Machine Learning in Land Use Prediction: A Comprehensive Review of Performance, Challenges, and Planning Applications

Abstract

The accelerated global urbanization process has positioned land use/land cover change modeling as a critical component of contemporary geographic science and urban planning research. Traditional approaches face substantial challenges when addressing urban system complexity, multiscale spatial interactions, and high-dimensional data associations, creating urgent demand for sophisticated analytical frameworks. This review comprehensively evaluates machine learning applications in land use prediction through systematic analysis of 74 publications spanning 2020–2024, establishing a taxonomic framework distinguishing traditional machine learning, deep learning, and hybrid methodologies. The review contributes a comprehensive methodological assessment identifying algorithmic evolution patterns and performance benchmarks across diverse geographic contexts. Traditional methods demonstrate sustained reliability, while deep learning architectures excel in complex pattern recognition. Most significantly, hybrid methodologies have emerged as the dominant paradigm through algorithmic complementarity, consistently outperforming single-algorithm implementations. However, contemporary applications face critical constraints including computational complexity, scalability limitations, and interpretability issues impeding practical adoption. This review advances the field by synthesizing fragmented knowledge into a coherent framework and identifying research trajectories toward integrated intelligent systems with explainable artificial intelligence.

1. Introduction

The accelerated global urbanization process has positioned land use/land cover change (LULC) modeling [1] as a critical component of contemporary geographic science and urban planning research [2]. Traditional LULC modeling approaches encounter substantial challenges when addressing the inherent complexity of urban systems [3,4], multiscale spatial interactions [5,6], and high-dimensional data associations, creating urgent demand for more sophisticated analytical frameworks [7,8]. Contemporary research indicates that machine learning technologies, through intelligent algorithm design and automated pattern recognition, provide unprecedented capabilities for solving complex urban applications across multiple scales, from land use coverage analysis to building environment configuration and architectural design optimization [9,10].

Land use prediction methodologies have evolved through distinct phases reflecting technological advancement and computational capabilities. Early applications relied on traditional statistical approaches and cellular automata models that established foundational frameworks for spatial simulation. The emergence of machine learning introduced sophisticated algorithmic capabilities, with traditional methods including random forest algorithms achieving overall accuracy exceeding 90% and support vector machine approaches demonstrating superior performance in spatial prediction tasks [11,12]. Advanced neural network implementations marked a significant methodological transition, with multilayer perceptron models achieving exceptional performance, significantly outperforming conventional approaches [13]. The deep learning revolution introduced unprecedented capabilities through convolutional-recurrent neural network integration, achieving remarkable performance metrics (overall accuracy of 99.18% and Kappa coefficient of 0.88) [14], while long short-term memory networks demonstrated superior capability in capturing temporal dependencies for urban expansion modeling [15]. Contemporary applications have progressed toward hybrid methodologies that systematically integrate multiple algorithmic components, with ensemble approaches utilizing random forest and Markov chain models demonstrating reliable performance [16,17] and advanced integration strategies, achieving substantial performance gains through comprehensive utilization of spatial and temporal information [18,19].

Machine learning applications in land use prediction have evolved significantly from conventional statistical approaches toward integrated intelligent systems. Random forest algorithms consistently achieve overall accuracy exceeding 90% with a Kappa coefficient around 0.88 in metropolitan applications [11], while support vector machine approaches demonstrate superior performance in spatial prediction tasks [12]. Advanced neural network implementations, including multilayer perceptron models, achieve exceptional performance with a Kappa coefficient of 0.9025, significantly outperforming traditional approaches in comparative assessments [13]. Deep learning technologies have introduced methodological innovations, with convolutional-recurrent neural network integration achieving exceptional performance metrics, including an overall accuracy of 99.18% and a Kappa coefficient of 0.88 [14]. Long short-term memory networks demonstrate superior capability in capturing temporal dependencies for urban expansion modeling [15], while cellular automata-artificial neural network integration achieves substantial accuracy improvements over traditional methods [20,21]. Ensemble approaches utilizing random forest and Markov chain models demonstrate reliable performance in understanding metropolitan urban growth trajectories [16,17].

