A Systematic Review of AI-Based Classifications Used in Agricultural Monitoring in the Context of Achieving the Sustainable Development Goals

Abstract

The integration of Artificial Intelligence (AI) into remote sensing data classification has revolutionized agriculture and environmental monitoring. AI is one of the main technologies used in smart farming that enhances and optimizes the sustainability of agricultural production. The use of AI in agriculture can involve land use mapping and crop detection, crop yield monitoring, flood-prone area detection, pest disease monitoring, droughts prediction, soil content analysis and soil production capacity detection, and for monitoring the evolution of forests and vegetation. This review examines recent advancements in AI-driven classification techniques for various applications regarding agriculture and environmental monitoring to answer the following research questions: (1) What are the main problems that can be solved through incorporating AI-driven classification techniques into the field of smart agriculture and environmental monitoring? (2) What are the main methods and strategies used in this technology? (3) What type of data can be used in this regard? For this study, a systematic literature review approach was adopted, analyzing publications from Scopus and WoS (Web of Science) between 1 January 2020 and 31 December 2024. By synthesizing recent developments, this review provides valuable insights for researchers, highlighting the current trends, challenges and future research directions, in the context of achieving the Sustainable Development Goals.

1. Introduction

Data-driven smart farming involves the use of recent technology developments, such as the Internet of Things (IoT) and Artificial Intelligence (AI) [1], and consists of a multidisciplinary effort to address the achievement of the Sustainable Development Goals (SDGs) [2]. These technologies include algorithm- and model-based tools [3] that require accurate geospatial data.

Agriculture 4.0 is characterized by the integration of advanced digital technologies to create a more efficient and sustainable farming sector. This implies big data analyses and the use of AI, the IoT and robotics to enhance the digitalization and automation of agricultural practices, improving productivity and resource management [4].

Various agricultural policies, such as the Common Agricultural Policy of the EU [5], promote the use of remote sensing as an appropriate control system for agriculture. Satellite images and aerial photographs, in combination with advanced algorithms and machine learning (ML) techniques, are used to assess if farmers are fulfilling the obligations outlined in the strategic plans [6].

Agriculture and vegetation are strongly interconnected, each influencing the other in ecological and economical aspects. Plant biodiversity can enhance ecosystem functions that benefit agriculture, while intensive farming can reduce vegetation diversity [7].

In addition to satellite and UAV (Unmanned Aerial Vehicle) images, LiDAR (Light Detection and Ranging) point clouds acquired through airborne laser scanning can contribute to the mapping and analysis of vegetation, as laser signals can penetrate through forest canopies [8].

The sustainable management of land has been a valid concern for a long time, being an issue especially relevant in the context of climate change [9]. In this regards, recent technological advancements have made the use of higher-resolution remote sensing data both feasible and cost-effective, significantly increasing their practical value [10].

Inefficient use of natural resources, such as soil and water, has been proven to cause economic losses and lead to environmental damage [11]. Remote sensing technologies can be valuable tools to support precision agriculture throughout different phases of crop production. Recent progress in sensor technologies, data processing and analytics has increased the accessibility of remote sensing in agriculture [12].

The accuracy of the methods applied in agriculture depends on multiple factors, including the image resolution and analysis technique [11].

Artificial Intelligence is increasingly being integrated into remote sensing applications for agriculture, providing efficient analysis techniques of large-scale satellite and UAV data. AI techniques, such as ML and DL (deep learning), are used for a wide range of tasks in remote sensing for agriculture, promoting agricultural reform [13]. The integration of big data into predictive models and decision-support tools allows stakeholders to mitigate the impact of climate change and other environmental threats [14]. Therefore, the use of AI in remote sensing for agriculture is transforming how data is analyzed, leading to more efficient, adaptive and cost-effective strategies applied across the entire agricultural field, while contributing to the achievement of the SDGs.

Based on the current pace of progress since their adoption in 2015, none of the 17 Sustainable Development Goals are expected to be fully achieved by 2030. Globally, SDG 2 (Zero Hunger) has been stagnating and faces major challenges (Figure 1) [15].

The SDGs are deeply interconnected [16,17], and the use of AI in remote sensing for agriculture offers a powerful tool to address multiple goals simultaneously. In Figure 2, the connections between SDG 6—Clean Water and Sanitations (a), SDG 9—Industry, Innovation and Infrastructure (b), SDG 12—Responsible Production and Consumption (c) and SDG 13—Climate Action (d) and the other 16 SDGs are highlighted. Access to clean water (SDG 6) is essential for achieving Zero Hunger (SDG 2), as agriculture depends heavily on water availability. AI models trained on satellite and UAV data can improve water resource monitoring and support precision irrigation, minimizing waste and increasing crop productivity. The link between SDG 2 and SDG 13 (Climate Action) is also relevant, as enhanced resilience to climate change can support more sustainable agriculture. AI-enhanced remote sensing enables the monitoring of climate variables—such as rising temperatures, changing precipitation patterns and the intensity and frequency of extreme weather events that profoundly affect agricultural fields. Furthermore, SDG 2 connects with SDG 12 (Responsible Consumption and Production), as the efficient use of natural resources supports efforts to increase agricultural productivity. AI-driven analytics can optimize resource use, reducing environmental impacts while improving crop yields. SDG 9 (Industry, Innovation and Infrastructure) plays an important role in enabling the technological innovation behind AI applications in agriculture. These innovations enable farmers to make data-driven decisions that contribute to food security.

Therefore, it is important to have a deeper understanding of how AI can transform agricultural monitoring, in order to achieve sustainable and resilient food systems. This systematic literature review (SLR) aims to synthesize the current research on AI-based classifications in agriculture and identify the knowledge gaps and opportunities for further research.

2. Materials and Methods

2.1. Selection Criteria

For this study, the papers that fulfilled the following criteria were selected:

• Addressed the topics of the research questions;
• Were published in peer-reviewed journals;
• Were published between 2020 and 2024;
• Were open access;
• Were written in English.

2.2. Information Sources

The methodology applied involved searching in the Scopus and Web of Science databases (

Table 1. Information sources.

Data Source	Type	URL
Scopus	Digital Library	https://www.scopus.com (accessed on 18 March 2025)
Web of Science	Digital Library	https://clarivate.com/academia-government/scientific-and-academic-research/research-discovery-and-referencing/web-of-science/ (accessed on 18 March 2025)

), as these met the peer-reviewed journals, use of English language and open-access criteria.

2.3. Search Strategy

The methodology used involved searching for three keyword groups (

Table 2. Groups of keywords.

Group	Keywords
Group 1	“Artificial Intelligence”, “Deep Learning”, “Machine Learning”
Group 2	“Remote Sensing”, “Image Classification”, “Point Cloud Classification”, “Agriculture”
Group 3	“Environmental Monitoring”, “Land Use Mapping”, “Flood-Prone Area Detection”, “Forest Monitoring”, “Pest Disease Monitoring”, “Droughts Prediction”, “Soil Content Analysis”, “Soil Production Capacity Detection”, “Vegetation Monitoring”, “Crop Detection”, “Crop Yield Monitoring”

). The first group of keywords included the types of algorithms that the research focused on, the second group focused on the domains of application for these algorithms and the third group reflected the specific focus of the studies that were analyzed.

