Machine Learning based Recommender Systems for Crop Selection: A Systematic Literature Review

Crop selection (CS) is one of the most critical elements that affects the nal yield directly. As a result, selecting an appropriate crop is always a critical decision that a farmer must make, considering environmental factors. Choosing an appropriate crop for a given farm is a dicult decision including a plethora of variables that inuence the nal yield. Experts are frequently consulted to assist farmers with CS; but, as this alternative is time consuming and expensive, it is not available to many farms. The use of recommender systems (RSs) in agricultural management has recently brought some captivating and promising results. We propose a systematic literature review (SLR) in this article to nd and provide the most relevant and high-quality publications ad- dressing the crop recommendation (CR) question. The core concept of this SLR is inspired from the guidelines of PRISMA 2020.The different CR approaches are discussed, as well as all the most important input features for recommendation, which are determined and classied. We also identied some of the biggest hurdles to using CR in agriculture. Besides, we made an inventory of the most used techniques for CR. Further, we made an inventory of evaluation criteria and evaluation approaches. stand-alone crop recommending that detects soil quality and recommends a list of crops based on FL models. combination technique case- based signicant number of CF approaches when developing RSs are observed. Over half of the reviewed studies indicated that CF is the most used approach, with a stronger emphasis on a MB method. Perhaps, the availability of historical datasets of farmers linked to the marked dominance of CF in the last years.

used algorithms are: Top Popular, which recommends the top items (e.g., movies) with the highest ratings, and Product Association, which recommends the best combinations of items that are frequently bought together [5].
Customized approaches are other techniques that provide a recommendation to users based on ratings or content information. All these techniques are either focused on characteristics of an individual (or group) or characteristics of the product or service they are buying for example. In customized RS the following methods are used: Knowledge-based (KB) ltering is a technique that employs explicit knowledge to identify user's preference based on knowledge of items, user, and matching between both. It is speci cally effective in cases with less data on activity history of users [6].
RSs that are based on demographic information of users suggest a list of items that have good feedback from the users that are demographically similar [7]. The advantage of a demographic technique is that it does not require a history of user's feedback.
The vast majority of prior research had adopted the collaborative ltering (CF) approach, which employs a family of algorithms that calculates the utility and the rating of an item for a given user. Two general classes as illustrated in Figure.1 were suggested in literature [8], Memory-Based (MB) algorithms, where the predicted value is estimated as a simple linear combination of ratings and weights, either explicitly or implicitly. The weights can re ect distance, correlation, or similarity between either users or items, this similarity function is the hyper-parameter that affects mainly the prediction quality. Several authors adopted Pearson, Euclidean and Cosine functions as similarity criteria, others used the development of genetic algorithms to nd the most suitable combination of weight vectors [9,10]. Typical examples of this approach are neighborhood-based CF and item-based/user-based top-N recommendations.
MB algorithm are not always as fast and scalable as they are meant to be, especially in the context of actual systems that generates realtime recommendations based on very large datasets. To achieve these goals, MB RSs are used. MB CF involves building a model based on the dataset of rating and have been used as a "model" to make recommendations without having to use the complete dataset every time. This approach potentially offers the bene probabilistic perspective, either by the calculation of the expected value of a rating given the user's historical data, or using dimensionality reduction techniques, such as: matrix factorization (MF); in order to model the latent factor spa studies in literature have examined the possibility of exploiting deep neural networks (DNN) architectures [11,12,13,14], Convolutional Neural Networks [15,16], Recurrent Neural Networks [17] and Auto-Encoders [18,19]. To capture more complex and nonlinear relations above the ratings dataset, there are many MB CF algorithms. Bayesian networks, clustering models, latent semantic models, such as: Singular Value Decomposition (SVD), Probabilistic Latent Semantic (PLS) analysis [20], Multiple Multiplicative Factor (MMF), Latent Dirichlet Allocation (LDA) [21] and Markov Decision Process (MDP) based models that includes Contextual Bandits approach [22,23] and Reinforcement Learning (RL) [24].
Nonetheless, CF algorithms suffer from three common problems, namely: Cold start re ects the inability of suggesting recommendation to a new user or item due to data absence.
Sparsity problem occurs when available data are insu cient to identify similar users.
Scalability problem happens when the RS's performance and latency decrease drastically with an increase in the number of users and items of the system. And nally, One way to transcend these hassles is the combination of CF and Content-Based Filtering (CBF) methods, which are another family of algorithms that makes recommendations based on user preferences of product features, in a Hybrid Filtering (HF) technique [25,26].
Due to the successful use of RSs in various advertisement sectors, they have been applied to solve a variety of problems in the agricultural sector. In [27] authors proposed an ontology based on an RS that helps to identify the pests affecting a crop and their treatments. In [28] a cultivation calendar RS for wheat cultivation in Egypt based on climate data is developed. In another work [29], a hybrid technique for recommendation of agricultural products to buyers is used. In [30] a CF web-based RS was designed to provide help, such as: nancial help, irrigation facilities and insurance to the farmer's crops.
CS is one of the fundamental issues that have a strong in uence on farmer's revenue, and the applications of recommendation techniques has shown a signi cant progress recently, especially for these speci c tasks. Henceforth, the remainder of this analysis will focus on scienti c papers CR.

