classification model proposed by Habaragamuwa et al. [84], Oishi et al. [85] developed an automatic detection system for abnormal potato plants using a portable camera. This study reduces the need for specialized expertise in monitoring potato plant health and provides convenience for potato agriculture in small plots. The conceptual diagram of the proposed potato abnormal plant detection system is shown in Figure 7C.

Figure 7. (A) Comparison effect of the model before and after the improvement on the detection of potato seedlings at different heights and at different stages [79]. (B) Patch-scale experiments for the classification of healthy potatoes and potatoes with late blight in (a) experimental site 1 and (b) experimental site 2. The example patches on the right illustrate the accuracy comparison between ground truth (GT) surveys and late blight prediction levels (PLs). Each value within the patches represents the disease ratio (late blight pixels/total pixels) [82]. (C) Conceptual diagram of the proposed potato abnormal plant detection system: An alert sound is emitted when an abnormal plant or leaf is detected, and the user can review the detection results on the terminal [85].

Figure 7. (A) Comparison effect of the model before and after the improvement on the detection of potato seedlings at different heights and at different stages [79]. (B) Patch-scale experiments for the classification of healthy potatoes and potatoes with late blight in (a) experimental site 1 and (b) experimental site 2. The example patches on the right illustrate the accuracy comparison between ground truth (GT) surveys and late blight prediction levels (PLs). Each value within the patches represents the disease ratio (late blight pixels/total pixels) [82]. (C) Conceptual diagram of the proposed potato abnormal plant detection system: An alert sound is emitted when an abnormal plant or leaf is detected, and the user can review the detection results on the terminal [85].

In this subsection, we systematically review the development of potato plant health monitoring technologies. Over the past decade, automated monitoring technologies for potato plant health have made significant advancements, with spectral imaging, machine learning, and deep learning being the primary technologies in this field. Deep learning models, due to their considerable advantages, have been widely applied in this area. However, several limitations have been revealed during their application:

• Most current research is still limited to theoretical studies such as model optimization, thus lacking the development of compatible APPs or devices, which restricts the practical application of deep learning technology in monitoring potato plant health.
• Due to morphological differences in potato plants at different growth stages, there is currently a lack of large, high-diversity datasets, thus resulting in insufficient robustness and generalization capabilities of the corresponding deep learning models.
• The ability of existing deep learning-based monitoring models to identify early symptoms of potato diseases still needs improvement.
• There is still room for improvement in the real-time performance and efficiency of practical applications.

3.2. Crop Yield Prediction

3.2.1. Deep Learning-Based Yield Prediction

The morphological, physiological, and biochemical parameters of plants are critical for predicting potato yields. Rozentsvet et al. [86] investigated the relationship between potato yields and various characteristics of 24 potato varieties, including plant height, leaf size and number, and root systems, as well as biochemical traits such as photosynthesis, respiration, and the storage of primary and secondary biosynthetic products. This research has laid a solid scientific foundation for potato yield prediction based on deep learning.

Previously, researchers have applied machine learning to predict potato yields. Hofstee et al. [87] proposed a model for bulk potato yield prediction using a line scan camera based on machine learning. Long et al. [88] used stereo vision to calculate depth information for measuring potato volume. Fan et al. [89] developed a semimechanistic model using proximal sensing and environmental variables to predict potato yields. Li et al. [90] employed RF regression models and full-spectrum partial least squares (PLS) regression models to estimate potato biomass and predict yields. Additionally, Kuradusenge et al. [91] and Kurek et al. [92] conducted similar studies using RF, SVM, and other ML models. However, deep learning technology can optimize the feature selection and node splitting of random forests, thus further enhancing the accuracy and robustness of these models [93,94]. M. Awad et al. [95] created an optimized model to evaluate potato yields, thus compensating for the limitations caused by high-spatial-resolution remote sensing images due to climatic conditions. The powerful performance of deep learning in image processing and feature extraction, along with its demonstrated potential, makes it a research hotspot in the field of potato yield prediction, thus addressing the shortcomings of traditional techniques. Lee et al. [96] proposed a Mask R-CNN network with ResNet-101 as the backbone, and they trained over 12,000 iterations on a dataset of 230 images to address the issue of distinguishing harvested potatoes from soil. This method achieved an average detection precision of 90.8% and is expected to be deployed on potato harvesters for yield prediction in the future.

Previously, researchers developed a pixel segmentation method based on VGG [97] and U-Net convolutional networks [98]. Building on this research, Wyniawskyj et al. [99] developed a precision agriculture cloud platform named KORE. This platform integrates data from satellites, drones, and ground sensors using CNN technology to assess potato field yields. This study not only demonstrated the potential of CNN technology in yield prediction but also showed that the KORE platform could support subsequent agricultural tasks such as planting, water and fertilizer management, and quantitative spraying. However, when collected data involved excessive plant overlap and obstructions, the model’s recognition performance deteriorated. The main issues with this model include the following: (1) poor prediction results due to obstruction by trees; (2) instances of duplicate counting of potato plants; (3) the continuous distribution of potato plants leading to repeated counting.