Contemporary applications demonstrate increasing sophistication in addressing multidimensional urban complexity through algorithmic integration and optimization techniques [22,23]. Hybrid methodologies consistently outperform single-algorithm implementations, with advanced integration strategies achieving substantial performance gains through comprehensive utilization of spatial and temporal information [18,19]. Vector-based cellular automata enhanced with convolutional neural networks achieve superior performance in urban land use change simulation [24], while temporal-enhanced approaches achieve precision improvements through fine spatial granularity utilization [25]. Deep learning architectures, including U-Net, enable accurate urban land use prediction with minimal parameterization requirements [26], and patch-generating land use simulation models demonstrate effectiveness in multi-scenario urban growth assessment [27,28]. Optimization algorithms integrated with cellular automata provide enhanced parameter optimization capabilities [29], while artificial intelligence applications demonstrate promising potential for pioneering urban spatial planning through synergistic integration approaches [30,31].

Three-dimensional modeling extensions address both horizontal expansion and vertical development complexity, representing significant methodological frontiers [32]. Enhanced simulation models achieve comprehensive 3D urban dynamics modeling under shared socioeconomic pathways [33] while building height prediction approaches demonstrate robust applicability for complex urban planning scenarios [34,35]. Large-scale applications demonstrate the capability to process comprehensive parcel-level datasets while maintaining robust performance across different spatial and temporal scales [36,37]. Global applications processing over 10,000 urban settlements demonstrate the feasibility of consistent boundary delineation using automated remote sensing approaches [38]. Cross-regional transferability studies indicate that advanced models can maintain performance across different geographic contexts [39,40]. Specialized modeling frameworks address diverse urban challenges, including carbon storage dynamics assessment [41], ecosystem service value prediction [42,43], and integrated disaster risk evaluation [44,45]. Advanced applications encompass flood risk assessment, achieving high accuracy in urban contexts [46,47], and a comprehensive landscape ecological risk assessment using driving mechanisms [48,49].

Machine learning applications in land use prediction have demonstrated substantial maturity and technical sophistication across diverse methodological implementations and geographic contexts. The field has evolved from experimental studies to operational applications, achieving consistently high-performance standards, with sophisticated approaches enabling comprehensive analysis of complex urban systems while maintaining robust performance [50,51]. Advanced modeling frameworks incorporate multi-objective optimization and uncertainty analysis, with integrated approaches combining multiple analytical capabilities, achieving superior performance in environmental assessment applications [52]. Specialized applications have expanded to address critical sustainability challenges, including carbon emission prediction for administrative regions [53], and a comprehensive urban thermal field assessment [54,55]. Global-scale validation across 11,581 cities in 61 countries confirms consistent urban scaling relationships, with remote sensing providing uniform standards for cross-regional comparisons [56]. International applications demonstrate broad geographic applicability, with successful implementations across diverse contexts, including European land cover change simulation [40], and developing economy applications, achieving substantial accuracy improvements [57,58]. These technological advances indicate that machine learning applications have transitioned from research innovations to essential tools for evidence-based urban planning, supporting sustainable development through comprehensive scenario analysis and adaptive management frameworks [59,60].

Despite substantial technical progress, contemporary applications face multifaceted constraints, including algorithm complexity trade-offs, scalability limitations, cross-regional transferability challenges, and model interpretability issues that limit widespread adoption in planning practice. Implementation challenges include varying complexity-accuracy trade-offs where advanced methods require larger computational resources compared to traditional approaches [61]. Model interpretability requirements for planning applications necessitate a balance between technical advancement and practical applicability [59,62]. This review addresses these challenges through comprehensive analysis of 74 research publications spanning 2020–2024, with three main objectives: (1) to characterize current applications, technical characteristics, and development prospects of machine learning technologies in land use prediction; (2) to elucidate implementation barriers including technical constraints, data quality issues, and model interpretability problems; (3) to identify future development trajectories toward integrated intelligent systems while providing recommendations for technology transfer, capacity building, and institutional adaptation in planning practice.