Research Formulation
To present a clear review of the recommendation techniques applied in agriculture, we followed the SLR protocol adopted in PRISMA [31], in its latest version (2020). In the following sections we formulate the Research Questions (RQs) which we will try to respond in this SLR, then, we explain the adopted search strategy for collecting papers, followed by the exclusion criteria that serve as a lter to select relevant papers for review, and nally in the data e phase, the information needed for the analysis of the selected papers is extracted.

1.Questions Formulation
In order to accomplish the objective of this SLR and get a full analysis of CR techniques, we de ned the following RQs: There are some synonyms that indicate CR Systems. In this SLR, we con-sider terms that replace "recommendation" by "selection" or "suggestion". To retrieve studies that use new techniques that are based on agricultural data, we used the terms "Arti cial Intelligence", "Machine Learning", "Deep Learning". All these terms were featured in the Search Query (SQ), which is presented as follows: Query = (Crop AND (Recommendation OR Selection OR Suggestion) AND ("Arti cial intelligence" OR "Machine Learning" OR "Deep Learning"))

Exclusion criteria (EC)
To strengthen the validity of the SLR, we considered only studies published between 2010 and 2020, and we maintained those that have no evaluation section due to the lack of research publications about the CR eld. We adopted the following EC: EC1. Studies must be peer-reviewed articles or proceedings.
EC2. Studies must be published in a conference, journal, press, etc.
EC3. Letters, notes, and patents are not included in the review.
EC4. Graduate reports are not considered for review.

EC5
. We considered only studies in English.
EC6. Studies that do not describe their proposed approach in a proper way were not considered in the review.

Data collection process
To answer the aforementioned RQs, data from the selected articles have been collected and structured. The extracted information focused on verifying whether

Paper reference
Year and type of publication The Indexing Database The country origin of study Models used to address the problem The inventory and description of crops and features used in each study Performance measures used to evaluate the proposed models

Results
This section presents the harvest of the selection process. Then, we present the information matrix along with results of each RQ.

Study selection
Based on the obtained results from the aforesaid scienti c databases, 89 results were identi ed. We followed a ltering process to eliminate the articles that do not match our selection criteria. We rst excluded 18 records because they were not indexed in well-known academic indexing services. Afterwards, we started scrutinizing the articles' content by reading the title, abstract and keywords, after that we kept 50 articles for further investigation, in addition to 2 more articles that were found in references. By full reading of the 52 articles of this study, we excluded another 12 articles, as they are either not clearly relevant, or are out of our scope. Thus, we ended up with 40 papers for synthesis and analysis. Fig 2 illustrates a ow chart of the paper ltering process.