Jang et al. [100], based on YOLOv5 and employing DeepSort, assigned unique IDs to harvested potatoes, thus achieving a counting accuracy of 99.97%. In their subsequent study [101], they deployed cameras in the sorting unit of a potato harvester to estimate the quality of continuously delivered potatoes. The recognition rate reached 95.2%, with a total measurement quality error of 9%. This research could be further integrated with GPS to create yield maps for underground crops. Przybylak et al. [102] proposed a potato tuber quality assessment method based on neural network image analysis. Their image processing approach involves several stages, starting from the source image, passing through segmentation, mask creation, and background removal, and culminating in the application of image filters to produce the final image.

Additionally, Al-Gaadi et al. [103] obtained satellite images during the potato growth stages and generated a Normalized Difference Vegetation Index (NDVI) and Soil-Adjusted Vegetation Index from the images. They successfully predicted potato yields, with the predicted yields differing from the actual yields by only 3.8%. This opens up possibilities for the application of deep learning models based on satellite imagery data for potato yield prediction. Jasim et al. [104] investigated the use of active and passive sensors to predict potato yields based on potato vegetation indices and plant pigment levels. Building on previous research [105], Liu et al. [106] estimated above-ground potato biomass using vegetation indices and texture features constructed from hyperspectral imagery captured by drones. However, the stability of the model proposed in this study was relatively low [106]. Subsequently, they conducted further research [107], but issues such as slow model speed and inadequate accuracy persisted. Future efforts could explore the integration of deep learning to develop more accurate and robust models.

Over the past few decades, nondestructive detection techniques such as spectroscopy, ML, and DL have been used for potato yield prediction. Among these, deep learning models have demonstrated excellent performance in this field due to their speed, high accuracy, and real-time capabilities. However, deep learning also has some limitations:

• Climate and soil conditions vary across different potato production regions, thus limiting the generalization ability of existing deep learning models for potato yield prediction. These models require a large amount of localized data and may struggle to be applied in other regions.
• Obtaining high-quality, comprehensive potato field data from different regions is challenging, and annotating such data requires significant human effort. The scarcity and imbalance of data greatly affect the performance of the models.
• Current deep learning-based methods for potato yield prediction still struggle to accurately address the problem of decreased accuracy caused by environmental and climatic changes. Additionally, deep learning models are complex, opaque, and lack interpretability, which may impact the formulation of agricultural strategies.
• The training and prediction processes of deep learning models require extensive computational resources and time. These high computational requirements may affect the economic viability of potato production.

3.2.2. Product Quality Inspection

Deep learning provides a simple, fast, and cost-effective method for assessing the quality of potatoes in production and food processing. Pathogens such as bacteria, fungi, phytoplasmas, viruses, and phytoplasmas can infect potatoes, and mechanical damage may occur during harvesting. Potatoes infected with these pathogens or severely damaged mechanically, even after harvest, are difficult to realize their original economic value [108]. Therefore, severe mechanical damage and pathogen-infected potatoes can reduce the yield and quality of potatoes, thus reducing their economic benefits. Based on this, the quality inspection of potatoes plays a crucial role in predicting potato crop yields.

Several image-based methods for potato product quality inspection have been reported previously. Riza et al. [109] utilized a genetic algorithm–PLS as a variable selection method and completed the detection and classification of potato mechanical damage, common scab, and others based on diffuse reflectance characteristics. Dacal-Nieto et al. [110] employed machine vision technology and infrared waves for potato product quality inspection tasks. Su et al. [111,112] achieved potato product quality detection using hyperspectral imaging technology in the visible and near-infrared regions. Dorokhov et al. [113] developed an optical detection system for potato tuber surface damage with a directional diffuse module. Additionally, Wu et al. [114] developed a potato decay and soft rot monitoring system using chemical reaction dyes based on the KNN algorithm.

Samatha et al. [108] utilized improved support vector machine (MSVM) and CNN models for potato image segmentation and feature extraction, thus proposing a deep neural network model for potato product quality inspection. Experimental results demonstrated an accuracy of 97% to 99%, thus providing technical support for the application of potato product quality inspection and subsequent crop yield prediction tasks. Furthermore, Oppenheim et al. [115] and Arshaghi et al. [116,117] employed CNN models for the detection and classification of potato tuber diseases such as black scurf, common scab, gray mold, and silver scurf, thus contributing to potato product quality inspection. Notably, Arshaghi et al. [116] compared the potato product quality detection CNN model with the SVM, thus validating the superiority of deep learning in this task. Additionally, in their subsequent research [117], Arshaghi et al. established a potato surface defect dataset comprising 5000 images, thus further enhancing the generalization ability and robustness of the proposed model.

Moreover, Wei et al. [118] proposed a model for potato quality grade monitoring and classification using laser backscatter and an improved VGG16 network, thus achieving an accuracy of 95.33%. The study compared this model with the original VGG16 and DenseNet121. Additionally, Zhang et al. [119] utilized an improved U-Net to detect surface defects on potatoes. Wang et al. [120], based on deep transfer learning models, developed three DCNN methods, namely SSD Inception V2, RFCN ResNet101, and Faster RCNN ResNet101, for potato surface defect detection. These studies provide scientific support for quality grade monitoring and classification in subsequent production processes such as the storage, transportation, sale, and processing of potatoes.