2. Research Methodology

2.1. Literature Search and Selection

This systematic review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA, Detailed information can be found in the Appendix A) 2020 guidelines. The review protocol was not registered. The analysis examined machine learning applications in land use prediction through a comprehensive examination of 74 research publications spanning 2020–2024. The final results were independently obtained by the three authors (Figure 1). Literature search was conducted using the Web of Science Core Collection database with the following search strategy:

TS = ((“machine learning” OR “deep learning” OR “artificial intelligence” OR “random forest” OR “neural network” OR “support vector machine” OR “gradient boosting” OR “XGBoost” OR “decision tree”) AND (“land use change prediction” OR “land use simulation” OR “land use modeling” OR “land use forecasting” OR “urban growth prediction” OR “urban growth modeling” OR “spatial prediction” OR “LUCC prediction” OR “cellular automata” OR “CA-Markov” OR “PLUS model” OR “CLUE-S” OR “FLUS” OR “urban expansion modeling”) AND (“spatial planning” OR “urban planning” OR “city planning” OR “urban development”)) AND PY = (2020–2024) AND LA = (English) AND DT = (Article OR Review).

Study selection followed a two-stage screening process. Three reviewers (T.S., T.L., and J.L.) independently screened titles and abstracts, followed by full-text assessment of potentially eligible articles. Disagreements were resolved through discussion, with final decisions made by the lead author (C.L.) when consensus could not be reached.

Inclusion Criteria: (1) Studies focusing on land use/land cover change prediction, urban growth modeling, or spatial land use simulation; (2) Studies investigating land use change prediction or urban expansion forecasting using quantitative approaches; (3) Studies employing machine learning methods either standalone or integrated with spatial models; (4) Studies addressing spatial planning applications or urban development planning; (5) Peer-reviewed articles and reviews published between 2020 and 2024 in English.

Exclusion Criteria: (1) Studies employing only traditional statistical methods without machine learning approaches; (2) Studies focusing primarily on land use classification rather than prediction; (3) Studies using machine learning solely for data preprocessing; (4) Studies lacking quantitative validation or accuracy assessment; (5) Studies with insufficient methodological details; (6) Studies where land use prediction was not the primary objective.

2.2. Data Extraction

A standardized data extraction form was developed to systematically collect information from included studies. Data extraction was conducted by one reviewer (C.L.) and independently verified by three additional reviewers (T.S., T.L., and J.L.). Extracted variables included study characteristics (geographical location, temporal scope, and data sources), methodological details (algorithm types, model parameters, and validation approaches), and performance outcomes (accuracy metrics, computational efficiency indicators). Discrepancies were resolved through consensus discussion.

Given the methodological diversity of included studies and the absence of standardized assessment tools for machine learning applications in spatial modeling, formal risk of bias evaluation was not conducted. Study quality was assessed based on methodological rigor, validation procedures, and reporting transparency.

2.3. Methodological Classification Framework

To address the complexity and diversity of machine learning approaches in land use prediction, this review establishes a systematic classification framework based on technological maturity and implementation complexity. The framework categorizes methodologies into three distinct groups that reflect both algorithmic sophistication and practical deployment considerations.

Traditional Machine Learning Methods encompass established algorithmic foundations with proven institutional adoption patterns. This category includes classical statistical learning approaches (Random Forest and Support Vector Machine) and conventional neural network architectures implemented through standard training procedures (Multi-Layer Perceptron and Artificial Neural Networks). The defining characteristic involves reliance on well-established mathematical foundations with predictable implementation requirements and institutional familiarity, operating as standalone algorithms or with minimal computational enhancement. Deep Learning Architectures represent advanced neural network innovations characterized by sophisticated automatic feature learning capabilities and complex spatial-temporal pattern recognition. These methods leverage recent advances in neural network design, including specialized convolutional architecture (U-Net and ConvLSTM), and demonstrate state-of-the-art performance through architectural sophistication rather than conventional training paradigms. Hybrid and Ensemble Methodologies constitute systematic algorithmic integration that deliberately combines multiple distinct computational paradigms through explicit coordination mechanisms. These approaches include neural network-cellular automata coupling, multi-algorithm fusion frameworks, and ensemble methods that demonstrate emergent capabilities exceeding individual component performance through strategic multi-paradigm integration and comprehensive optimization frameworks.

Classification decisions prioritize the core algorithmic innovation over implementation details. Methods involving systematic integration of multiple distinct algorithmic paradigms are categorized as hybrid methodologies regardless of individual component complexity, while computational enhancements of established algorithms remain within their base category when the fundamental algorithmic approach is unchanged.

3. Results

3.1. Current Status and Characteristics of ML Applications in Land Use Prediction

A total of 74 studies met the inclusion criteria and were included in this review (Figure 1). The included studies demonstrated diverse geographical coverage and methodological approaches in machine learning applications for land use prediction.