Literature Review Matrix (LRM)
One of the major problems that the farmers face in the beginning of every agricultural season is the selection of a suitable crop that would produce a better yield. This process is usually done based on the farmer's experience, or with the help of an agronomist. Assisting this process was the objective of a huge [32] 2014 In this paper, a Features: Twenty-seven input criteria, namely s input (6 sub criteria), support, facilities, and risk [37] 2016 The proposed system is designed to predict the best suit-able crops for a given farm, and to suggest farming strategies, such as: mixed cropping, spacing, irrigation, seed treatment, etc. along with fertilizers and pesticides.
Crops: 44 crops that have been considered Features: Crop name, suitable rain-fall, temperature, cost, soil, and pH.
[38] 2016 The authors applied the majority voting technique using Random tree, CHAID, Knearest neighbors (KNN) and Naïïve Bayes as base learners for CR.
[39] 2017 Proposed a system that can detect the user's location then recommend top-k crops based on the seasonal information and crop production rate (CPR) of each crop of similar farms.
Crops: Not mentioned.

Features: Crop
Growing period database, Thermal zone database, Physiographic database, Seasonal crop database and CPR database. [40] 2017 In this article, an at-tempt is developed to predict crop yield and price that a farmer can obtain from his land, by analyzing patterns in past data.
Crops: Not mentioned.
Features: Crop areas, types of crop cultivated, n and the overall crops consumed.
Reference Year Contribution Dataset [41] 2017 In this study, a FL-based expert system is proposed to auto-mate the CS for farmers based on parameters, such as, the climatic and soil conditions. Crops: 20 crops.
Features: 23 features. [42] 2017 In this paper, a decisionmaking tool is developed for selecting the suitable crop that can be cultivated in each agricultural land.
Crops: Paddy, groundnut, and Features:26 input variables w [43] 2017 This paper develops a fuzzy based agricultural decision support system which helps farmers to make wise decisions regarding CS.
Crops: Not mentioned.
[44] 2017 This paper proposed two mathematical formulations, the rst one for the determination of crop-mix that maximizes the farmer's expected pro t, and the second model that maximizes the average expected pro t under a prede ned quantile of worst realization. [49] 2018 The article proposes a system that gives the farmer a prior idea regarding the yield of a particular crop by predicting the pro-duction rate according to the location of the farmer and the past data of weather conditions.
[50] 2019 This study proposed to design a KB solution for building an inference engine for recommending suitable crops for a farm.
Crops: Not mentioned.
Features: Elevation, temperature, fertilizer t rainfall, eld type, seed type and soil.

