In the SLR, a protocol was first created using (Kitchenham & Charters, 2007) popular review guidelines to ensure transparency, rigor, and reproducibility. The review protocol, including research questions, inclusion and exclusion criteria, search strategy, data extraction scheme, and quality assessment procedures, was defined before the search process. The protocol outlines the research questions to be addressed to limit the scope of the study before searching for relevant studies through selected databases. The databases used were ScienceDirect, Scopus, SpringerLink, Wiley, and Google Scholar because of their excellent scientific journals and trustworthy reference materials. A set of exclusion and quality criteria was used to screen and evaluate the pertinent studies. Subsequently, data from these studies were extracted and combined to answer the research questions.

The three steps and procedures used in the current study are planning, carrying out, and reporting the review (Figure 1). Planning the review, or the first step, involved deciding on research questions, creating a protocol, and testing the protocol to see if the chosen strategy was workable (Carver et al., 2013). A predefined review protocol was established to ensure methodological rigor, specifying the research questions, inclusion and exclusion criteria, search strategy, and data extraction procedures. This step started by determining the research questions that will direct the review’s scope and direction to guarantee that the study tackles pertinent and particular problems in the field. Another crucial task is creating a protocol that specifies the methodology and criteria of the review and guarantees an organized and methodical approach (Smith et al., 2011).

The protocol includes details such as the publication locations to be considered, the initial search string to find relevant studies, and the selection criteria for publications to help identify the most relevant and high-quality research. To determine whether the chosen strategy is workable and realistic, the protocol must be validated at this point. This entails examining the protocol to ensure that it is feasible within the parameters of the review and that it successfully gathers the required data. The planning stage lays a strong foundation for the later phases of the SLR by ensuring that the protocol is appropriate and comprehensive. Finally, updating the protocol considering preliminary results and input guarantees that it remains applicable and suitable for its intended use. This iterative process aids in improving the strategy by addressing any problems or gaps found during the initial planning. The planning phase is essential for defining precise goals, a strong methodology, and a verified protocol, all of which together guarantee the methodical and exhaustive character of the literature review (Snyder, 2019).

The primary objective of the second stage of the SLR is to select and examine relevant publications (Van et al., 2021). The first step in this process is a comprehensive search of all selected databases to identify studies that meet the predefined criteria. The objective of this study is to compile as many publications as possible that pertain to the research inquiries. The characteristics of the pertinent publications are extracted after they have been located. These characteristics include important information such as the authors, year of publication, publication type (journal article, conference paper, etc.), and other particulars relevant to the research questions. This step is essential because it facilitates the methodical organization and classification of the data. Following the extraction of these characteristics, the information is combined to find trends, patterns, and gaps in the current literature. Analyzing the gathered data to derive significant insights and draw conclusions is the process of this synthesis. The goal of this study is to provide a thorough summary of the current state of research on mapping food crops using remote sensing and ML and DL algorithms. By synthesizing the data, the review can highlight the most important studies, common methodologies, important findings, and areas that require more research. In the end, this phase produces an extensive synopsis of the state of the field’s knowledge, providing insightful analysis and direction for future research projects.

The research questions of the SLR last phase are documented by the findings and by offering thorough responses to the primary objectives (Shaffril et al., 2021). This phase is essential for converting the collected data and synthesized information into a meaningful and cohesive narrative that meets the main goals of the review. First, the results are meticulously recorded, frequently containing thorough explanations of the conclusions, pertinent information, and any observed trends or patterns. The information in this documentation is logically and clearly arranged to make it easy for readers to access and comprehend. Data can be successfully presented using visual aids, such as tables, charts, and graphs. The primary goal of this phase is to answer the research questions. By methodically examining the results, the review offers concise, fact-based responses to the initial research questions. This process entails connecting the synthesized data to each research question to ensure that every aspect is fully covered. The results of this phase greatly advance our knowledge of how well ML and DL algorithms work with remote sensing to map food crops. These insights can promote the development of knowledge by highlighting effective methodologies, filling in the gaps in the current research, and providing suggestions for new lines of inquiry.

Researchers, policymakers, and practitioners can greatly benefit from the documented results and responses. The review can help these stakeholders make well-informed decisions by offering evidence-supported conclusions. For example, researchers may find new areas for study, and policymakers may use the insights to create guidelines or allot funds for agricultural monitoring. All things considered, the last phase of the SLR ensures that the research findings are fully recorded, the research questions are fully addressed, and the acquired knowledge is successfully conveyed to aid decision-making and promote the advancement of knowledge in the field.