3.1.1. Application Distribution and Method Selection Patterns

Figure 2 presents the global distribution of machine learning applications in land use prediction based on our analysis of 74 studies, revealing significant regional concentrations and methodological preferences across different continents. Contemporary machine learning applications demonstrate clear evolutionary trajectories across three distinct methodological categories, each exhibiting unique performance characteristics and application preferences. Traditional machine learning methods maintain substantial adoption through proven reliability and interpretability, with random forest algorithms consistently achieving overall accuracy exceeding 90% and a Kappa coefficient around 0.88 in metropolitan-scale applications [11]. The Random Forest model demonstrates superior performance with AUC-ROC values of 99.5%, Kappa of 98%, and accuracy of 90% in landslide susceptibility mapping applications, significantly outperforming other approaches, including decision tree (93% AUC), XGBoost (98% AUC), and k-nearest neighbor (91% AUC) methods [46]. Support vector machine applications achieve mixture discriminant analysis performance with 83.2% AUC, followed by random forest (80.6%), boosted regression tree (78.0%), and multivariate adaptive regression spline (75.5%) in groundwater spring potential mapping [12].

Deep learning architecture represents the most sophisticated category, demonstrating exceptional capabilities in complex pattern recognition and multiscale analysis. Convolutional-recurrent neural network integration achieves remarkable performance metrics, including 99.18% overall accuracy, a Kappa coefficient of 0.88, and a figure of merit of 0.13, with improvements attributed to multiscale neighborhood information powered by U-Net and time series information of historical urban expansion uncovered by LSTM [14]. The Beijing–Tianjin–Hebei application demonstrates urban land projected to peak at 8736–9155 km² during 2039–2043, representing increases in the range of 10.99–16.31% compared to 2020 levels. Three-dimensional modeling innovations achieve Figure of Merit values ranging from 0.21 to 0.35 for horizontal expansion, overall accuracy values of 83% for refinement of urban functional types, and root mean squared error values of 5–7 m for built-up height simulations across Beijing–Tianjin–Hebei, Yangtze River Delta, and Pearl River Delta regions [33].

Hybrid and ensemble methodologies have emerged as the dominant paradigm through algorithmic complementarity that systematically addresses limitations of individual approaches while amplifying respective strengths. The PST-CA model integrates self-organizing map partitioning with spatiotemporal 3D convolutional neural networks, achieving overall accuracy improvements of 4.66–6.41% compared to four traditional models, including logistic regression-CA, support vector machine-CA, random forest-CA, and artificial neural network-CA [19]. Advanced integration strategies demonstrate substantial performance gains, with K-means-CNN-LSTM-CA approaches improving FoM index by 9.86–19.43% compared to traditional logistic regression-CA and artificial neural network-CA models, while temporal dependency consideration increases FoM index by 0.98–3.51% and spatial heterogeneity consideration contributes 1.08–5.15% improvements [18]. The Temporal-VCA framework achieves precision improvements up to 22.12% through comprehensive utilization of fine spatial and temporal granularity information of cadastral plot temporal data, outperforming regular VCA models and traditional raster CA models [25].

3.1.2. Technical Integration and Innovation Strategies

Technical integration strategies demonstrate increasing sophistication in addressing multidimensional urban complexity through intelligent algorithm fusion and adaptive optimization frameworks. Spatial partitioning methodologies employ advanced clustering techniques where a self-organizing map, coupled with hierarchical clustering and patch generation land use simulation models, automatically determine optimal partition schemes, achieving FoM improvements compared to traditional approaches [63]. The approach addresses spatial heterogeneity by dividing study areas into homogeneous sub-regions using machine learning-based partitioning strategies, with spatiotemporal 3D convolutional neural networks extracting spatiotemporal neighborhood features for enhanced land use change simulation.

Optimization techniques have evolved beyond traditional parameter tuning toward intelligent algorithm fusion with real-time adaptation capabilities. The artificial fish swarm algorithm integrated with cellular automata demonstrates superior performance for optimizing parameters of variables in urban growth models, outperforming genetic algorithm-cellular automata and other competing models [29]. Dynamic spatiotemporal rolling prediction models achieve an average adjusted R² of 0.89 through integrated historical trends, neighborhood status, and spatial proximity analysis, representing paradigm shifts from static macroscopic to dynamic microscopic prediction approaches that offer valuable insights for future urban development and planning decisions [50].