Potato product quality inspection essentially involves a feature selection task. Currently, researchers have developed various optimization algorithms such as Artificial Bee Colony (ABC) [121], Ant Colony Optimization (ACO) [122], Particle Swarm Optimization (PSO) [123], and Buzzard Optimization Algorithm (BUZOA) [124]. Arshaghi et al. [125], based on BUZOA’s feature selection method, utilized these features for correlation segmentation, feature extraction, and classification, thus proposing a potato surface defect detection method for product quality inspection. Experimental results demonstrate that the BUZOA algorithm performs well in achieving global optima and rapid convergence. DL can extract deep and complex image feature information. In the future, it may be beneficial to explore the integration of deep learning with optimization algorithms such as ABC and BUZOA to enhance the speed and quality of potato product quality inspection tasks.

Through a comprehensive review, we have observed that researchers in the field of potato product quality inspection have extensively employed deep learning, as well as new approaches combining traditional techniques with deep learning models. Despite the demonstrated effectiveness and significant potential of deep learning in potato product quality inspection, there are still some shortcomings:

• Different varieties of potato tubers may exhibit morphological differences, and surface defects caused by the same disease may vary in appearance. Currently, deep learning models used for potato product quality inspection lack detection capabilities tailored to different potato varieties, thus limiting their practical application in production.
• Due to the poor robustness of deep learning models to image noise, changes in lighting conditions, and background interference, detection results may be unstable. Since potato product quality inspection often encounters these issues in practical production environments, it is necessary to continue refining and optimizing models in subsequent research to improve their robustness.
• Deep learning models deployed on potato production machinery such as harvesters demand high real-time performance. This requirement poses challenges to the application of deep learning in this field, thus necessitating ongoing optimization efforts.

3.3. Intelligent Irrigation and Fertilization Management

In the production process of potatoes, irrigation and nitrogen fertilizer management play a crucial role in sustainable production and yield enhancement [126,127,128]. Producers need to adjust irrigation and fertilization strategies in a timely manner based on environmental factors such as climate, changes in potato plant growth status, and other factors. Cheng et al. [129] investigated the effects of irrigation and fertilization methods on potato yield, water absorption, and productivity, thus laying a scientific foundation for the application of deep learning in irrigation and fertilization tasks.

3.3.1. Application of Deep Learning in Irrigation Strategy Formulation

As a shallow-rooted crop, proper irrigation management is crucial for potatoes [130]. For irrigation, a well-designed irrigation plan helps farmers improve water resource utilization efficiency, reduce production costs, and ultimately enhance economic benefits, which are essential for potato production [131,132,133]. Like other crops, the potato is influenced by environmental factors such as temperature, humidity, climate variations, and rainfall. Characterizing the changes in each environmental factor allows for timely adjustments in irrigation water quantity and frequency, thus facilitating potato growth and development while mitigating yield reduction caused by environmental changes [134,135,136]. Currently, deep learning has achieved significant success in irrigation strategy formulation.

Qu et al. [137] developed a method using UAV multispectral data and vegetation indices to monitor potato canopy leaf water content and soil moisture content, thus providing a solid scientific basis for intelligent irrigation management of potatoes. Li et al. [138] investigated the impact of prolonged drought duration on potato plant growth, as well as yield and tuber quality, thus paving the way for the formulation of potato irrigation strategies. Jiménez-López et al. [33] utilized three deep learning models, namely 1-DCNN, LSTM, and a hybrid CNN–LSTM model (as shown in Figure 8), with climate variables such as temperature, soil moisture content, wind speed, and crop evapotranspiration as inputs to develop a method for potato irrigation strategy formulation. Through validation, this model proved to be helpful in accurately formulating potato irrigation prescriptions, thereby significantly improving water resource management.

The potato, as a typical shallow-rooted crop, has relatively weak water and nutrient absorption capabilities [139,140], and its root system exhibits different dynamic characteristics depending on the soil texture and the growth stage of the potato [141,142]. Therefore, studying the spatiotemporal distribution patterns of potato root systems is crucial for the scientific management of soil moisture and nutrients. Wu et al. [143] proposed a method for potato root image segmentation based on an improved DeepLabv3+ semantic segmentation network. They calculated the length of potato roots in different soil layers based on the segmented images, thus obtaining the root length of potatoes in different soil layers during various growth stages. Experimental results showed that the model achieved a mean intersection over union of 94.05% and a mean pixel accuracy of 95.72%. Additionally, Wu et al. [144], in combination with new sensing technologies, developed an intelligent irrigation system capable of real-time monitoring of soil moisture conditions in potato fields. This system provides a scientific basis for determining irrigation water usage and irrigation strategies for potatoes.