This SLR aimed to fully comprehend the effectiveness of the latest ML and DL algorithms for mapping food crops using remote sensing. Therefore, relevant studies were selected and analyzed from several dimensions to answer the following specific research questions:

All databases were searched on March 5, 2026, covering publications from January 2017 to March 2026. ML and DL have limitless areas of application, meaning that many published studies may not be within the scope of this review article. Consequently, the search began by narrowing the basic concepts to only relevant ones. In the first step, the words “machine learning” OR “deep learning” AND “crop mapping” were entered, and the search was conducted automatically on the five databases. Keywords were used to obtain a broad overview of the topic of interest. Then, the abstracts of the returned articles were examined to identify synonyms for the three keywords. After processing the initial search results and applying the exclusion criteria, a more complex search query was created to ascertain if any relevant studies were missed. The final search string was as follows: ["machine learning" OR "deep learning"] AND ["crop mapping" OR "crop classification" OR "crop identification" OR "crop monitoring" OR “crop type”] AND “remote sensing”: anywhere. Consequently, 1,589 published works were obtained by executing this search string.

Two reviewers conducted the screening process independently to minimize selection bias. The procedure consisted of three stages: title screening, abstract screening, and assessment of full-text eligibility. Disagreements between reviewers were resolved through discussion and consensus. The chosen research was evaluated using the criteria for exclusion to define the parameters of the systematic review to eliminate published works that were deemed unnecessary. The criteria for exclusion (EC) were as follows:

EC 1: This publication is not related to food crop mapping, ML, or DL.

EC 2: English is not used in the manuscript.

EC 3: Duplicates publications that have been retrieved from other databases.

EC 4: The full text of the publication is not accessible.

EC 5: The article is a survey/review.

Figure 2 illustrates how to choose and exclude articles from the database for a PRISMA review. The number of publications obtained from the five databases is displayed in Table 1 using the string search and selection criteria. Of the 1,589 published works captured using the string search, only 1,580 studies remained after screening using the first three exclusion criteria. With the five exclusion criteria, only 80 studies were retained for further analysis. Data were extracted from these selected studies and then combined and synthesized to determine whether the studies should be omitted or kept based on the exclusion criteria to answer the three research questions.

For many reasons, the Google Scholar, Springer Link, Science Direct, Scopus, and Wiley databases were carefully and strategically chosen. These databases provide thorough coverage and easy access to excellent peer-reviewed literature when combined. Google Scholar is a great resource for a variety of literature because it offers a vast array of scholarly publications, including books, theses, articles, and conference papers, from a range of disciplines. The extensive collection of scientific and technical content provided by Springer Link is particularly strong in technology, agriculture, and environmental science. Science Direct’s vast database of top-notch, peer-reviewed journals and articles in the domains of science, technology, and medicine ensures reliable and credible research studies. Scopus is one of the largest abstract and citation databases of peer-reviewed literature and is well-known for its exacting indexing standards and thorough coverage of research outputs, including papers, conference proceedings, and patents. Accessing comprehensive studies and reviews in agricultural science and remote sensing technologies is made possible by Wiley’s extensive collection of scientific, technical, medical, and scholarly research journals. These databases also cover a wide range of disciplines related to food crop mapping, such as agriculture, remote sensing, environmental science, computer science, and ML, to ensure a multidisciplinary approach. They also provide access to a large archive of scholarly publications, citation tracking, and sophisticated search features to facilitate effective and efficient literature searches. These databases allow researchers to conduct a thorough, multidisciplinary, and high-quality review of the literature on remote sensing and ML for mapping food crops, capturing the most impactful, up-to-date, and pertinent studies.

A structured data extraction form was developed. The following variables were recorded from each eligible study: publication year, study location, crop type(s), remote sensing platform (satellite/UAV), sensor type (optical, SAR, multispectral, hyperspectral), algorithm(s) used (ML or DL), and accuracy metrics (e.g., overall accuracy, F1-score, Kappa, IoU). This structured extraction ensured consistency and comparability across studies. The attributes of each of the 80 retrieved publications were analyzed to answer the research questions, as explained below:

Research Question 1: The analysis considered the type of platform (satellite, unmanned, or manned aerial vehicle) as well as the sensing technology (radioactive detection radar, Lidar, multispectral, and hyperspectral). When it came to satellite imagery, the name and attributes of the satellite were recorded to determine whether it was a suitable sensor for a given crop species.