Three-dimensional modeling innovations represent a significant methodological advancement through vertical urban development integration with traditional horizontal expansion analysis. The enhanced Future Land Use Simulation model in a 3D version simulates continuous 3D dynamics of real-world urban development, with the distinctive characteristic of concurrently updating 3D information of developed land grids during simulation processes [33]. Building height prediction utilizes random forest regression algorithms to predict building height growth in the vertical direction, with crucial factors affecting building heights, and simulation results of future urban 3D expansion hotspot areas providing scientific support for urban spatial planning decisions [34]. The R² was 0.8628, RMSE was 2.0180, MAE was 1.3707, and EV was 0.8632 for building height prediction accuracy in Shenzhen applications.

3.2. ML Performance in Land Use Change Driving Mechanism Identification

3.2.1. Driving Factor Identification Capability Assessment

Machine learning algorithms demonstrate exceptional capabilities in identifying and quantifying complex driving mechanisms underlying land use change processes through advanced feature importance analysis and comprehensive relationship discovery techniques. Margin-based random forest approaches prove more reliable and sensitive than traditional importance measures when detecting driving mechanisms behind land use change, with importance values and ranking orders remaining stable regardless of similarity measure selection [64]. The methodology successfully identifies core driving factors of urban land use change using multitemporal global land cover products and point-of-interest data, revealing that topographic conditions persistently affect urban land use change while transportation factors gradually become the most important human driving factors.

Advanced ensemble learning frameworks enhance driving factor analysis through integrated importance ranking and cross-validation approaches across multiple algorithmic components. Random forest algorithms achieve the best prediction accuracy among selected machine learning methods for regional-level urban growth prediction, with the methodology demonstrating high prediction accuracy that confirms most variability of urban growth can be described by past observations of self and neighboring changes [65]. The role of random forest and Markov chain models in understanding metropolitan urban growth trajectory reveals significant shifts in land use and land cover influenced largely by socioeconomic factors, with overall accuracy exceeding 90% and a Kappa coefficient around 0.88 [11].

Deep learning architecture enables the discovery of latent driving relationships previously undetectable through conventional statistical approaches, demonstrating particular strength in processing unstructured spatial data and complex geographic patterns. Convolutional neural networks integrated with vector-based cellular automata extract high-level features of driving factors within neighborhoods of irregularly shaped cells, discovering relationships between multiple land use changes and driving factors at the neighborhood level while obtaining more details on morphological characteristics of land parcels [24]. Deep forest algorithms integrated with vector-based cellular automata achieve best simulation performance at parcel-level (Figure of Merit = 39.88%), pattern-level (similarity = 96.47%), and community-level (correlation coefficient = 0.9269) through capacity to directly mine high-level features from unstructured data [66].

3.2.2. Complex Relationship Modeling and Discovery

Deep learning architecture excels in modeling nonlinear relationships and complex interaction patterns through sophisticated feature extraction and temporal dependency analysis frameworks. Integrated convolutional neural networks and recurrent neural networks capture multiscale neighborhood information powered by U-Net and time series information of historical urban expansion uncovered by LSTM, with improvements attributed to these complementary capabilities, enabling simultaneous simulation of edge expansion and leapfrog expansion [14]. U-Net deep learning algorithms automatically learn complex urban development patterns and processes, such as neighborhood influence, gravity effects of large cities, and the tendency for linear development, without requiring explicit parameterization or forcing data [26].

Sophisticated hybrid modeling approaches enable comprehensive relationship discovery through algorithmic complementarity and multidimensional analysis frameworks. The coupling of cellular automata with area partitioning and spatiotemporal convolution extracts spatiotemporal neighborhood features using a machine learning-based partitioning strategy and spatiotemporal 3D convolutional neural network, achieving overall accuracy improvements of 4.66–6.41% compared to traditional models [19]. K-means-CNN-LSTM-CA models demonstrate that considering temporal dependencies increases FoM index by 0.98–3.51%, while considering spatial heterogeneity increases FoM index by 1.08–5.15%, with comprehensive integration improving FoM index by 9.86–19.43% compared to traditional approaches [18].