Additionally, Biswal et al. [145] found that the use of drip irrigation and straw mulching can influence the microclimate of potato fields, thereby increasing potato yield and water use efficiency. In future research, it would be beneficial to incorporate deep learning to explore the optimal straw mulching rate and drip irrigation volume for potato production, thus aiming to increase potato yields while alleviating pressure on water resources [131,146]. Similar findings were reported by Zhao et al. [147]. Yin et al. [148] utilized the AquaCrop model (a process-based model for simulating crop biomass and yield response to water [149]) to simulate the effects of 30 different irrigation schemes under two irrigation scenarios on potato production and tuber yields, thus aiming to explore the most suitable irrigation strategies for potatoes in arid regions. Batch simulations of various irrigation schemes are undoubtedly labor-intensive and expensive. Introducing deep learning models into the AquaCrop model to investigate the optimal irrigation strategies for potatoes would greatly reduce research costs. Moreover, irrigation strategy optimization studies based on manual parameter selection [150,151,152,153,154] could also benefit from the integration of deep learning techniques to reduce research costs and enhance efficiency.

Alibabaei et al. [155] utilized RNNs, including bidirectional LSTM (BLSTM) and Gated Recurrent Units (GRUs), to develop an end-to-end decision support system based on climate big data. This system is designed for formulating irrigation strategies such as irrigation water volume, irrigation frequency, timing, and yield prediction.

The main basis for formulating irrigation strategies includes environmental parameters, plant status parameters, and field images. Therefore, the application of deep learning technology in intelligent irrigation for potatoes mostly relies on tasks such as image processing, time series prediction, and reinforcement learning. However, there are the following limitations in the current practical application:

• Every year, it is necessary to collect images of the entire potato field and continuously update local environmental parameter data, which limits the generalization ability of current deep learning-based intelligent irrigation decision models for potatoes and makes it difficult to transfer them for use in different regions.
• To train the model, a large amount of calibration is required for the collected images, which is a labor-intensive and time-consuming task. The scarcity of datasets can result in poor performance, thereby affecting the formulation of irrigation strategies.
• Factors such as changes in lighting, the overlap of potato plants, leaf coverage, and variations in image proportions all negatively impact the effectiveness of the model [156], thus constraining the development of potato intelligent irrigation decision models based on deep learning.

3.3.2. Precision Fertilization Management

Unscientific fertilization often fails to meet the growth and development needs of potatoes, thus resulting in low yields. As a major agricultural input, the price of fertilizers has risen sharply since the end of 2020 [157], which has further increased the production costs of potatoes due to extensive nutrient management [158,159]. Amare et al. [160] found that applying nitrogen and phosphorus fertilizers can more than double potato yields. Scientifically precise fertilization management can increase potato yields by 50% [161]. Additionally, Xu et al. [162] reported that scientific and rational fertilization management can reduce environmental risks while increasing potato yields.

Previously, researchers have applied many nondestructive testing methods to precision fertilization management for potato. The canopy nitrogen content of potatoes is an important indicator for assessing their growth status and guiding precise fertilization management. Guo et al. [163] and Zhou et al. [164] used a hyper- and multispectral imagery method to estimate the canopy nitrogen content in potatoes. Additionally, Fan et al. [165] constructed 12 texture indices based on hyperspectral texture features and used an arbitrary variable combination optimization algorithm to optimize these indices for monitoring the nitrogen content in potato plants at multiple growth stages and guiding fertilization. Di et al. [166] used the RF algorithm to evaluate the relative importance of main N management indicators on potato yields and environmental indices (EI_Ys), thus providing a scientific basis for precision fertilization management and sustainable production in potato cultivation.

Unlike past methods such as the AdaBoost algorithm [167], the SIFT algorithm [168], and the Tamura Texture Feature algorithm [169], which relied on manually extracting features, deep learning provides an effective framework for automatically extracting target features. Xia et al. [170] proposed a simple and practical potato deep learning network called PCANet. After image segmentation, grayscale processing, binarization, and filtering, they achieved feature extraction training and testing on 48 × 48 potato plant images. The study achieved a detection speed of 1.22–1.31 images per second, thus providing a scientific basis for efficient and precise fertilization and potentially increasing potato yields. Guo et al. [171] proposed a near-infrared visible light spectral potato dry matter online detection method based on 1-DCNN, thus highlighting the significant potential of introducing deep learning models into spectral detection methods to more accurately and efficiently evaluate agricultural quality and guide precision fertilization management.

Scientific and rational fertilization management of potatoes not only enhances potato yields and economic benefits but also effectively reduces environmental risks. Although deep learning has demonstrated excellent performance in other potato production areas, there are still some shortcomings in precision fertilization management tasks:

• Although deep learning has many advantages and has been widely applied in other potato production processes, the research on precision fertilization management models based on deep learning for potatoes is limited and not widely utilized. Many studies and practices are focused on other crops, with few related cases for potatoes.
• Traditional precision fertilization management has accumulated a wealth of experience and data in potato production, but the integration of deep learning with traditional methods is still insufficient. The better integration of traditional techniques with deep learning would greatly enhance the effectiveness of precision fertilization.
• Currently, most research is still limited to theoretical validation, thus lacking practical applications and large-scale field experiments.
• The standardization process of deep learning technology in precision fertilization management for potatoes has not yet been established. There are significant differences between different studies and applications, thus making it difficult to share experiences and consequently affecting the promotion and application efficiency of the technology.