Filters for the publication date (between 1 January 2020 and 31 December 2024), language (English) and access type (open-access papers) were also applied.

The expected results were papers that used AI, DL or ML techniques applied to remote sensing, image classification, point cloud classification or agriculture in general, that were used for agricultural and environmental monitoring-related tasks, such as land use mapping, flood-prone area detection, forest monitoring, pest disease monitoring, droughts prediction, soil content analysis and production capacity detection, vegetation monitoring, crop detection or crop yield monitoring.

For each search, certain search queries were applied (

Table 3. Search queries.

Digital Library	Search Algorithm
Scopus	(TITLE-ABS-KEY(“Artificial Intelligence” OR “Deep Learning” OR “Machine Learning”)) AND (TITLE-ABS-KEY(“Remote Sensing” OR “Image Classification” OR “Point Cloud Classification” OR “Agriculture”)) AND (TITLE-ABS-KEY(“Environmental Monitoring” OR “Land Use Mapping” OR “Flood-Prone Area Detection” OR “Forest Monitoring” OR “Pest Disease Monitoring” OR “Droughts Prediction” OR “Soil Content Analysis” OR “Soil Production Capacity Detection” OR “Vegetation Monitoring” OR “Crop Detection” OR “Crop Yield Monitoring”))
WoS	TS = (“Artificial Intelligence” OR “Deep Learning” OR “Machine Learning”) AND TS = (“Remote Sensing” OR “Image Classification” OR “Point Cloud Classification” OR “Agriculture”) AND TS = (“Environmental Monitoring” OR “Land Use Mapping” OR “Flood-Prone Area Detection” OR “Forest Monitoring” OR “Pest Disease Monitoring” OR “Droughts Prediction” OR “Soil Content Analysis” OR “Soil Production Capacity Detection” OR “Vegetation Monitoring” OR “Crop Detection” OR “Crop Yield Monitoring”)

). The search results showed 472 papers on Scopus and 309 papers on WoS.

2.4. Selection Process

The selection process (Figure 3) first involved removing the papers that were found in both databases. A total of 201 duplicates were identified and excluded. An initial screening was then conducted for the 580 remaining papers. The screening involved analyzing whether the papers were relevant to this research or not by reviewing their abstracts. During this step, 426 papers were excluded. The papers that were excluded fell outside the primary scope of the research, addressing topics such as wildfires which, while related, belong to different areas of study. The next stage involved full-text assessment of the remaining 154 papers. This included sorting the papers into categories and analyzing the data, methodology and algorithms that were used for each study. Following this in-depth evaluation, 46 additional papers were removed that were not relevant to this research. Although AI and remote sensing were mentioned in these papers, they were not applied to data classification or analysis, with some studies relying on soil samples instead of remote sensing data. Finally, 108 papers were included in this review. PRISMA Checklist is available in Supplementary Materials.

During the selection process, an increased interest in the research topic (Figure 4) was noticed; the number of papers published by year increased by 537.5%, from 8 papers in 2020 to 51 papers in 2024.

The geographic distribution of the study areas (Figure 5) showed a significantly higher number of case studies conducted in China, followed by a considerable number of papers where the study area was not specified. Countries such as Brazil, USA, Germany, France and Finland showed a moderate level of research activity in this domain. In particular, the map in Figure 6 highlights a clear lack of studies in Eastern Europe. This finding is especially relevant as Eastern Europe is renowned for its strong agricultural sector, particularly in countries like Ukraine, Romania, Hungary and Bulgaria, as they possess extensive arable lands and have a long-standing tradition of farming, contributing significantly to both regional and global food production. In 2023, the main cereals-producing member states of the European Union, alongside countries like Poland, Germany and France, were Romania, Bulgaria and Hungary [18]. Despite the importance of agriculture in these countries, there appears to be a gap in the research focused on the integration of advanced technologies into the agricultural sector.

3. Related Work

While existing studies have addressed individual challenges in agriculture and smart farming (

Table 4. Related work.

Main Topic	Focuses on Problems Related to Agriculture	Focuses on Problems Related to Vegetation and Forests	Focuses on Types of AI Algorithms for Data Extraction	Focuses on Available Data Types	Focuses on Data Sources	Reference
Applications of IoT in smart farming	✓	x	x	x	x	https://www.sciencedirect.com/science/article/pii/S1364032123007165 (accessed on 9th July 2025) [1]
Applications of IoT in agricultural sector	✓	x	x	x	x	https://www.sciencedirect.com/science/article/pii/S0301479721015504 (accessed on 9th July 2025) [3]
The use of remote sensing in agriculture	✓	x	✓	x	✓	https://www.sciencedirect.com/science/article/pii/S0167739X24006551 (accessed on 9th July 2025) [20]
The use of remote sensing in agriculture	✓	x	x	✓	x	https://www.sciencedirect.com/science/article/pii/S0034425719304213 (accessed on 9th July 2025) [10]
Water resources use and management in agriculture	✓	x	x	x	x	https://www.sciencedirect.com/science/article/pii/S0378377425000617 (accessed on 9th July 2025) [19]
The use of remote sensing and machine learning in agriculture	✓	x	✓	x	x	https://www.mdpi.com/2073-4395/14/9/1975 (accessed on 9th July 2025) [13]

), focusing on IoT technology [1,3], on the models and algorithms used, remote sensing for water resource management and smart irrigation [19,20], nutrient testing and scientific fertilization, disease and pest monitoring and ecosystem balance, crop-type classification and weed control, development stage monitoring and yield prediction, ref. [20] and evaluation of remote sensing platforms and sensors [10], none of these studies have integrated all the relevant aspects related to agriculture and vegetation, such as the data types used, types of AI algorithms for data processing and the sources of data, which defines its spatial resolution. This study fills this gap by offering a comprehensive review that takes into consideration all these relevant aspects.

Other reviews fail to comprehensively address issues regarding the use of AI and remote sensing in agricultural monitoring, as they are generally application-specific and do not compare how different domains—such as land use mapping and crop detection, crop yield monitoring, flood-prone area detection, forest and vegetation monitoring, pest disease monitoring, droughts prediction and soil content analysis and soil production capacity—overlap or diverge in terms of methods and resources. Furthermore, existing work rarely examines in an integrated way the data types employed (satellite imagery, UAV data), the AI algorithms used for data processing, the software products implemented and the data sources. By contrast, our review systematically considers all these aspects side by side across the main categories of agricultural and vegetation studies, allowing us to highlight common trends and methodological preferences.

4. Results

4.1. Answer to the First Research Question (“What Are the Main Problems That Can Be Solved Through Incorporating AI-Driven Classification Techniques in the Field of Smart Agriculture and Environmental Monitoring?”)

Technologies like ML, DL, Data Science and the IoT help select the optimal crops for specific regions by identifying threats, such as pests and diseases, and can be applied across several agricultural phases, such as soil preparation, land management and environmental factors monitoring, water monitoring, pesticide recommendation and crop damage assessment, to enhance the overall productivity of crops and reduce costs and resource waste [21].

Therefore, for this study, the papers were divided into the following categories (Figure 7):

• Land use mapping and crop detection;
• Crop yield monitoring;
• Flood-prone area detection;
• Forest and vegetation monitoring;
• Pest disease monitoring;
• Droughts prediction;
• Soil content analysis and soil production capacity detection.

Figure 7. Number of papers by research category.