Reference Year
Contribution Dataset [51] 2019 A new datamining technique was proposed to cluster crops based on their suitability compared to soil nature of different areas.
Features: Soil, crop, tempe [52] 2019 In this paper, a multi-class classi cation-based decision model has been developed to assist the farmer in selecting suitable crops using rough, fuzzy, and soft set approaches.
[53] 2019 This study proposes an intelligent agriculture platform that manages and analyses sensors data to monitor environmental factors, which provides the farmer a better understanding of crop suitability.
Crops: Celery, water spinac Features: Temperature, h conductivity (EC), soil mois [54] 2019 This paper suggests the using of ANN and SVM for crop prediction considering the environmental parameters.
Crops: Not mentioned.
Features: Rainfall, minimu [55] 2019 The paper proposes a mobile application that will allow farmers to predict the region's production for a speci c crop. [63] 2020 The article proposes using hierarchical fuzzy model to reduce the classical system complexity with the huge number of generated rules.
Crops: Not mentioned.
[64] 2020 This work developed an application that helps selecting the most convenient type of crops in a certain zone considering the climate conditions of that zone, the production, and the needed resources for each crop. 3. How did research about CR evolved over time?
The uprise of new technologies in solving agricultural problems is an immi-nent fact. We can see in Figure 3a the distribution of selected studies over time, and we notice a remarkable increase in the number of studies related to CR from 2014 to 2020. Figure 3b illustrates the distribution of papers based on source database. Most of the selected papers were published in IEEE or Springer, and fewer papers were found in Wiley database. Furthermore, Figure 3c presents the proportion of each type of publication, nearly 62% of papers were published in imminent conferences, and about 31% comes from journal issues, while only 7.14% are book chapters, which enforce the quality of the publications included in the SLR. 4. What are the main techniques that were used in the literature for CR?
RSs are generally classi ed into three types: CBF, CF, or HF. The CF based model was used the most in our literature review, as discussed in section 2, CF tries to compress the entire database into a model, then performs its recommendation task by applying reference mechanism into this model. We identify two common approaches for MBCF: clustering and classi cation. Clustering CF assumes that users of the same group have the same interest, so they are partitioned into groups called "clusters". The authors in [61] proposed K-means clustering (KMC) algorithm, which is an unsupervised learning algorithm used to nd out fertilizers with NPK contents that are the nearest to the requirements of a speci c crop. First it calculates the required amount of the fertilizer, then the algorithm forms clusters of nearby fertilizers based on the Euclidean distance. Therefore, fertilizers in clusters with minimum distance are recommended to farmer. The recommendation task can be viewed as a multiclass classi cation problem, which uses a classi er supervised learning algorithm that maps the input data to a speci c output, a variety of these classi ers were tested on agricultural data. Each classi er provides its own set of class labels with an acceptable accuracy. The labels class of individual base learners are combined using the majority voting technique. The CR system classi es the input soil dataset into the recommendable crop type, Kharif and Rabi (Autumn and Spring). One of the most promising models in CF is FL, which extracts IF-THEN rules from the provided data using a membership function and linguistic variables that expresses the human knowledge. The authors in [49] proposed a fuzzy based model that uses 27 rules with 3 modalities: Low, Medium, and High. In this traditional single-layer fuzzy system, the rules are exponentially increased when the system's parameter increased, and a larger rule base will affect the system performance and transparency. Therefore, [63] developed a multi-layer system by using the fuzzy hierarchical approach. The hierarchical fuzzy model was ap-plied in the same Mamdani[1] fuzzy inference system for a suitable CR system. The CR system has 12 input variables, and it was decomposed into six fuzzy subsystems, then arranged by priority. The results show that a hierarchical CR system provides a better performance than a traditional fuzzy CR.
CFMB models have a low frequency in the reviewed studies, even though they represent the most common approach in RSs. Yield prediction is based on similarity relationships among items (farms or crops), in terms of collected production yield. For instance, [39] proposed a model consisting of calculating the PCS between farms using the information stored in the crop growing period database, the thermal zone database, and the physiographic database, then, they select the top-n similar farms. The seasonal information and CPR of each crop of the similar farms are used for ltering the rst appropriate list to the context. Finally, they recommend the top-k crops to each user respectively. Another study [49] used the similarity approach, in which they developed a system that gives the farmer a prior idea regarding the yield of a particular crop, by predicting the production rate. The COS measure is used to nd the similar farmers in terms of location from the database. Then, the resulting farmers that are like the querying farmer form the database for the fuzzy algorithm.