3.4. Deep Learning Models for Potato Price Forecasting

In the potato market, price is a crucial factor. Unfortunately, the imbalance between the supply and demand of potatoes leads to frequent and severe price fluctuations, thus causing numerous issues for both producers and consumers and even posing threats to food security and economic stability [172]. According to Jalonoja et al.’s report [173], price was shown to exhibit elastic responses to annual potato yield shocks, with a 10% increase in potato yields leading to a 20% decrease in market prices. Therefore, the development of scientific and accurate potato price prediction models is of paramount importance for guiding potato production.

Previously, researchers have applied various mathematical analysis methods to forecast the market price of potatoes. Zhang et al. [174] employed symbolic regression equations to construct a potato price prediction model from the perspective of supply and demand relationships. Additionally, based on [175,176], Jannat et al. [177] incorporated climate variables to build a mathematical model of potato supply and demand, thereby predicting its price. In subsequent research [178], they applied the nonlinear autoregressive distributed lag technique based on autoregressive distributed lag models to explore the impact of Bangladesh’s per capita GDP on its domestic potato price. However, traditional mathematical analysis has several limitations that severely constrain its development [179]:

• Neglect of multivariable factors: Traditional mathematical analysis usually considers only a few factors, while potato price is influenced by a complex interplay of multiple factors such as weather, pests and diseases, and market demand.
• Static nature: Traditional methods assume that market conditions and influencing factors remain constant. However, in reality, these factors are dynamic, which limits their accuracy.
• Linear assumptions: Traditional methods assume linear relationships between variables, whereas nonlinear relationships may exist in reality. This causes traditional methods to perform poorly when dealing with complex nonlinear relationships.
• High data dependency: Traditional methods rely heavily on data, and incomplete or inaccurate data can affect the reliability of the prediction results.
• Lack of real-time capability: Traditional methods cannot process and analyze data in real time, while market price forecasting requires quick responses to market changes. This leads to lagged and inaccurate predictions.

Deep learning can handle high-dimensional data, consider the interrelationships among multiple variables, simultaneously input various influencing factors (such as weather, policies, historical prices, etc.), and automatically learn the complex relationships between these factors. Moreover, deep learning possesses dynamic characteristics, which are capable of capturing nonlinear relationships within the data, thus allowing for nonlinear modeling. It can also process large-scale datasets and extract useful features. Furthermore, deep learning has strong real-time prediction capabilities, thereby enabling real-time data processing and prediction [180,181]. These advantages endow deep learning with great potential for application in potato price forecasting. Alzakari et al. [172] introduced an advanced LSTM–RNN model for predicting potato prices and compared it with traditional ML methods such as KNN and RF, as well as traditional mathematical analysis methods, thus validating the feasibility of deep learning in addressing the problem of potato price prediction. Notably, they collected a historical potato price database and other economic variables using the Z-score normalization method to ensure data consistency and reliability. Additionally, Eed et al. [182] proposed a method for predicting potato consumption based on stacked LSTM. The experimental results further confirmed that potato prediction models based on DL significantly outperform traditional ML methods.

In practice, due to factors such as the seasonality and production uncertainty of potato, accurately predicting potato prices is crucial. Avinash et al. [183] proposed a hidden Markov-guided deep learning potato price prediction model (HM-DL), which achieved high prediction accuracy and made significant contributions to potato price forecasting and agricultural production planning. Additionally, Ray et al. [184] proposed a model that combines random forest and CNN (RF–CNN) to predict wholesale potato prices.

Accurate and effective potato price prediction models can not only provide producers with efficient guidance for production strategies but also play a critical role in stabilizing potato prices, thus ensuring food security and addressing global hunger issues. Despite the successful application of deep learning in this field, there are still the following limitations:

• Deep learning models typically require a large amount of historical data for training. If the historical data on potato prices are insufficient or incomplete, the prediction performance of the model will be significantly compromised.
• Potato prices are influenced by various factors such as weather, pests, diseases, and market demand. If the input data contain noise or errors, the prediction results of the model will also be affected.
• Deep learning models are often considered black box models, thus making it difficult to interpret their internal working mechanisms.
• Potato prices may be affected by unexpected events (such as natural disasters or policy changes). These changes may not have clear patterns in the historical data, and deep learning models may perform poorly in coping with such dynamic changes.

We summarized the main applications of deep learning in the potato production chain, as shown in

Table 2. Summary of applications of deep learning in potato production chain.