Figure 7. Number of papers by research category.

4.1.1. Land Use Mapping and Crop Detection

According to Food and Agriculture Organization of the United Nations (FAO) [22], in 2021, agricultural land occupied about a third of the world’s total land area. In an effort to achieve the SDGs proposed by the United Nations [23], specifically Goal 2—Zero Hunger, some of the targets that were set were to double agricultural productivity by 2030 and maintain the genetic diversity of seeds and cultivated plants [24]. In order to achieve these goals, it is necessary to have an efficient monitoring system for land use and crops, through the use of remote sensing and Artificial Intelligence. Numerous studies have been published that assess the efficacy of involving AI and IoT in agricultural practices.

4.1.2. Crop Yield Monitoring

Crop yield monitoring is an essential part of precision agriculture. It helps farmers take immediate action according to the crops’ needs, to increase productivity, reduce waste and make further data-driven decisions, contributing to the achievement of SDG 12 (Responsible Consumption and Production). Moreover, AI models can be used to predict crop yield, enabling better resource use planning and early detection of potential issues [25].

4.1.3. Flood-Prone Area Detection

Agriculture is highly vulnerable to the impact of climate change and the hydrological disasters associated with it. Rapid flood impact assessments can lead to better emergency response actions [26]. Periodic flood mapping can help develop flood control measures and enhance flood management, contributing as well to agricultural yield estimation [27]. Moreover, in dryland areas, farming systems can rely on supplementary water derived from floods [28].

4.1.4. Forest and Vegetation Monitoring

In a study by Chavarria et al. [29], it was proven that forests have a positive impact on the performance of neighboring farms and their productivity. There is evidence that forest management practices favor the preservation of species in agricultural lands [30]. Sustainably managed forests contribute to reducing soil erosion and minimizing the risk of landslides, ensuring a balanced water cycle, which is essential for agricultural development and food security [31]. Therefore, forests play an essential role in preserving the natural environment’s biodiversity and ecosystem.

4.1.5. Pest Disease Monitoring

Pests are the main biological threat to agricultural and forestry production [32] and identifying areas of insect damage from remote sensing data through AI models is an important part of forest and crops health monitoring [33]. Pest and disease monitoring is a fundamental practice in smart agriculture, aimed at protecting crops and ensuring food security. It involves the detection, identification and tracking of harmful organisms that can reduce crop yield or quality.

4.1.6. Droughts Prediction

In regions characterized by water shortages and frequent drought conditions, a forecasting system for water irrigation demand needs to be developed. Such a system integrates remote sensing techniques and a soil water balance model to serve as a monitoring tool and support water resources management [34]. Comprehensive studies examining the impacts of changing water use patterns resulting from agricultural expansion, particularly increased irrigation, are still limited [35].

4.1.7. Soil Content Analysis and Soil Production Capacity Detection

Accurate soil moisture estimation plays an important role in enhancing agricultural productivity and supporting effective environmental management [36]. Assessment of soil macronutrients [37] and soil organic carbon stocks [38] can also provide a comprehensive understanding of soil health, leading to more informed decisions in smart agriculture and environmental management.

4.2. Answer to the Second Research Question (“What Are the Main Methods and Strategies Used in This Technology?”)

After analyzing the AI algorithms used in the selected papers, they were split into the categories shown in Figure 8. Decision trees-based (DTs) models were the most frequent algorithms found in this SLR, with Random Forest (RF) and other derived algorithms, such as XGBoost, CART (Classification and Regression Tree), GTB (Gradient Tree Boost), Cubist or IF (Isolation Forest), found in 55 papers. A DT-based model is a supervised ML method used for classification and regression tasks. It builds a tree-like model of decision, guiding the algorithm to classify the data or predict a value by splitting the data into smaller subsets until the predefined criteria is met [39]. Deep Neural Networks (DNNs) are supervised algorithms that use labeled data and are structured in three steps. The input layer takes in a file as the data, the hidden layers pass the information from one layer to the next and the output layer produces the final result. Convolutional Neural Networks (CNNs) are a type of DNN that specialize on data with a grid-like structure, particularly images [40], and were the most used type of DNN found in this SLR. CNNs, such as DenseNet, ResNet, TempCNN, ATCNN, SLMFNet, Deeplab V3+, SegNet, U-Net, U-Net++, AlexNet, DABNet, PSPNet and other DNNs were found in 49 papers. Support Vector Machine (SVM) is a supervised ML algorithm used for classification and regression tasks. It tries to find the best boundary, known as the hyperplane, that separates the different classes in the data, with the main goal of finding the optimal boundaries between classes [41]. SVM was found in 19 papers. A significant number of papers in which hybrid models were used were also found. These models, found in 13 papers, combined different AI techniques to overcome the limitations of using a single AI method. Statistical models, such as Auto-ARIMA, GLM (Generalized Linear Model), GNB (Gaussian Naïve Bayes), LASSO (Least Absolute Shrinkage and Selection Operator), LR (Linear Regression) and NB (Normal Bayes), were found in nine papers. Among other classic supervised learning models, kNN (k-Nearest Neighbor) was the most used algorithm. It finds the closest training examples to a new, unclassified case and assigns it the most common class among those neighbors [42]. Recurrent Neural Networks (RNNs) were found in seven papers. They are DL algorithms designed for processing sequential data like time series, text or speech [43]. Transformers- and attention-based models can prioritize the center of attention and focus on the key features according to their degree of importance in complex scenarios [44]. Algorithms such as BETAM (Border-Enhanced Triple Attention Mechanism) were found in five papers in this SLR. The Fully Connected Neural Network (FCNN) MLP (Multilayer Perceptron), a type of ANN (Artificial Neural Network) with multiple layers of interconnected neurons [45], was found in four papers. Clustering algorithms, such as K-means, are unsupervised algorithms used to group similar data points into clusters by identifying the patterns and structures within the data [46], and were found in two papers. Optimization algorithms that update the parameters throughout the learning process, such as OG-WOA (Optimal Guidance-Whale Optimization Algorithm), which is applied to increase the exploitation during the search process and select the relevant features for classification [47], and QHPO (Quantum Hippopotamus Optimization algorithm), which selects the best features for the final classification using NN (Nearest Neighbor) classifiers [48], were found in two papers. Lastly, Generative Adversarial Networks (GANs), with the potential for reconstructing high-resolution images from low-quality data through a process known as super resolution [49], were found in two papers.

The implementation tools used for algorithm development and data processing were also analyzed, as illustrated in Figure 9. Python programming emerged as the most commonly used tool, appearing in 45 papers, either for training the models or for processing the data. Notably, the PyTorch open-source deep learning framework [50] was found in the majority of these papers. Google Earth Engine (GEE) was the second most frequently used platform, mentioned in 23 papers, followed by ArcGIS (Pro or Online) in 21 papers. Other notable software products included Agisoft, QGIS and SNAP (Sentinel Application Platform), each found in 10 papers. R programming was used in eight papers, Pix4D in five papers, ENVI and Matlab computing in four papers, Google Earth Pro in three papers and OTB (Orfeo Toolbox) in two papers. Other software/frameworks found in only one paper each were DrakNet Framework, C++ programming, Go programming, Grass GIS, Iota2, MODIS Reprojection Tool, SAS, Minitab 18, Sen2Cor, OpenStack, CentOS Linux programming, GDAL, ArcMap and SCP (Semi-automated Classification Plugin).