CBF, one of the most signi cant models in RSs, which are of a high importance for CR, as well as yield estimation. Examples of CBF model applications included in [66], the model of this study is based on the contents that use soil and crop properties, then suggests the list of ve high priority crops based on the corresponding properties between the crop and the land for matching soil properties. The algorithm takes two inputs, the land soil details and the re-quired property value for each crop. Primitively, the algorithm computes the similarity between the land and the crop, based on their properties to prede ne a range. If the comparison falls in a prede ned range, they generate a rank for the combination of crops and lands. In another study [51], authors developed a new data mining technique to cluster the crop based on the suitability of a crop against the soil nature of areas. Features are extracted from the datasets using ve different feature extraction metrics, such as, pH distance calculation, NPK (Macro nutrition distance calculation), MICRONUT (Micro nutrition distance calculation), water requirements, and temperature requirements. Then, the crops are clustered using hierarchical clustering based on the vectors into three groups, namely: most suitable, less suitable, or least suitable.
HF is another signi cant category of models used in CR. In the rst study [64], authors presented a new method, which is integrated within an IOT system, that is developed to advise farmers which crop type will generate more y clustering technique is proposed to the obtained groups that have been characterized by their weather conditions. The extracted knowledge forms the model and the rules engine. Finally, the RS generates an ordered list of crops that are suitable in descending order. In the second study [37], the authors developed a CR (hybrid) system, which utilizes FL to choose from 44 crop rules. The system is based on FL, which gets input from an ANNs based weather prediction module. An agricultural named entity recognition module is developed using conditional random eld to extract crop conditions data. Further, cost prediction is established based on a LR equation to aid in ranking the crops recommended. Table 3 shows how many studies describe an approach in each of the classes described earlier in section 2, as well as the studies themselves according to the approach category. As an outcome, a signi cant number of CF approaches when developing RSs are observed. Over half of the reviewed studies indicated that CF is the most used approach, with a stronger emphasis on a MB method. Perhaps, the availability of historical datasets of farmers linked to the marked dominance of CF in the last years. Figure 4 traces the timeline of publications, this latter con rms that CF with a MB method has a continual growth. The graph shows that there has been a slight increase in the two recent years in this eld and the number of studies is likely to increase after 2020.
Another important conclusion drawn from Table 3 and Figure 4 is the scarcity of research efforts focused on other ltering methods. Nevertheless, some studies showed that the CBF and HF give more accurate recommendations overall than all other types of ltering. However, throughout the years, the research pace on these types of ltering has been relatively low.  [1] First introduced as a method to create a control system by synthesizing a set of linguistic control rules obtained from experienced human operators. In a Mamdani system, the output of each rule is a fuzzy set.  ML models are data-depending models, without a constitution of high-quality training data, even the most performant algorithms theoretically will not give the expected results. Indeed, robust ML models can be useless when they are trained on inadequate, inaccurate, or irrelevant data. In the same con-text a wide variety of inputs were suggested in the reviewed articles, Figure 5 shows the classi cation of these parameters in six categories, viz.: Geography: This category of inputs indicates the agroclimatic regions, which is a land unit suitable for a certain range of crops and cultivars. Table 7 shows that 19 papers built their RS using geographic data among other variables which con rms the importance of this type of inputs, mainly because it works as an identi er that is unique to every farm.
Weather conditions (WCs): Weather plays a major role in determining the success of agricultural pursuits. For farmers, timing is critical in the obtainment of resources, such as: fertilizer and seed, but also forecasting likely weather in the upcoming season, informing on how much irrigation is needed, as well as temperature that can affect crop growth. These factors can be determined by recording hourly, daily, or weekly, temperature, rainfall, solar radiation, wind speed, evaporation, relative humidity, and evapotranspiration. In this SLR, WCs were used in 75% of the reviewed articles as Table 7 indicate.
Soil propriety (SP): All soils contain mineral particles, organic matter, water, and air. The combinations of these components determines the soil's quality, which depends both on its physical properties (texture, color, type, porosity, bulk density, etc.) and chemical properties (soil pH, soil salinity, nutrients availability, soil electrical conductivity etc.). Table 7 con rms that soil characteristics are the mandatory inputs on which researchers-built crop RSs.
Soil physical properties: Soil texture: Refers to the size of the particles that make up the soil and depends on the proportion of sand, silt and clay-sized particles and organic matter in the soil, it can in uence whether soils are free draining, whether they hold water and how easy it is for plant roots to grow.
Soil color: The surface soil varies from almost white through shades of brown and grey to black. Light color indicates law organic matter content while clave color indicates a high organic matter content.
Soil type: It describes the way the sand, silt and clay particles are clumped together. Organic matter (decaying plants and animals) and soil organisms like earthworms and bacteria in uence soil structure. it is important for plant growth, regulating the movement of air and water, in uencing root development and affecting nutrient availability.
Soil porosity: It refers to the pores within the soil. Porosity in uences the movement of air and water. Healthy soils have many pores between and within the aggregates. Poor quality soils have few visible pores, cracks, or holes.
Bulk density: It is the proportion of the weight of a soil relative to its volume. Bulk density is an indicator of the amount of pore space available within individual soil horizons and it re ects the soil's ability to function for structural support, water and solute movement, and soil aeration.