Tasks	Metrics	Networks	Datasets	References
Potato diseases detection and diagnosis	99.75% (accuracy)	YOLOv5	PLD dataset and PlantVillage	[52]
	97.00% (accuracy)	ResNet-50	Self-constructed dataset (2152 images)	[57]
Potato pest detection and diagnosis	91.00% (accuracy)	Inception V3	Self-constructed dataset (495 images in 8 classes)	[62]
	88.07% (accuracy)	YOLOv3	Self-constructed dataset (976 images in 20 classes)	[65]
Potato health status monitoring	95.70% (precision)	Faster R-CNN	Self-constructed dataset (400 images)	[70]
	96.70% (precision)	ECLF-CS	PlantVillage	[85]
Potato yield prediction	90.80% (precision)	ResNet-101	Self-constructed dataset (230 images)	[96]
	95.20% (detection rate)	YOLOv5	Actual production trials	[101]
Potato product quality inspection	95.34–100% (accuracy)	CNN	Self-constructed dataset (5000 images in 5 classes)	[117]
	98.70% (accuracy)	RFCN ResNet-101	Self-constructed dataset (2770 images)	[120]
Potato irrigation strategy formulation	MSE < 0.0670 RMSE < 0.258	CNN–LSTM	Meteorological station data	[33]
Potato fertilization management	95.40% (prediction rate)	1D-AlexNet	Experimental data	[171]
Potato price forecasting	0.0920(MAE), 0.0145(MSE), 0.980(R2)	LSTM-RNN	Self-constructed dataset	[172]
Potato price forecasting	3.28(RMSE) 3.02(MAE)	RF–CNN	Self-constructed dataset (229 data points)	[184]

.

4. Discussion

4.1. Advantages of Deep Learning Technology in Potato Production

The application of deep learning in agricultural production is increasingly widespread, especially in various aspects of the potato production chain. Compared to traditional manual methods, deep learning can quickly process large amounts of data. Using deep learning models allows for rapid classification and analysis of a large number of potato images, thus far exceeding the speed of manual detection and analysis. For example, automated pest and disease detection systems can analyze large field data in seconds [185,186], while manual methods may take hours or even days. Although the initial development and cost of deep learning systems are relatively high, they can significantly reduce costs in long-term potato production operations [187]. For instance, utilizing drones and deep learning algorithms for crop monitoring can reduce labor and time input [188], thus lowering operational costs. Compared to ongoing payments required by manual methods, the maintenance costs of deep learning systems are lower. Additionally, employing deep learning can reduce the labor dependency of agricultural production, thus allowing agricultural workers to invest time and effort in other high-value-added tasks.

In comparison to traditional nondestructive detection methods such as spectral imaging and ML, deep learning has significant advantages in handling complex and high-dimensional data. When combined with spectral imaging technology, deep learning can more effectively extract useful information from hyperspectral data [189]. For instance, deep learning can automatically learn features from raw spectral data, whereas traditional machine learning methods require manual feature extraction [190,191]. Furthermore, deep learning models are typically trained on large-scale datasets, thus resulting in higher accuracy and robustness when applied to various processes in the potato production chain. In contrast, machine learning methods may be limited by feature selection and model complexity, thus leading to lower model accuracy. Unlike machine learning methods that require continuous expert involvement and adjustments, the advantages of deep learning models in learning the inherent structure and features of data without human intervention make them more capable of automation in the potato production process.

4.2. Challenges of Deep Learning Technology in Potato Production

Currently, many deep learning methods have been successfully applied in the field of pest and disease detection in crops such as potatoes. Deep learning can obtain good feature representations from large datasets through various linear and nonlinear transformations, thereby discovering relationships in complex data based on specific self-supervised and supervised learning approaches [192]. However, the limitations of DL have also been observed in practical applications: (1) Due to the unstructured nature of potato production environments, deep learning models applied in agriculture have high time complexity and a large number of parameters, thus requiring significant computational resources. (2) The potato production domain lacks datasets with both sample quantity and sample diversity, which may lead to overfitting and other issues during model training, thus restricting the application of deep learning in potato production processes. (3) The datasets currently used for crop pest and disease detection classification tasks include IP102 [193], PlantVillage [194], and the New Plant Diseases Dataset [195]. However, these datasets are not specifically designed for potato pest and disease detection, and the quality of the data needs further investigation, thus greatly affecting the generalization and robustness of potato pest and disease detection classification models and limiting the development of deep learning models for potatoes. (4) It is challenging to obtain scale images of plant diseases and pests in real field potato environments, and it is not possible to obtain images of multiple diseases and pests in one area.

It is worth noting that [196] proposed a method for automatically extracting knowledge graphs for potato pests and diseases, thus providing scientific support for the automated construction of large-scale knowledge bases for potato pests and diseases and laying a new foundation for data acquisition in the field of potato pest and disease detection. The authors in [197] constructed a dataset containing seven types of potato leaf diseases, including fungi, bacteria, and late blight (a total of 3076 images)—breaking the limitations of the current single dataset (PlantVillage) for potato disease detection tasks—thus promoting the progress of current research on potato leaf disease recognition.