Advanced relationship discovery techniques utilize generative approaches and attention mechanisms to uncover complex urban transformation patterns. GAN-based land use and land cover change prediction models using multiscale local spatial information achieve the highest accuracy in both short-time interval and long-time interval scenarios, with results closest to ground truth from a landscape pattern perspective [67]. Temporal vector cellular automata frameworks utilizing coupled temporal data and vector cellular automata achieve precision improvements up to 22.12%, outperforming regular VCA models and traditional raster CA models while revealing complex nonlinear temporal patterns within cadastral-scale urban development processes [25].

3.3. ML Prediction Performance and Method Applicability Assessment

Contemporary machine learning applications demonstrate distinct performance hierarchies across methodological categories, with comprehensive comparative assessments revealing clear accuracy-complexity trade-offs (Table 1). Traditional approaches provide reliable baseline performance, with Random Forest achieving consistent metropolitan applications (OA: 90.89–91.19%, Kappa: 0.87–0.88) and multilayer perceptron demonstrating superior comparative results (Kappa: 0.9025 vs. CA-MARKOV: 0.6941) [7,27]. Deep learning architectures achieve exceptional accuracy through sophisticated pattern recognition, with convolutional-recurrent integration reaching peak performance (OA: 99.18%, Kappa: 0.88, FoM: 0.13) and 3D modeling maintaining reliability across multiple metropolitan regions (OA: 83%, FoM: 0.21–0.35) [68,69]. Hybrid methodologies consistently outperform individual algorithms, with integrated approaches improving performance indices by 9.86–19.43% through systematic algorithmic complementarity [26].

Computational efficiency and practical applicability vary significantly across approaches, with critical trade-offs influencing implementation decisions. GPU acceleration enables substantial scalability improvements, achieving 16,000× spatial processing acceleration while maintaining 92% accuracy for large-scale parcel analysis [2]. However, complexity-performance analysis reveals that simpler approaches often provide better practical outcomes, with CA-Markov achieving a 97.20% Kappa coefficient compared to ConvLSTM’s 94.50% while requiring substantially lower computational resources [63]. Parameter-free approaches like U-Net algorithms enable accurate predictions with minimal forcing data requirements, offering practical solutions for resource-constrained applications [42]. Implementation success correlates strongly with methodological selection based on institutional capacity, data availability, and computational constraints rather than peak performance metrics alone, with mixed-cell approaches providing effective solutions for balancing simulation performance and computational intensity in large-scale applications [70].

Table 1. Algorithm Performance Comparison here.