4.2.1. Land Use Mapping and Crop Detection

Out of the 44 papers in this category, only 7 were dedicated to crop detection, the other 37 being related to land use and land cover mapping. Most of them implemented DT-based algorithms. Hamidi et al. [51] proposed a method that uses an Optic-SAR (Synthetic Aperture Radar) fusion for crop mapping based on remote sensing imagery, using RF, SVM, SAE (Sparse Auto-Encoder) and GFSAE (Guided Filtered Sparse Auto-Encoder) AI algorithms and methods, to classify crops such as corn, pea, canola, soybean, oat and wheat. In a study by Snevajs et al. [52], RF was also used for crop classification based on remote sensing imagery. Gackstetter et al. [53] used RF, RNN and a transformer model for crop-type mapping based on satellite imagery. In a study by Di Tommaso et al. [54], RF was used for mapping tall and short crops based on satellite images and LiDAR data. Zhang et al. [55] used DSSGT (Dilated Spectral–Spatial Gaussian Transformer Net) to map corn, cotton, sesame, broad-leaf soybean, narrow-leaf soybean, rice and water crops. In a study by Yue & Tian [56], crops and soil were evaluated with a DT-based algorithm, RF. A specialized RF model—CRRF (Crop-Residue Random Forest)—was also proposed by Yue & Tian [56] for classifications in vegetation–soil ecosystems for crop monitoring. Alotaibi et al. [57] proposed a DTODCNN-CC (Dipper-Throated Optimization with Deep Convolutional Neural Networks-based Crop Classification) approach, applied to satellite imagery, to accurately classify crops into the following classes: maize, banana, legume and other.

Most studies in this category focused on land use and land cover classifications. For this challenge, DT-based models were preferred, with RF being the most used algorithm [58,59,60,61,62,63,64,65,66,67,68,69]. Other algorithms based on decision trees were XGBoost [70,71], CART and GTB [72].

CNNs and DNNs were the second most preferred. Rajesh et al. [73] used a CNN for land cover/land use classification from satellite imagery. In other studies, multiple types of CNNs and DNNs were used, such as MSFA-Net [74], Inception-V3 and DenseNet121 [75], ResNet and ResNet50 [47,48,76,77], TempCNN and ATCNN [70], SLMFNet [78], Deeplab V3+ [44,79], FCN and SegNet [80], UNet and UNet++ [79,81], AlexNet [71], PSPNet and DABNet [79], CMFNet [82] and other CNNs and DNNs [42,83,84,85,86,87].

In a study by Lemenkova [88], an SVM algorithm was used to map the land cover dynamics.

Hybrid models that combine different types of algorithms were also found in the literature. Zhang et al. adopted an approach based on an SACA (Semantic-Aware Contrastive Adaptation) framework for land cover classification [89]. Dantas et al. [70] proposed the use of an REFeD deep learning framework (data Reuse with Effective Feature Disentanglement) for land cover mapping, which incorporates an effective supervision scheme, reinforcing feature disentanglement through multiple levels of supervision. A KD-KNN algorithm was used by Mourtadha et al. [59] to perform land cover and land use classification on satellite imagery. Another hybrid model was Swin-UNet, a semantic segmentation network model, based on the Swin-Transformer block, used by Hao et al. [80] to classify land use based on remote sensing imagery. MCSGNet is a network that solves the issues of low precision and poor generalization ability, proposed by Hu et al. [76] for classifying land cover types from remote sense imagery. Vinaykumar et al. [47] proposed the OG-WOA-Bi-LSTM (Bi-directional Long Short-Term Memory) technique, an advanced satellite image classification approach that combines deep feature extraction, optimization algorithms and sequence-based classification to effectively reduce overfitting in land use and land cover classification.

Other classic supervised learning models, such as kNN, were also found in the literature. In a study by Souza and Rodrigues [61], kNN was used to classify satellite imagery into different land use and land cover classes. The MD (minimum distance) model is another classical classification model found in the literature [72].

RNNs, which are DL models designed for sequential data and are able to maintain a memory of previous inputs, are suitable for analyses where the order and context of data are very important [90]. In a study by Nigar et al. [83], a comparison of RF, CNN and RNN models was performed for land use and land cover classification. The results showed that the RNN model provides an accuracy similar to that of the CNN model.

Optimization algorithms were also found in the literature. A study by Albarakati et al. [48] proposed the use of QHPO optimization to select the best features for final classification of land use and land cover.

Transformer algorithms and attention-based models, such as BETAM, are also used for land cover classifications based on remote sensing images. BETAM consists of a combination of three forms of attention and was used by Wang et al. [44] to perform semantic segmentation for land cover classification.

In a study by Aljebreen et al. [85], the statistical model NB was applied to the training dataset, along other algorithms (SVM, RF, DT and ANN) to perform land use and land cover classification using PlanetScope multispectral imagery.

Clustering algorithms, such as K-means and the supervised classification model LVQ (Learning Vector Quantization), were used by Ouchra et al. [91] with Google Earth Engine for land cover classification.

The overall results of this analysis are found in Figure 10.

For land use mapping and crop detection, Python programming was the most commonly found environment for algorithm development and implementation (Figure 11). ArcGIS was also preferred, found in nine papers, while GEE software, dedicated to land surface monitoring, was found in eight papers. SNAP software was used in six papers and QGIS was chosen in five cases. Matlab Computing and R Software (programming) were each mentioned in three papers. As for the less frequent trends, Google Earth Pro and OTB were used in two papers and Agisoft and ENVI were mentioned only once each. Seven papers used other software products, while in two cases no processing software was specified.

4.2.2. Crop Yield Monitoring

DT-based algorithms were the most used for crop yield monitoring (Figure 12). Algorithms like RF were used to estimate millet FCover [92], to perform image preselection to eliminate unsuitable images for crop monitoring [93] and to predict the yields of corn, rice and soybean [94].

A CNN was used in a study by Sabo et al. [95] to forecast crop yield for barley, durum and soft wheat crops. ConvLSTM, an RNN, alongside a 3D-CNN, a fully convolutional architecture, were used by Nevavuori et al. [25] for crop yield monitoring and prediction. SVM algorithm was used by Lv et al. [96] in a study related to winter wheat fAPAR (fraction of Absorbed Photosynthetically Active Radiation) modeling.

For crop yield monitoring, five software applications/environments were found in this SLR (Figure 13). Among these, the programming language Python was used in four articles. Interestingly, Pix4D was mentioned in two papers, while Agisoft, QGIS and R software (programming) were each used in one paper.

4.2.3. Flood-Prone Area Detection

Few studies were found related to flood-prone area detection. CNNs and other DNNs, transformers- and attention-based models, decision trees-based models and Support Vector Machine models were found in one paper each. In a study by Wu et al. [97], floodwater extraction based on UAV imagery was performed using a transformer model. Pech-May et al. [98] classified flooded areas using satellite images and a U-Net neural network. Atchyuth et al. [27] used RF, CART and SVM models to map flood inundations.

For flood-prone area detection, three software products were found. ArcGIS was the most preferred, found in two articles, while Python programming and GEE were found in one study each.