In the field of potato yield prediction, deep learning has made significant progress and has been widely applied. However, due to external factors such as the unstructured and variable environment and the shortcomings of the models themselves [2,198], the application of deep learning in yield prediction is also limited:

• Significant differences exist in environmental parameters such as temperature and humidity in different regions, and the environmental parameters at different times in the same region are not the same, thus resulting in the insufficient generalization ability of existing deep learning models and making it difficult to transfer applications between different regions and different times in the same region.
• In addition, environmental data and field images need to be updated and supplemented regularly, and acquiring high-quality, comprehensive potato field data is extremely challenging. The scarcity of datasets will seriously affect the model’s capabilities.
• Deep learning models are complex, have strong black box characteristics, and have poor interpretability [199], which will greatly affect the formulation and application of potato planting strategies.
• Different potato varieties have morphological differences, and their manifestations of the same disease are not the same [200,201], which will affect the robustness of the model [90]; as well, relevant research is still lacking.

Deep learning has been applied to some extent in the fields of irrigation strategy formulation and precision fertilization management. However, there are still relatively few cases related to potatoes, especially in the area of precision fertilization. Currently, researchers use methods based on spectral imaging, ML, and manual experiments to explore reasonable irrigation and fertilization methods. Compared to manual and traditional nondestructive detection methods, deep learning has many advantages. However, deep learning still faces some common problems in practical applications, such as the large amount of data required and the insufficient interpretability of models. Therefore, in future research, we can try to address these issues and integrate deep learning with traditional methods to better optimize irrigation and fertilization management strategies.

In the field of potato price forecasting, deep learning has made significant progress and has been widely applied. However, deep learning-based potato price forecasting models typically require large amounts of historical data for training. The current potato price databases are still imperfect, with issues such as incomplete data, insufficient data points, poor data quality, and incomplete influencing factors. These problems severely affect the accuracy of existing price prediction models. Additionally, deep learning has a strong black box nature, thereby making it difficult for researchers to interpret its internal working mechanisms. Potato prices can also be affected by unexpected events, and deep learning performs poorly in coping with such dynamic changes.

4.3. Future Directions of Deep Learning Technology in Potato Production

Deep learning models still face the challenge of high model complexity. In future research, efforts should be directed towards rendering the models applied in various stages of the potato production chain to be more easily adjustable [53,62], thus serving as the foundational networks for other target detection techniques. This enhancement aims to improve model performance, thereby increasing the economic benefits of potato production. Specifically, in subsequent research, methods such as enhancing the depth and width of DL networks, adding attention mechanisms (such as Self-Attention, SE modules, etc.), introducing an automated model search function, improving optimizers, and incorporating regularization can be employed to improve the performance of potato deep learning models.

Furthermore, fertilizers and water are two interrelated environmental factors in the potato production process [202,203]. Future research should explore deep learning models capable of simultaneously integrating precise fertilization management and irrigation strategy formulation. This approach aims to ensure the stability of potato yields, enhance the utilization efficiency of water and fertilizers, and consequently improve the economic efficiency of potato production. In future studies on precise fertilization management of potatoes based on deep learning, it is advisable to integrate research on fertilizer application rates and crop growth modeling [204,205,206]. This integration aims to develop deep learning-based potato fertilization management models that align with the actual potato production needs of the local context.

In the potato production process, crucial stages such as planting, weeding, spraying, sorting, harvesting, and processing significantly impact the final yield of potatoes. Currently, researchers have developed some automated machines for these production processes, such as low-damage harvesters, plastic mulch planters, and potato–soil separation devices, to enhance production efficiency [207,208,209,210,211,212,213]. However, in potato production, mechanized equipment often causes mechanical damage to potato tubers, which severely affects potato yields and negatively impacts economic benefits. Deep learning also holds significant potential in optimizing mechanical structural parameters [214,215,216]. In future research, the integration of deep learning techniques for parameter optimization could be explored to further reduce potato damage rates and enhance economic benefits. Inspired by [217], which utilized LSTM to visualize the moisture content in potato slices during the drying process, future research can further develop deep learning models to visualize production processes at various stages of the potato production chain.

Currently, potatoes have developed into multiple varieties, each with different suitable processing methods. Accurately and timely identifying crop varieties is of crucial importance for effective crop management and yield prediction [218,219]. Based on this, Li et al. [220] proposed a potato breed detection method based on a spectral imaging analysis algorithm utilizing GADF image encoding combined with the ConvNeXt V2-CAP model. However, there is currently limited research on deep learning models for potato breed detection, which restricts its development and application potential in this field. In subsequent research, the DL technologies mentioned earlier in this paper can be integrated into or independently developed for potato breed detection algorithms, thus aiming to formulate more rational and scientific production strategies for potato in various regions. Although Guo et al. [221] combined Wilks’ Λ and partial least squares regression to propose a feature wavelength extraction strategy for hyperspectral information, thus successfully achieving the detection and classification of different potato breeds, this traditional nondestructive detection method still has limitations such as reliance on manual processes, insufficient accuracy, and poor real-time performance. Therefore, future research could integrate DL networks such as VGG and DenseNet with optimization algorithms like ABC to further enhance the robustness and generalization capabilities of detection models.