Categories	Algorithm/ Method	Study Area	Performance Metrics	Technical Capabilities	Implementation Requirements
Traditional Machine Learning Methods	Random Forest [11]	Islamabad	OA: 90.89–91.19% Kappa: 0.87–0.88	Large-scale reliability Stable performance over decades	Traditional ML limits Slower processing
	Multi-Layer Perceptron [13]	Kolkata	OA: 92.78% Kappa: 0.9025 vs. CA-MARKOV: 0.6941 vs. STCHOICE: 0.5392	Nonlinear mapping Superior performance Best performer among methods	Data requirements Complex architecture
	GPU-Accelerated ANN [36]	Florida	OA: 92.0% AUC: 0.902 FoM: 0.053 F1: 0.373	Massive scalability 16,000× faster processing GPU acceleration	Hardware needs Technical expertise
	Support Vector Machine [12]	Wadi Az-zarqa	AUC: 80.2% (SVM) MDA: 83.2% (best) RF: 80.5% BRT: 78.0%	High-dimensional data processing Robust classification Highest MDA accuracy	Parameter tuning Computational cost
Deep Learning Architectures	CNN-RNN Integration (U-Net + LSTM) [14]	BTHUA	OA: 99.18% Kappa: 0.8812 FoM: 0.1323	Multiscale feature Temporal information Highest accuracy combination	High computational cost Large training data
	3D-FLUS Model [33]	BTH, YRD, PRD	OA: 83% Kappa: >0.77 FoM: 0.21–0.35 Height RMSE: 5–7 m	3D simulation capability First concurrent modeling Horizontal–vertical integration	Complex calibration High data needs
	ConvLSTM [61]	Casablanca	Kappa: 94.50% vs. CA-Markov: 97.20% vs. MLP-Markov: 89.40%	Advanced deep learning Spatiotemporal feature Regional transferability	Higher complexity Large data volumes
	U-Net Architecture [26]	Shenzhen	OA: 91.0% (median) AUC: 0.81 (median) FoM: 0.20 Hit rate: 0.33	Parameter-free approach Minimal tuning requirements Auto-learned patterns	Limited interpretability Data quality dependent
Hybrid and Ensemble Methodologies	ANN + CA-Markov [20]	Irbid	OA: 90.04% vs. CA-MC: 86.29%	Integration benefit Accuracy improvement Performance enhancement	Implementation complexity Training needs
	CNN + Vector CA [24]	Shenzhen	FoM: 0.361	Vector representation Irregular neighborhoods High-level features	Implementation complexity Limited metrics
	PST-CA (SOM + 3D CNN) [19]	Shanghai	OA: +4.66–6.41% vs. 4 traditional models	Spatial partitioning Temporal extraction SOM partitioning strategy	Complex implementation Parameter sensitive
	K-means-CNN-LSTM-CA [18]	Hangzhou	FoM: +9.86–19.43% Temporal: +0.98–3.51% Spatial: +1.08–5.15%	Multi-algorithm fusion Highest accuracy gains Comprehensive integration	High complexity Coordination challenges
	Temporal-VCA Framework [25]	Shenzhen	Precision: +22.12% vs. regular VCA & raster CA	Cadastral precision Fine temporal patterns Comprehensive utilization	Data requirements Limited availability
	Random Forest + CA [17]	Shanghai	OA: 94.88% Kappa: 0.8772	Scenario modeling High consistency Pattern simulation	RF limitations Pattern aggregation
	LSTM-CA Integration [71]	Lanzhou	OA: 91.01%	Long-term effects Temporal modeling LSTM transition rules	Data requirements Training complexity
	CA(GBDT) Ensemble [72]	YRD	OA: >89% FoM: >27%	Ensemble strength Nonlinear patterns Gradient boosting integration	Model complexity Interpretability issues
	Deep Forest-VCA [66]	Shenzhen	FoM: 39.88% Similarity: 96.47% Correlation: 0.9269	Multi-level performance Deep ensemble Best simulation performance	Computational intensity Parameter tuning

Table 1. Algorithm Performance Comparison here.

Categories	Algorithm/ Method	Study Area	Performance Metrics	Technical Capabilities	Implementation Requirements
Traditional Machine Learning Methods	Random Forest [11]	Islamabad	OA: 90.89–91.19% Kappa: 0.87–0.88	Large-scale reliability Stable performance over decades	Traditional ML limits Slower processing
	Multi-Layer Perceptron [13]	Kolkata	OA: 92.78% Kappa: 0.9025 vs. CA-MARKOV: 0.6941 vs. STCHOICE: 0.5392	Nonlinear mapping Superior performance Best performer among methods	Data requirements Complex architecture
	GPU-Accelerated ANN [36]	Florida	OA: 92.0% AUC: 0.902 FoM: 0.053 F1: 0.373	Massive scalability 16,000× faster processing GPU acceleration	Hardware needs Technical expertise
	Support Vector Machine [12]	Wadi Az-zarqa	AUC: 80.2% (SVM) MDA: 83.2% (best) RF: 80.5% BRT: 78.0%	High-dimensional data processing Robust classification Highest MDA accuracy	Parameter tuning Computational cost
Deep Learning Architectures	CNN-RNN Integration (U-Net + LSTM) [14]	BTHUA	OA: 99.18% Kappa: 0.8812 FoM: 0.1323	Multiscale feature Temporal information Highest accuracy combination	High computational cost Large training data
	3D-FLUS Model [33]	BTH, YRD, PRD	OA: 83% Kappa: >0.77 FoM: 0.21–0.35 Height RMSE: 5–7 m	3D simulation capability First concurrent modeling Horizontal–vertical integration	Complex calibration High data needs
	ConvLSTM [61]	Casablanca	Kappa: 94.50% vs. CA-Markov: 97.20% vs. MLP-Markov: 89.40%	Advanced deep learning Spatiotemporal feature Regional transferability	Higher complexity Large data volumes
	U-Net Architecture [26]	Shenzhen	OA: 91.0% (median) AUC: 0.81 (median) FoM: 0.20 Hit rate: 0.33	Parameter-free approach Minimal tuning requirements Auto-learned patterns	Limited interpretability Data quality dependent
Hybrid and Ensemble Methodologies	ANN + CA-Markov [20]	Irbid	OA: 90.04% vs. CA-MC: 86.29%	Integration benefit Accuracy improvement Performance enhancement	Implementation complexity Training needs
	CNN + Vector CA [24]	Shenzhen	FoM: 0.361	Vector representation Irregular neighborhoods High-level features	Implementation complexity Limited metrics
	PST-CA (SOM + 3D CNN) [19]	Shanghai	OA: +4.66–6.41% vs. 4 traditional models	Spatial partitioning Temporal extraction SOM partitioning strategy	Complex implementation Parameter sensitive
	K-means-CNN-LSTM-CA [18]	Hangzhou	FoM: +9.86–19.43% Temporal: +0.98–3.51% Spatial: +1.08–5.15%	Multi-algorithm fusion Highest accuracy gains Comprehensive integration	High complexity Coordination challenges
	Temporal-VCA Framework [25]	Shenzhen	Precision: +22.12% vs. regular VCA & raster CA	Cadastral precision Fine temporal patterns Comprehensive utilization	Data requirements Limited availability
	Random Forest + CA [17]	Shanghai	OA: 94.88% Kappa: 0.8772	Scenario modeling High consistency Pattern simulation	RF limitations Pattern aggregation
	LSTM-CA Integration [71]	Lanzhou	OA: 91.01%	Long-term effects Temporal modeling LSTM transition rules	Data requirements Training complexity
	CA(GBDT) Ensemble [72]	YRD	OA: >89% FoM: >27%	Ensemble strength Nonlinear patterns Gradient boosting integration	Model complexity Interpretability issues
	Deep Forest-VCA [66]	Shenzhen	FoM: 39.88% Similarity: 96.47% Correlation: 0.9269	Multi-level performance Deep ensemble Best simulation performance	Computational intensity Parameter tuning