Research Question 2: The analysis documented the model, basic algorithm, and crop species discussed or investigated in the literature. In publications where multiple models were used, an overall accuracy evaluation metric was created.

Research Question 3: This analysis explains the challenges faced by the authors and the potential solutions proposed or reported in each study.

A risk-of-bias assessment was conducted to enhance methodological rigor. Each study was evaluated based on the following criteria: Clarity of dataset description, Transparency of algorithm configuration, Reporting of validation strategy, Reporting of class imbalance handling, and Completeness of performance metrics. Each criterion was scored on a scale of 0–2 (0 = not reported, 1 = partially reported, 2 = clearly reported). Studies with insufficient methodological transparency were excluded from the quantitative analysis. This quality assessment ensured that only studies meeting minimum methodological standards were included in the final analysis.

The different findings gathered to address the research questions for the study are presented and discussed in this section. Distribution of selected publications by year (Figure 3). Most articles were retrieved from Scopus, Google Scholar, and SpringerLink. Additionally, the number of scientific publications on the mapping of food crops using ML and DL algorithms has been steadily rising annually. In 2022 and 2023, the number of publications increased significantly from 6 to 12 and 14, respectively, indicating that this topic received more attention. There are several reasons for the notable rise in publications in recent years about mapping food crops using DL and ML algorithms. First, the worldwide demand for sustainable and efficient agricultural practices has increased interest in innovative technologies that can improve crop monitoring and yield prediction. ML and DL algorithms provide comprehensive information about crop health, growth trends, and possible problems, such as pests or diseases. Large amounts of data from remote sensing and other sources can be analyzed using these methods (Wang et al., 2022).

Second, advances in processing capacity and the accessibility of sizable datasets have made it easier to apply these complex algorithms to real-world agricultural problems (Alibabaei et al., 2022). Easy access to high-resolution satellite imagery and other remote sensing data has also contributed to the expanding use of these technologies in agriculture. Furthermore, people are more conscious of the significance of food security, especially in light of climate change and population growth (Fonta et al., 2011). Researchers and policymakers are focusing more on creative solutions to guarantee consistent food supplies, and ML and DL are viewed as important facilitators in this endeavor. Furthermore, more funding opportunities and cooperative projects in this field have resulted from the recognition of the potential of these technologies to transform conventional farming methods (Altieri & Nicholls, 2017). Studies examining the relationship between artificial intelligence and agriculture have also gained more attention from conferences, workshops, and journals, which has increased the number of publications.

Figure 4 shows the spatial distribution of publications on ML and DL mapping of food crops. With 35 publications, China had the most pertinent works, followed by India (n = 13) and the United States (n = 5). Fewer than three studies have been conducted on this subject in other nations. China, India, and the United States are known to make significant investments in research and development, particularly in the fields of science and technology, because of their many research centers, top-notch universities, and government policies. Furthermore, as part of the “Made in China 2025” plan, China has boosted its investment in technology, including artificial intelligence, recently (Wübbeke et al., 2016).

Table 2 shows the profile of each selected publication, including the authors, algorithms used, crop type, sensors, and maximum overall accuracy (%). This information serves as an initial reference in selecting and determining the most effective algorithm for mapping different food crops when using certain sensors to produce the desired level of accuracy. The reviewed studies were reclassified into two analytical categories to avoid misleading comparisons across heterogeneous platforms: (1) satellite-based crop mapping and (2) UAV-based crop mapping.

Satellite imagery typically provides medium- to high-resolution data (e.g., 10–30 m for Sentinel-2 and Landsat) with broad spatial coverage, making it suitable for regional- and national-scale monitoring. However, in heterogeneous agricultural landscapes, medium-resolution imagery is more prone to mixed-pixel effects and spectral confusion. In contrast, UAV imagery offers centimeter-level spatial resolution, enabling highly detailed crop discrimination at the field scale. Nevertheless, UAV-based studies are often limited in terms of spatial extent and operational scalability. Because of these substantial differences in spatial resolution, scene complexity, and mapping objectives, the accuracy values reported for UAVs and satellite platforms are not directly comparable. Therefore, performance trends are interpreted within rather than across platforms.