4.2.4. Forest and Vegetation Monitoring

In most of the papers regarding forest and vegetation monitoring, CNNs and other DNNs were the most used algorithms (Figure 14). Xiang et al. used U-Net, ResUNet and TeranusNet to monitor forest changes using UAV data [99], and Ecke et al. used EfficientNet CNN architecture to monitor forest health based on UAV data [100]. A TempCNN model was used by Perbet et al. in a study regarding boreal forest disturbances monitoring [101]. A mask R-CNN (Mask Region-based Convolutional Neural Network) model with a ResNet-50 backbone was used by Lucas et al. to detect individual trees from aerial imagery [102]. SegForest was proposed as a segmentation model for remote sensing images by Wang et al. to estimate forest area from satellite imagery [103]. Slagter et al. used a standard CNN model to monitor tropical forest disturbance from satellite data [104]. Retallack et al. used various CNNs to detect an indicator arid shrub in UAV imagery [105], Guo et al. used an SNN (Siamese Neural Network) for forest cover change extraction from satellite images [106] and Lahssini et al. [8] used a CNN, along with other algorithms, to estimate the basal area and wood volume in forests based on LiDAR and satellite data. In a study by Mäyrä et al., a workflow for tree species classification from LiDAR data was presented, and it showed that CNN models outperform other algorithms, such as RF, GBM and SVM [107]. He et al. proposed an OUDN (Object-based U-Net-DenseNet-coupled) network for intelligent mapping of urban forests from remote sensing imagery [108]. Li et al. proposed an MSCIN (Multifeature Synthesis Perception Convolutional Network) for vegetation extraction from multisource remote sensing images [109]. The results were compared with results from other models, such as the CNN HRNet, the hybrid model OCBDL (Off-Center Bayesian Deep Learning) and RFC (Random Forest Classifier); MSCIN achieved robust vegetation extraction, having the ability to overcome internal fragmentation and unclear boundaries.

The U-Net CNN was found in multiple studies for identification of certain species of trees [110], for multi-temporal forest monitoring [111], for NDVI (Normalized Difference Vegetation Index) estimation [112], for coastal forest cover change detection [113] and for deforestation detection [114], along with other CNNs, such as DeepLabv3+, SegNet, Res-UNet and FC-DenseNet.

The CNN DeepLabv3+ was used by Andrade et al. to detect deforestation from satellite imagery [115]. Other DNNs and CNNs were found in various studies for classifying deforestation patterns [116], for NDVI reconstruction and vegetation cover study [117] and for assessing changes in boreal vegetation [118].

RNNs were found in two studies for vegetation behavior forecasting based on satellite imagery [119] and for generating dense NDVI time series and covering the gaps resulting from cloud cover [120].

DT algorithms were the second most found type of algorithm during the literature review. RF was used for forest cover change detection from satellite imagery [121], to monitor changes in boreal peatland vegetation from satellite imagery [122], to map forest habitat [123] or forest canopy height [124] with satellite imagery as the base, for a forest cover dynamics assessment from satellite imagery [125], to map landscape-scale dynamics of the plant area index from radar data [126], to monitor oak dieback from satellite imagery [127], for vegetation monitoring of mountainous regions from satellite imagery [128], to map coastal mangrove forests from satellite big data [129], for forest/nonforest segmentation based on satellite data [130], for vegetation species detection from remote sensing imagery [131] and to map vegetation cover based on UAV imagery for guided species conservation [132]. Other DT algorithms used in the literature were XGBoost [124,133], GBM [129], SS-RF (Stochastic Spatial Random Forest) [134], CART [131] and IF [116].

SVM algorithm was used for forest cover change detection [121] and for vegetation cover mapping [132], and other similar algorithms, such as SVR (Support Vector Regressor) [124] and OC-SVM (One-Class Support Vector Machine) [116], were also found in the literature.

Statistical models, such as ARIMA [119], QDA (Quadratic Discriminant Analysis) and GNB [130] and LR [132], were also found in the literature.

The classic supervised learning model kNN was used for forest canopy height mapping from satellite imagery [124], detection of different vegetation species from remote sensing images [131] and a basal area and wood volume estimation based on satellite images and LiDAR data [8].

A Fully Connected Neural Network MLP was found in two papers for deforestation patterns classification [116] and for basal area and wood volume estimation [8].

Perbet et al. [101] used a transformer-based model to detect and classify boreal forest disturbances from satellite imagery. Kong et al. used a generative AI model—a dual RSS-GAN network (Remote-Sensing Super-Resolution Generative Adversarial Network)—for long-term vegetation monitoring [49]. Liu et al. used the clustering model K-Means++ for tree species classification [135].

Other papers implemented the use of hybrid models, such as SiamHRnet-OCR [136] and ChangeFormer [137], for deforestation detection from satellite imagery. Zheng et al. used the Near Pseudo model for vegetation species detection from remote sensing imagery [131], Li et al. used the OCBDL model for remote sensing image vegetation extraction [109] and Liu et al. used the YOLOv7 (You Only Look Once) model for tree species classification from UAV data [135].

For forest and vegetation monitoring, after Python programming, which was found in 15 papers; GEE was the most used software product, found in 12 papers; followed by ArcGIS, used in 5 papers (Figure 15). Agisoft and QGIS were found in three papers each, while ENVI, Pix4D and R programming were found in two papers each. Google Earth Pro and SNAP were used in only one paper each.

4.2.5. Pest Disease Monitoring

For pest disease monitoring, CNNs or other DNNs, found in three papers, were the most used models; decision trees-based models were found in two papers and statistical models were found in one paper. In a study by Turkulainen et al. [138], bark beetle-induced spruce damage from UASs (Unoccupied Aircraft Systems) images was determined using DNNs. Zhang et al. [32] used Unet++ with a ResNeSt model as feature extraction backbone for pest-infested forest damage detection. Kislov et al. [139] used a DCNN (Deep Convolutional Neural Network) to detect forest disturbance based on satellite images.

RF was also found in the literature. Junttila et al. [140] used an RF to map spruce tree decline due to bark beetle infestation from multispectral imagery and Iordache et al. [141] used an ML approach based on RF to detect pine wilt disease from airborne spectral imagery.

A statistical model—ARIMA (Autoregressive Integrated Moving Average)—was used by Alkan & Aydin [33] to predict insect damage spread.

For pest disease monitoring, Agisoft was the most used software, found in three papers, followed by ArcGIS, used in two papers. In this case, Python programming and SNAP were found in one paper each, while one paper involved the use of other software products.

4.2.6. Droughts Prediction

Two studies on droughts prediction were found during the search, and neither of them implemented AI algorithms. Villani et al. [34] presented the iCOLT climate service, used for seasonal predictions of irrigation. It is a probabilistic approach, based on soil water balance modeling. Laipelt et al. [35] used the geeSEBAL model, an open-source implementation of the Surface Energy Balance Algorithm for Land (SEBAL) using Google Earth Engine [142], to assess changes in evapotranspiration resulting from cropland expansion.

4.2.7. Soil Content Analysis and Soil Production Capacity Detection

For soil analyses, DT-based models were found to be the most used in the literature (Figure 16). RF was used to estimate the soil moisture content based on UAV images [143] or on satellite data [36], to estimate the soil organic carbon based on satellite data [144,145], to assess the soil macronutrients status based on satellite imagery [37] and to map soil pollution using drone image recognition on an arsenic-contaminated agricultural field [146]. Other decision trees-based models, such as GBDT (gradient descent-boosted regression tree), which was used to estimate soil organic carbon stocks based on satellite imagery [38], Cubist [37] and ERF (Extreme Random Forest) [146], were found in the literature.