Soil chemical properties:
Soil pH: Soil reactivity is expressed in terms of pH and is a measure of the acidity or alkalinity of the soil. More precisely, it is a measure of hydrogen ion concentration in an aqueous solution and ranges in soils from 3.5 (very acid) to 9.5 (very alkaline). The effect of pH is to remove from the soil or to make available certain ions.
Soil salinity: It is the salt content in the soil; the process of increasing the salt content is known as salinization. Salts occur naturally within soils and water. Salination can be caused by natural processes such as mineral weathering or by the gradual withdrawal of an ocean. Soil Electrical Conductivity (SEC): It is an indirect measurement that correlates very well with several soil physical and chemical properties. Electrical conductivity is the ability of a material to conduct (transmit) an electrical current. As measuring soil electrical conductivity is easier, less expensive, and faster than other soil properties measurements, it can be used as a good tool for obtaining useful information about soil.
Crop propriety (CP): Some crops are very labor-intensive. Some crops require more skill than others. Some crops are riskier than others (high pro t if it's a good year but high chance of crop failure if the weather is bad), and some farmers are more able to cope with those risks. Each crop has its suitable amount needed of nutrients, optimal weather conditions and optimal soil properties. Unfortunately, there is no universal structure or data source for this kind of crop information, so researchers in different reviewed papers uses data mining techniques to extract knowledge from raw data, where FL shows high quality results, because of its rules generating model.
CPR: There are a lot of crop types produced in farms not all of them are suitable for producing in all areas. So, considering CPR of each one of them for every farm is very important to recommend and predict the crop productivity. Almost 90% of the reviewed papers are using supervised learning, where crop yield or crop pro tability, in ton/hectare or kg/hectare, were used as the dependent variable.
Market: Even with a high yield, decision about recommending the crop cannot be taken without knowing its price for the period of sell, as well as what is its cost. The price of a speci c crop is determined through demand/supply in the market; however, it can be predicted using historical data. In the other hand, cost can only be given by the farmer himself, ultimately it remains di cult to gather such data. Table 7 con rms this claim with just four papers using market information.
Several evaluation metrics have been used. Table 11 gives information about metrics used for evaluation techniques in the reviewed studies. This SLR restricted to the CS, which make classi cation metrics such as: Accuracy, Precision or Recall, the most popular performance metrics used in the studies of this SLR.
Accuracy is the proportion of true results among the total number of cases examined, which has the highest number of occurrences in our SLR, more than precision, which refers to the fraction of relevant recommendation among the retrieved crops, and recall, which refers to the fraction of retrieved recommendation among all relevant crops.
Another important result is the existing of studies evaluating their results by using regression error metrics such us: RMSE, MSE and the mean absolute error (MAE) metrics. The reason is that a study may use CPR as an output of the model developed, then, it chooses the crop which has the highest rate.
What is striking in the Table 11 is the remarkable amount of papers that are not evaluated by any performance criteria. The most likely cause of this result is the di culty to verify whether the recommended crop is truly the correct on this reason, most of studies seek the help of experts or farmers to judge the relevance of the suggested crops [50].  This section puts into terms challenges encountered in the extant research, these challenges are observed through three layers: the proposed algorithm, the used data and the evaluation preferences. Most studies have almost exclusively focused on the exploitation of CF algorithms for classi cation and clustering, more precisely: DTs, SVM, ANNs and NBC. Several gaps and shortcomings were identi ed in these techniques, namely the cold start problem, where the system cannot draw any inferences for users or items about which has not yet gathered su cient information. A closer look to the literature, reveals that these proposed adaptations are very classical, since the eld of RS has known a signi cant improvement due to entertainment companies, new algorithms were developed. For Instance, the Net ix Prize was an open competition for the best CF algorithm to predict user ratings for movies [72]. On September 21, 2009, the grand prize was given to the BellKor's Pragmatic Chaos team, which bested Net ix's own algorithm for predicting ratings by 10.06% [73]. During this competition MF became widely known due to its effectiveness, and important steps were taken in later years towards some very successful algorithms, which share the same basis of latent factor and user/item representations. Unfortunately, there was no article that suggested an implementation of these modern techniques for CR. A potential barrier that may face researchers in their quest to solve this delay is the available data, Figure 6 shows that more than 50% of the reviewed papers were published by Asian researchers, and more precisely Indian researchers, where SP parameters, CPR parameters and parameters for a very long period of time and by different states/districts, are available in government o cial websites. Notwithstanding, these data are inaccessible for foreign scientists. Furthermore, the proposed structure of data presents another challenge, where the vast majority of the well performed RS are tted by a user/item rating matrix, and the user's demographic data for hybrid systems, where ratings are either explicit or implicit (number of clicks, number of page visits, the number of times a song was played, etc.) So different complex preprocessing techniques are required. Even though, the sparsity remains a potential challenge to deal with. Finally, the high correlation degree that was observed for performance metrics and evaluations of the model, based on anonymous experts raises a very important question about the reliability of these results, hence the proposed algorithm.