Table 3 shows the frequency distribution of algorithms used to address the first research question (RQ1). The three most popular algorithms in food crop mapping studies were Random Forest (RF), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM). In comparison, few studies used alternative modeling, architectures, and algorithms. In ML, an ensemble learning method called RF builds a combination of the outputs of several decision trees to increase precision and resilience (Appiahene et al., 2020). Its main advantage is its capacity to manage high-dimensional data and deliver strong performance even with a high feature count. Furthermore, RF is less prone to overfit than individual decision trees. However, its main drawbacks are that it is computationally taxing, particularly when working with large datasets, and its complexity may make it more difficult to comprehend.

The SVM component of ML is a powerful classification technique that selects the most effective hyperplane to split the feature space into different classes (Pisner & Schnyer, 2020). SVMs offer reliable performance with both linear and nonlinear data through kernel trick extensions, making them extremely effective in high-dimensional spaces. The main advantages of SVMs are their high accuracy and capacity to manage intricate class boundaries. However, choosing the right kernel and hyperparameters can be difficult and require extensive tuning, and SVMs can be memory-intensive and slow to train, particularly with large datasets.

In DL, time-series data and sequences are modeled using a recurrent neural network (RNN) type known as long short-term memory (LSTM), which circumvents the vanishing gradient issue and captures long-term dependencies (Kim et al., 2018). LSTMs perform exceptionally well in tasks involving sequential data, including time-series prediction, language modeling, and speech recognition, because they can retain and apply long-term context. The main advantage of LSTMs is that they outperform traditional RNNs on temporal tasks. However, LSTMs require a significant amount of training time and resources and are computationally costly. Because of their intricate architecture, they can be difficult to fine-tune and require a large amount of training data to perform well.

RF, SVM, and LSTM performed well in various mapping and image analysis studies (Filho et al., 2020). For this reason, scholars commonly rely on algorithms with proven performance to investigate topics with similar contexts of application. Besides, there have been many examples of scientifically sound implementations and tools for RF, SVM, and LSTM. Several popular ML libraries and frameworks, such as scikit-learn for RF and SVM and TensorFlow or PyTorch for LSTM, are easy to use and enable straightforward implementations (Chaudhary & Kumar, 2022). Last but not least, RF and SVM tend to be easier to interpret, which might be one of the primary reasons behind their broad applications, particularly in food crop mapping that may involve stakeholders with varying degrees of technical expertise ((Belgiu, 2016); (Dang et al., 2021); (Noi, 2018)).

The algorithms in the order of performance in food crop mapping to answer RQ2 are presented in Fig. 5. Here, high performance or effectiveness is based on the level of accuracy of a model or an algorithm when applied several times by researchers in different locations using various sensors to distinguish between types of food crops. As shown in the table, U-Net was the most effective algorithm, with an average overall accuracy of 97.72%, followed by the value-guided perception model algorithm and one-dimensional convolutional neural network (Conv1D), with average overall accuracies of 95.73% and 92.96%, respectively.

Despite the promising performance reported in many studies, remote sensing-based crop mapping still faces several technical and operational challenges that influence classification accuracy and model generalizability. One of the most frequently reported limitations relates to sensor-related constraints, particularly the mixed-pixel problem in medium-resolution satellite imagery. A single pixel may represent multiple crop types or land-cover classes in heterogeneous agricultural landscapes, which complicates classification and reduces thematic accuracy. Moreover, spectral similarity among crop species during certain phenological stages often leads to confusion between crops with comparable spectral signatures, particularly when using optical multispectral imagery.

Another important issue concerns inconsistencies in validation strategies across studies. While some studies employ hold-out validation, others rely on k-fold cross-validation or independent test datasets. These methodological differences can significantly influence the reported accuracy metrics and limit the results’ comparability across studies. Furthermore, the growing use of UAV imagery introduces additional methodological trade-offs. UAV platforms provide very high spatial resolution, enabling detailed crop discrimination at the field scale. However, UAV-based approaches typically cover smaller spatial extents and require intensive data processing workflows, whereas satellite imagery provides broader coverage but often suffers from mixed pixels and lower spatial resolution.

To address the third research question (RQ3), every chosen publication was examined to identify the difficulties encountered by researchers in mapping food crops using ML and DL algorithms. In addition, improvements made or proposed to the algorithms or models were documented. Accurately classifying and differentiating crops can be made more difficult by the variability in spectral signatures caused by various crop types, growth stages, and environmental factors (Potgieter et al., 2021). Atmospheric factors, such as haze and cloud cover, can degrade the quality of remote sensing data, producing incomplete or erroneous imagery (Xia & Jia, 2022).