SVM was also used to estimate the soil moisture content based on satellite data [36] and to map soil pollution using drone image recognition on an arsenic-contaminated agricultural field [146]. Another similar algorithm, SVR-RBF (Support Vector Regression with Radial Basis Function) was found in a study on organic carbon estimation in soil [144].

The statistical model GLM was used to assess soil macronutrients status based on satellite imagery [37] and to assess atrazine metabolite effects from soil contamination based on UAV images [147]. Another statistical model—LASSO—was found in the literature, implemented in a soil organic carbon estimation study [144].

Other algorithms included FCNNs [38,146], the RNN LSTM [36] and the hybrid model GA-RF (Genetic Algorithm-Optimized Random Forest) [143].

For soil content analysis and soil production capacity, Python programming and ArcGIS were preferred (Figure 17), found in three papers each. Agisoft, SNAP, GEE and R programming were used in two papers each, while QGIS, Matlab, Envi and Pix4D were found in one paper each.

4.3. Answer to the Third Research Question (“What Type of Data Can Be Used in This Regards?”)

The data sources were classified into the categories found in Figure 18. Satellite imagery was by far the most used data source, found in 78 papers. Next, UAV imagery was found in 18 papers and Radar data in 14 papers. Five studies used unspecified remote sensing imagery, while LiDAR data was found in four papers. In a single study, a stereo camera system placed at the study location was used.

The sources of data used in the selected papers were also investigated (Figure 19). Among the many available open-access satellite datasets, Sentinel data was the most commonly used, found in 43 studies. The Sentinel-1 mission, part of the Copernicus Initiative by the European Commission (EC) and the European Space Agency (ESA), comprises a constellation of two sun-synchronous polar-orbiting satellites, flying in the same orbital plane but phased 180° apart, equipped with C-band SAR systems, enabling all-weather, day-and-night Earth observations. Sentinel-1 operates in four distinct imaging modes, offering spatial resolutions down to 5 m and swath widths up to 400 km [148]. The Sentinel-2 mission is a European initiative focused on wide-swath, high-resolution, multispectral Earth observations. It consists of a pair of satellites achieving global coverage every 5 days at the Equator. Each satellite is equipped with a single payload: the Multispectral Instrument (MSI) is capable of capturing imagery across 13 spectral bands. These include the following: four bands at a 10 m resolution, six bands at a 20- m resolution and three bands at a 60 m resolution, with a swath width of 290 km [149].

Landsat data was found in 31 papers. Landsat-8, launched by NASA in 2013, is equipped with the Operational Land Imager (OLI), which captures Earth imagery across the visible, near-infrared (NIR) and shortwave infrared (SWIR) spectrum. The satellite system features two additional spectral bands: Band 1 is deep blue and optimized for coastal zone observation and Band 9 is optimized for cirrus cloud detection in the NIR. Landsat-8 can ensure a 15 m resolution for panchromatic images and 30 m for multispectral ones [150]. Landsat 4, built and launched by NASA–NOAA, is equipped with the Landsat 4 Thematic Mapper (TM) sensor which features seven spectral bands, capturing data across the blue, green, red, NIR, mid-infrared (two bands) and thermal infrared (TIR) regions of the electromagnetic spectrum [151].

Gaofen (GF) is a series of Chinese high-resolution Earth imaging satellites, launched as part of the China High-Resolution Earth Observation System (CHEOS), and GF data was found in six papers, including images from the Gaofen-1 (GF-1) mission, Gaofen-2 (GF-2) mission and Gaofen-6 (GF-6) mission. GF-1 is equipped with two main imaging systems: a panchromatic and multispectral camera (PMC) and a wide-field imager (WFI). The PMC is a high-resolution pushbroom imager featuring Time Delay Integration (TDI), consisting of two cameras that provide a swath width of 69 km at nadir. It captures imagery in one NIR and three visible bands—blue, green and red—with a spatial resolution of 8 m. Additionally, it provides 2 m resolution panchromatic images in the visible range. It makes observations in three visible bands and one NIR band, offering 16 m spatial resolution [152]. A GF-2 satellite is equipped with two PAN/MS (panchromatic/multispectral) cameras. It is capable of capturing high-resolution imagery with a Ground Sampling Distance (GSD) of 0.81 m in the panchromatic band and 3.24 m in the multispectral bands across a 45 km-wide swath [153]. A GF-6 satellite is equipped with a panchromatic/hyperspectral camera offering resolutions of 2 m and 8 m, respectively, with an image swath exceeding 90 km. Additionally, it carries a wide-angle camera with a resolution of 16 m and an expansive swath of 800 km. Both cameras employ three-mirror anastigmatic telescopes and are capable of imaging from the visible to the NIR spectrum, covering wavelengths of approximately 450–900 nm [154].

Planet Scope data was found in four papers. The Planet Scope satellite camera operates in eight spectral bands—Coastal Blue, Blue, Green I, Green, Yellow, Red, RedEdge and NIR—with a GSD (at nadir) between 3 m and 4.1 m [155].

WorldView data was found in three papers, including WorldView-2, WorldView-3 and WorldView-4. WorldView-2 is capable of capturing high-resolution imagery with a GSD of 0.46 m in the panchromatic and 1.8 m in the eight multispectral bands [156]. WorldView-3 captures imagery at 0.31 m resolution in the panchromatic, 1.24 m in the eight VNIR (visible-to-near-infrared) bands, 3.7 m in the eight SWIR bands and 30 m in the CAVIS (Clouds, Aerosols, Vapors, Ice and Snow) bands [157]. WorldView-4 was designed to expand European Space Imaging’s offerings with ultra-high-resolution Earth observation capabilities, and it captures imagery at 31 cm resolution in panchromatic mode and 1.23 m resolution in multispectral mode [158].

SPOT (Satellite pour l’Observation de la Terre) data was used in two papers found in this SLR. Initiated by the French National Centre for Space Studies (CNES), five satellites were deployed between 1986 and 2015. These missions offered detailed views of the Earth’s surface, enabling significant advances in areas such as cartography, vegetation monitoring, land use and land cover analysis and disaster impact assessments. All satellites in the series feature both panchromatic and multispectral imaging capabilities, covering a swath width of 60 km. SPOT-6 and SPOT-7 cover a GSD of 1.5 m at nadir in panchromatic mode and 6 m at nadir in multispectral mode [159].

NASA’s UAVSTAR (Uninhabited Arial Vehicle Synthetic Aperture Radar) was found as the data source in two papers. It uses airborne radar and operates by transmitting and receiving radio waves to gather information about the Earth’s surface features. UAVSAR has collected data since 2007 in the following fields of study: vegetation, ice and glaciers, soil moisture, oceanography, earthquakes, volcanoes and other [160].

The Global Ecosystem Dynamics Investigation (GEDI) is a LiDAR instrument mounted on the International Space Station, designed to deliver high-resolution observations of Earth’s three-dimensional forest structure and was used in two papers in this study. It captures data at a spatial resolution of approximately 30 m per footprint, with samples spaced every 60 m along its orbital path. Additionally, it offers a spectral resolution of 1 m [161].