Discussion
During the last decade, the use of technology to enhance the agricultural processes has been very remarkable. For CR, we can see clearly from Table 4.2 that there has been some success in this direction. Exploration of the used techniques shows us that most of the times, the problem is formulated as a classi cation problem, where algorithms as DTs and NBC, etc. give remarkable results. Fuzzy systems have been used in other cases, to model the uncertainty in input variables that were categorized and analyzed to facilitate the choice for future researchers. This study reveals also some challenges that are being faced when creating a CS method, like the unavailability of the data, more speci cally benchmarking datasets to compare the models, the input variables are quite different from a study to another, and the di culty in measuring the performance of the proposed methods; some papers compare their model predictions to what the farmer has actually cultivated, while others compare to domain experts recommendations. Historical evidence shows that great scienti c achievements were guided by industrial needs, unfortunately, precision agriculture and CR haven't yet gained the attention that it deserves from different stakeholders, basically in emergent countries where agriculture is the most valuable resource. Major improvements in agricultural domain will certainly appear by integrating the successful algorithms of RS that were developed for entertainment companies, whom humankind might bene t differently. Finally, in this paper we've studied 40 well selected articles from different reliable sources. Nevertheless, this number remains statistically insigni cant and more similar works are needed to illuminate the path for new researchers that are willing to innovate, achieve, and contribute to the eld.

Conclusion And Future Work
In this SLR we have presented a detailed analysis of 40 articles published from 2010 to 2020 about the CR problem as well as the main achievements, and current challenges. The SLR was conducted with the aim of providing insights into the kind of solutions that were proposed in the recent years for the CS task.
Such insights are valuable in suggesting new directions for research studies and in providing a good understanding of the recent research trends. For our feature work, we look forward to proposing new methods that are inspired from the current development of RS in other domains and tackling the present speci c challenges in agriculture.

Declarations
Ethics approval and consent to participate Not applicable

Consent for publication
We con rm our consent for the publication of the present paper.
Availability of data and materials Not applicable.

Competing interests
No, we declare that the authors have no competing interests as de ned by Springer, or other interests that might be perceived to in uence the results and/or discussion reported in this paper. Author's contribution The authors equally contributed.
Classi cation of recommendation systems techniques.

Figure 2
Flow chart of identi cation process of relevant CR articles.  Timeline of articles by recommendation approaches.

Figure 5
Input variables categories used for CR. Articles distribution by country.