Another important factor to consider is temporal resolution; the revisit time of satellites may limit the frequency of imaging required to effectively track the growth and health of crops (Li et al., 2014). Integrating multiple data sources, such as combining optical and radar imagery, requires sophisticated processing techniques and expertise to extract meaningful information (Joshi et al., 2016). Spatial resolution is also a key consideration. While high-resolution images provide detailed information, their coverage is often limited and can be costly to acquire, whereas lower-resolution images might not capture the finer details required for accurate mapping (Zhao et al., 2020). Last but not least is the difficulty of ground-truth validation, which necessitates the labor-intensive and time-consuming process of gathering ground data to validate and calibrate remote sensing models. According to (Mura et al., 2015), overcoming these obstacles calls for the development of remote sensing technology, better data processing algorithms, and reliable techniques for combining various types of data.

The challenge also relates to improving algorithms and models used by incorporating variations in input parameters (such as NDVI and EVI) and applying them to a wider range of study areas with various types of food crops ((Dharma et al., 2022); (Wang et al., 2022)). The more algorithms used to address differences in meteorological, topographical, phenological factors, and crop practices, the better the resulting model will perform. However, this potential should be accompanied by efforts to increase the overall accuracy of the model (Aashi & Vema, 2025). Another challenge is the generally low average accuracy; more studies are needed to investigate the less explored features of SAR data, including time-series features, gray features, textures, and correlation coefficients, that can be used to monitor plant growth, specifically for modeling in cloudy or tropical areas (Tian et al., 2019). The use of DL models is challenging because it requires more combinations of optical and SAR image data to train appropriate architectures, including extracting and organizing thousands of spatiotemporal spectral features to improve plant classification ((Ajadi et al., 2021); (Chen et al., 2020); (Kordi & Yousefi, 2022)).

Many promising solutions can be generated using DL methods in mapping food crops with remote sensing. Advanced data fusion methods enable the integration of optical, radar, and multispectral images, contributing to a more complete description of the crop status and reducing atmospheric attenuation problems (Himeur et al., 2022). Deep image processing, such as CNNs, improves classification accuracy by learning complex patterns independently of variable conditions (Rawat & Wang, 2017). Temporal resolution challenges are effectively addressed by analyzing crop growth from multiple timestamps with LSTM networks (Rußwurm & Körner, 2020). The augmentation of existing datasets to enhance the model’s performance using synthetic data produced with GANs is also beneficial. Automatic feature extraction also minimizes the dependence on manual work and thus speeds up data processing. Cloud computing provides the necessary computing capacity for a scalable analysis approach for large datasets. Transfer learning (using pre-trained models) increases accuracy when labeled data are scarce. Strong ground-truth validation, supported by semi-supervised and active learning, accelerates training with fewer labeled examples. Finally, attention mechanisms and other techniques designed to maintain model explainability and interpretability should be utilized to make results interpretable and actionable for stakeholders (Shah & Konda, 2021). These methods improve the accuracy, reliability, and scalability of agricultural monitoring systems with ML and DL.

Based on the synthesis of the reviewed literature, several research directions can be identified to further advance the application of machine learning and deep learning in remote sensing-based crop mapping. First, developing foundation models for remote sensing represents a promising direction. These large-scale pre-trained models can learn generalized representations from massive datasets and may enable cross-regional transferability of crop classification models.

Second, explainable artificial intelligence (XAI) is gaining increasing attention in agricultural monitoring. While deep learning models often achieve high accuracy, their decision-making processes are frequently difficult to interpret. XAI methods can improve model transparency and support decision-making processes for agricultural management and policy development. Third, the integration of UAV and satellite data in real-time monitoring systems represents another important research direction. UAV imagery provides high spatial resolution for detailed field-level analysis, whereas satellite imagery provides consistent temporal coverage over large areas. The combination of these platforms could enable more comprehensive and timely crop monitoring systems.

Finally, future research should focus on the development of standardized benchmark datasets and validation protocols to improve comparability across studies. Harmonized evaluation frameworks would allow a more reliable assessment of algorithm performance and support the development of robust operational crop mapping systems. These research directions highlight the need for continued methodological innovation to improve the accuracy, scalability, and operational applicability of remote sensing-based crop mapping.