Inspired by the work of Péra et al. [222] on constructing a potato postharvest loss information system using database technology, in future research, deep learning can be integrated into database technology for data preprocessing, storage, real-time data processing, and database querying to develop loss information systems for various processes in the potato production chain. Additionally, potato price fluctuations are primarily influenced by factors such as climate and pest issues, production costs (including seed supply, labor intensity, production conditions, and yield potential), seasonal factors, policy factors, and geographical factors [172,174]. Future research could integrate the prediction results of other deep learning models within the potato production chain into potato price prediction models to explore more effective and robust forecasting models. Lu et al. [223] utilized the ensemble empirical modal decomposition model to investigate the characteristics of potato price fluctuation cycles and studied more comprehensive price influencing factors. In subsequent research, deep learning could be further applied to the study of potato price fluctuation cycle characteristics to identify additional price influencing factors, thereby constructing more comprehensive price prediction models and establishing a more extensive price prediction database. Notably, in the exploration of the optimal potato price prediction model, the findings of Adudotla et al. [224] differ from other studies, thus requiring further experimental validation to verify their authenticity.

Furthermore, DL exhibits significant advantages in handling complex and high-dimensional data, thus enabling the more efficient extraction of useful information from spectral data. Therefore, in subsequent areas of potato production such as disease and pest detection, plant status monitoring, yield prediction, and water–fertilizer management, deep learning can be combined with image detection technologies based on spectral imaging principles. By leveraging the rich information from hyperspectral and multispectral images, as well as harnessing the powerful image processing and data processing capabilities of deep learning, the complementary strengths of these approaches can achieve higher accuracy and production efficiency [225,226,227,228].

High-quality, comprehensive datasets are crucial for the application of DL in various areas of the potato production chain. Here are some common methods for constructing datasets [229]:

• Data collection: Gather data from public datasets such as Kaggle and OpenDataLab, or collect data through web scraping and surveys.
• Data cleaning: The raw data collected may have low reliability and require cleaning operations such as handling missing values, removing duplicate data, detecting outliers, and standardizing data.
• Data labeling: Use manual labeling or employ pretrained models for automatic labeling.
• Data augmentation: Expand image datasets through methods like rotation and scaling, and expand text datasets using methods like synonym replacement.
• Data splitting: Divide the dataset into training, validation, and test sets according to a certain ratio.

Notably, in our previous research [230], we constructed the GPID-22 dataset, which covers 199 different classes of pests and diseases with a total of 205,371 images. In future research, the data related to potatoes can be extracted from this dataset and supplemented with climate data, historical potato price data, and other relevant information to construct deep learning model datasets for various processes in the potato production chain.

5. Conclusions

The potato, as one of the major global food crops, possesses high yields and significant nutritional value, thus serving as a pivotal weapon against hunger. Over the past few decades, researchers have explored optimization methods for various processes in the potato production chain using manual methods, spectral imaging, machine learning, and other approaches. In the last decade, deep learning has been widely applied in potato production fields such as disease and pest diagnosis and detection, plant status monitoring, yield estimation, product quality inspection, and water–fertilizer management.

In the domain of potato disease and pest diagnosis, researchers have employed the powerful image processing capabilities of deep learning to detect and classify various diseases. However, the lack of high-quality potato disease and pest datasets currently hampers the generalization ability and robustness of models. Future research should focus on constructing high-quality datasets for different potato varieties, thus considering various manifestations of diseases and different physiological forms of pests to enhance model robustness, accuracy, and generalization ability. In the areas of potato plant status monitoring and yield estimation, deep learning has made breakthrough progress, yet existing research remains primarily theoretical, with limited practical application and insufficient practicality. Furthermore, due to the variability of environmental parameters and differences in planting conditions each year in potato production, frequent data updates are necessary. Subsequent research should involve the development of targeted applications and continuous model optimization to improve the transferability and practicality of deep learning methods. Potatoes have evolved into numerous varieties with morphological differences among them, and their responses to the same biological stressors vary. In the domain of potato product quality inspection, DL models have been widely applied and achieved significant success. However, existing research still lacks the preclassification of different potato varieties, and deep learning models deployed on harvesters suffer from insufficient real-time capability. Thus, future research should focus on developing preprocessing algorithms capable of classifying different potato varieties and improving model real-time performance. Irrigation and fertilization strategies have a decisive impact on potato yields, yet research cases using deep learning in this field remain scarce. Previous studies have mostly focused on exploring optimal irrigation and fertilization strategies using manual methods or traditional nondestructive detection methods such as spectral imaging and machine learning. In future research, integrating deep learning to enhance the efficiency of exploring optimal potato water–fertilizer management strategies is recommended. Additionally, combining deep learning with multispectral images, hyperspectral images, and other technologies can leverage the powerful learning capabilities of deep learning to supplement the deficiencies of traditional methods, thereby advancing the intelligent formulation of potato irrigation and fertilization management strategies. In the field of potato price prediction, although deep learning has made some progress, current research generally faces issues such as insufficient data volume, low data quality, and incomplete influencing factors. Moreover, there is a lack of highly credible and widely recognized databases. In subsequent research, researchers could focus on constructing high-quality, open-source potato price databases and integrating them with datasets from other areas of the potato production chain, thereby creating a comprehensive database for the entire potato production process. This approach would not only improve the accuracy of price predictions but also provide robust data support for the various stages of potato production.