It is worth noting that MAXAR satellite imagery was found in one paper. Although it offers a spatial resolution ranging from 30 cm to 1.2 m and supports spectral analysis across the visible, NIR and RedEdge bands [162], the high cost of the images can be a limiting factor for broader use.

In 20 papers, UAVs and other personal data sources were used. The spatial resolution of UAV imagery is typically very high, often better than satellite imagery, depending on the flight altitude, sensor quality and camera settings, with some of the most advanced cameras ensuring a GSD up to 3 cm [163].

Other sources, each found in only one paper, include ESA’s RapidEye data, SRTM DTM mission, PolSAR data, UC Merced Land Use Dataset, ICESat-2, ISPRS Potsdam Benchmark dataset, DeepGlobe dataset, ZY-3 dataset, Fuyang remote sensing image dataset, NASA SMAP mission data, NWPU-RESISC dataset, Bing Maps aerial imagery, NASA MODIS data, Planet Fusion CubeSat imagery, NICFI satellite data, Pleiades imagery, VHR satellite imagery, LISS IV images, NASA EO-1 Hyperion sensor, NCALM sensor, QuickBird satellite imagery, IKONOS satellite imagery, ESA’s RADARSAT data, Xiongan dataset and SWISSIMAGE HIST 1946 data.

4.3.1. Land Use Mapping and Crop Detection

For land use mapping and crop detection, satellite imagery was by far the most preferred data type (Figure 20). Sentinel and Landsat data were the most used (Figure 21) due to their wide availability. Other satellite data sources were Gaofen, SPOT and WorldView, found in two papers each. Other types of data found in the SLR included radar data, UAV imagery, unspecified remote sensing imagery and LiDAR data. The sources of this data were either NASA UAVSTAR, NASA GEDI, UAVs and other personal sources or other sources less frequently found in this study.

4.3.2. Crop Yield Monitoring

For crop yield monitoring, satellite imagery and UAV imagery were equally preferred (Figure 22), found in three papers each, while radar data was used in only one paper and in one study a stereo camera system was used. In this case, Sentinel data was also the most used (Figure 23) due to its availability and high resolution. Landsat, Planet Scope data and NASA UAVSTAR data were found in one paper each, while three studies used UAVs or other personal sources.

4.3.3. Flood-Prone Area Detection

For flood-prone area detection, radar data was preferred and used in two studies, while satellite and UAV imagery were found in one study each. For the data sources, Sentinel data was used in two studies, while Gaofen and UAV imagery were used in one study each.

4.3.4. Forest and Vegetation Monitoring

Satellite imagery was widely used for forest and vegetation monitoring and was found in 31 papers included in this SLR (Figure 24). UAV imagery was used in five papers, radar data in three papers and LiDAR data in two papers. In two studies, unspecified remote sensing images were used. Regarding the data sources (Figure 25), Sentinel and Landsat data were the most used, being found in 17 and 14 papers, respectively. Other satellite data sources included Gaofen, PlanetScope, WorldView and Maxar, which was used in only one paper. NASA GEDI data was used in one paper, while five other studies included UAV and other personal data sources.

4.3.5. Pest Disease Monitoring

For pest disease monitoring, satellite imagery was found in four papers, with Sentinel being the main data source, while UAV imagery was used in three papers and one study involved the use of other data source.

4.3.6. Droughts Prediction

While performing the initial analysis of the selected papers for this SLR, two papers regarding draughts prediction were found, which were excluded as they did not include the use of AI. In both papers, satellite imagery data sources Sentinel and Landsat were used.

4.3.7. Soil Content Analysis and Soil Production Capacity Detection

For soil content analysis and soil production capacity, satellite and UAV imagery were the most used data types (Figure 26), with Sentinel data found in three papers and Landsat data in one paper (Figure 27). Two papers included the use of radar data.

5. Discussion

During this systematic literature review, 472 papers from Scopus and 309 papers from Web of Science were identified. After removing 201 duplicates, an initial screening of titles and abstracts was conducted on the remaining records, leading to the exclusion of 426 papers. A full-text assessment was then performed on the remaining articles, resulting in the exclusion of an additional 46 papers. Ultimately, 108 studies met the inclusion criteria and were retained for the final review.

The first significant finding of the systematic literature review is the growing interest in the application of AI in remote sensing for agricultural research, with the number of relevant publications increasing from 8 in 2020 to 51 in 2024. In terms of geographical distribution, China stands out as the country with the highest number of studies. In contrast, regions such as Eastern Europe remain underrepresented in research on the use of AI and remote sensing in agriculture. This gap is particularly noteworthy given that countries like Romania, Ukraine and Bulgaria are recognized for their extensive arable lands and significant agricultural sectors.

Based on the search queries, to answer the first research question—“What are the main problems that can be solved through incorporating AI-driven classification techniques in the field of smart agriculture and environmental monitoring?” (Figure 28)—the papers were divided into the following categories: land use mapping and crop detection, crop yield monitoring, flood-prone area detection, forest and vegetation monitoring, pest disease monitoring, droughts prediction and soil content analysis and soil production capacity detection. The categories with the highest research interest were land use mapping and crop detection (44 papers) and forest and vegetation monitoring (41 papers). In contrast, fewer studies addressed soil content analysis and soil production capacity detection (eight papers), pest disease monitoring (six papers), crop yield monitoring (six papers) and flood-prone area detection (three papers). Notably, no studies meeting the inclusion criteria were found in the category of droughts prediction. This absence highlights a significant research gap and underscores the need for further studies in this area, especially considering the increasing impact of climate change on agricultural productivity. Furthermore, other categories, such as flood-prone area detection, pest disease monitoring, crop yield monitoring and soil content analysis and soil production capacity detection were also underrepresented. These findings highlight the need for more research in these areas to support sustainable and data-driven agricultural practices.

To answer the second research question—“What are the main methods and strategies used in this technology?” (Figure 29)—the AI algorithms found in this SLR were divided into the following categories: decision trees-based models, Convolutional Neural Networks and other Deep Neural Networks, Support Vector Machine, hybrid models, statistical models, other classic supervised learning models, Recurrent Neural Networks, transformers and attention-based models, Fully Connected Neural Networks, clustering algorithms, optimization algorithms and generative models. Among these, decision trees-based models (found in 55 papers) and Convolutional Neural Networks and other Deep Neural Networks (found in 49 papers) were the most commonly employed, highlighting their large-scale applicability. The widespread use of CNNs, in particular, can be attributed to their ability to effectively process grid-like data structures, such as imagery. Hybrid models are especially interesting (found in 13 papers), as they combine different AI techniques to overcome the limitations of using a single AI method. Regarding the implementation tools used for algorithm development and data processing, Python was the most commonly used programming language, mentioned in 45 papers. Its popularity is likely due to its open-source nature, extensive libraries for machine learning and remote sensing and overall versatility in handling large-scale data. Google Earth Engine was found in 23 papers in total, reflecting its relevance as a cloud-based, free for non-commercial use platform, for processing and analyzing large volumes of remote sensing data. Interestingly, ArcGIS was used in 21 papers, despite its high costs, highlighting its robust geospatial analysis capabilities and user-friendly interface, which, in addition to the available geoprocessing tools that simplify the overall classification workflows, allows for Python scripts integration for advanced algorithm development. The main methods used in each category were further studied. Decision trees-based models were the most preferred for land use mapping and crop detection, crop yield monitoring and soil content analysis and soil production capacity detection, while for the forest and vegetation monitoring and pest disease monitoring categories Convolutional Neural Networks and other Deep Neural Networks were the most used. In the three papers on flood-prone area detection, one algorithm from each of the following four categories was found: CNNs and other DNNs, transformer and attention-based models, decision-trees-based models and SVM. Python programming was the most used environment for land use mapping and crop detection, crop yield monitoring, forest and vegetation monitoring and soil content analysis and soil production capacity detection; for flood-prone area detection, ArcGIS software was the most used; and for pest disease monitoring, Agisoft was the most frequently found.

To answer the third research question—“What type of data can be used in this regards?” (Figure 30)—the data sources were classified into the following categories: satellite imagery, being by far the most frequently found type of data (78 papers); UAV imagery (found in 18 papers); radar data (found in 14 papers); unspecified remote sensing imagery (used in 5 papers); LiDAR data (found in 4 papers) and a stereo camera system (found in 1 paper). The widespread use of satellite imagery highlights its accessibility, broad spatial coverage and the availability of free and high-resolution datasets, making it a preferred choice for large-scale agricultural monitoring, while UAV data is preferred for local, more accurate observations. Overall, the open-source Sentinel data was the most used, found in 43 papers, due to its wide availability and high-spatial and -temporal resolutions, followed by the Landsat data, used in 31 papers, with a slightly lower spatial resolution. This type of data can support multi-temporal analyses, benefitting from a long historical archive. Other high-resolution data sources, such as QuickBird, were used in some studies, but their applicability is often limited by their high acquisition costs and restricted temporal coverage. In contrast, Sentinel and Landsat provide free, openly accessible data with global coverage, making them more practical for large-scale and multi-temporal studies, which explains their prevalence in the literature. Satellite imagery was the preferred data source for land use mapping and crop detection, forest and vegetation monitoring and pest disease monitoring. For crop yield monitoring and soil content analysis and soil production capacity detection, satellite and UAV imagery were equally preferred, while for flood-prone area detection, radar data was the most frequently used, due to its capacity to penetrate clouds and operate in all kinds of weather conditions. In all categories, Sentinel data was the most used data source, due to its availability, lack of costs, high resolution and versatility.

6. Conclusions

This systematic literature review highlights the growing interest in research on the integration of AI and remote sensing techniques into agriculture from 2020 to 2024. The increase in publications reflects a rising global interest in developing tools to address agricultural and environmental challenges. This review revealed that the most studied topics were land use mapping and crop detection and forest and vegetation monitoring, where AI and remote sensing proved to be highly effective. However, notable research gaps were identified, particularly in droughts prediction, gaps that are especially relevant given the context of climate change and food security.

CNNs and DT-based models emerged as the most frequently used AI techniques in the analyzed studies. Their popularity is likely due to their ability to handle the complex and large datasets commonly used in remote sensing applications. Satellite imagery—particularly from open-access platforms such as Sentinel and Landsat—played an important role in these analyses, offering high-spatial, -spectral and -temporal resolutions. The accessibility and continuity of satellite data make it especially relevant for monitoring agricultural dynamics, providing up-to-date insights into land use changes and crop health.

In

Table 5. Most used algorithms, software products, data type and data source for each category.

Category	Most Used Algorithm Type	Most Used Software Product	Most Used Data Type	Most Used Data Source
Land use mapping and crop detection	DT-based models	Python programming	Satellite imagery	Sentinel data
Crop yield monitoring	DT-based models	Python programming	Satellite imagery	Sentinel data
Flood-prone area detection	CNNs and other DNNs	ArcGIS	Radar data	Sentinel data
Forest and vegetation monitoring	CNNs and other DNNs	Python programming	Satellite imagery	Sentinel data
Pest disease monitoring	CNNs and other DNNs	Agisoft	Satellite imagery	Sentinel data
Droughts prediction	-	-	-	-
Soil content analysis and soil production capacity detection	DT-based models	Python programming	Satellite imagery	Sentinel data

, the most used algorithms, software products, data type and data source for each category are highlighted, indicating the current trends in land use mapping and crop detection, crop yield monitoring, flood-prone area detection, forest and vegetation monitoring, pest disease monitoring, droughts prediction and soil content analysis and soil production capacity detection.

By facilitating smarter, more efficient and data-driven agricultural practices, the integration of AI and remote sensing contributes directly to the achievement of several SDGs, including SDG 2 (Zero Hunger), SDG 9 (Industry, Innovation and Infrastructure), SDG 12 (Responsible Consumption and Production) and SDG 13 (Climate Action). The findings are relevant for SDG 6 (Clean Water and Sanitation), as remote sensing and AI techniques applied to flood-prone area detection and droughts prediction can improve water resource management and reduce the risks associated with water scarcity or excess. They also contribute to SDG 9 (Industry, Innovation and Infrastructure), as digital technologies, AI-driven analyses and open-source satellite platforms such as Sentinel and Landsat can foster innovation in agricultural monitoring and support the development of resilient digital infrastructures for sustainable farming. Furthermore, the subject aligns with SDG 12 (Responsible Consumption and Production), since AI-based tools for soil content analysis, crop yield monitoring and pest and disease monitoring provide insights that encourage efficient resource use, minimize waste and support sustainable production systems. Finally, the study is closely linked to SDG 13 (Climate Action), as AI and remote sensing are essential for monitoring climate-related issues, such as droughts and floods.

Possible limitations of this SLR include the exclusive use of Scopus and Web of Science databases for the literature retrieval, which may have resulted in the omission of relevant studies indexed elsewhere. Additionally, only papers written in English were considered, potentially excluding relevant research published in other languages. This review focused solely on publications from 2020 onwards, and therefore may not have captured earlier foundational work in this field. During the selection process, some topics—such as mapping wildfires with AI—were excluded for being outside the scope of the SLR. Other subjects, such as the influence of wildlife on agricultural activities, were not considered in our search strategy. While agricultural activities can indeed be influenced by wildlife from surrounding forests, and AI algorithms are increasingly being applied for wildlife detection [164,165], this line of research does not necessarily rely on remote sensing data.

Finally, we observed an uneven geographic representation of the studies, with certain regions—particularly Eastern Europe—remaining underrepresented. While this could be partly due to our study’s limitation to using only Scopus and WoS, it could also be the result of differences in policy support, research funding and national priorities, and variations in data availability and openness, as open-access satellite imagery and AI-ready datasets are more widely used in regions with established digital infrastructure and strong governmental support for smart agriculture initiatives.

To support the wider adoption of AI and remote sensing for sustainable agriculture, particularly in developing countries, several policy measures can be considered. Governments and institutions can promote the use of free, open-access satellite data such as Sentinel and Landsat, which provide high-quality imagery without significant financial investment. Policies that encourage data sharing, collaboration between research institutions and public–private partnerships can further enhance access to necessary infrastructure and expertise. By adopting such measures, developing countries can implement low-cost remote sensing and AI technologies to improve agricultural and vegetation monitoring, optimize resource use and increase resilience to climate-related risks.

We propose that future activities focus on filling the research gaps and continuing to promote the use of open-source data and tools, particularly in underrepresented regions, to ensure more efficient and sustainable agriculture.