3.4. Translation Mechanisms of ML Applications in Planning Decision Support

3.4.1. Decision Support Effectiveness and Planning Applications

Machine learning applications demonstrate substantial effectiveness in translating technical predictions into actionable planning insights through comprehensive scenario-based frameworks and policy-relevant modeling capabilities. Multi-scenario approaches enable systematic evaluation of alternative development strategies, with 3D future land use simulation models successfully applied to simulate future evolution of 3D urban dynamics under Shared Socioeconomic Pathways, effectively illustrating the influence of each SSP on 3D urban development [33]. The predicted urban land use shows planned development scenario projections of 242.10, 312.69, 363.80, and 400.72 km² compared to unplanned development scenario projections of 242.91, 314.31, 366.23, and 403.98 km² during 2021, 2031, 2041, and 2051 [73].

Decision support translation mechanisms demonstrate particular strength in addressing integrated urban challenge assessment through comprehensive technical frameworks. Contemporary applications across major urban regions demonstrate these capabilities through diverse case implementations. The Beijing–Tianjin–Hebei urban agglomeration exemplifies advanced applications through convolutional-recurrent neural network integration, achieving 99.18% overall accuracy and a Kappa coefficient of 0.88. Modeling results project urban land expansion to peak at 8736–9155 km² during 2039–2043, representing increases of 10.99–16.31% compared to 2020 baseline levels [14]. Comparative framework applications for spatially explicit urban growth modeling successfully monitor urban land use efficiency and sustainable urban development, with multilayer perceptron neural network models enabling projection of urban expansion dynamics while providing a scientific basis for technical selection and method optimization [13]. Three-dimensional simulation models provide scientific support for decisions in urban spatial planning by effectively simulating horizontal expansion and vertical growth of real cities in 3D space, with crucial factors affecting building heights, and simulation results of future urban 3D expansion hotspot areas [34].

Temporal–spatial decision support capabilities enable dynamic planning adaptation through sophisticated rolling prediction frameworks. Multidimensional effects modeling using dynamic spatiotemporal rolling prediction models achieves an average adjusted R² of 0.89, representing shifts from static macroscopic to dynamic microscopic prediction of urban land demand, offering valuable insights for future urban development and planning decisions [50]. Multi-scenario-based urban growth modeling provides promising guidelines for urban planners and conservation scientists to implement robust artificial intelligence-based hybrid geo-simulation models for compact, organized, and integrated land use-transportation development [73].

3.4.2. Implementation Challenges and Standardization Requirements

Implementation challenges persist despite technical advances, particularly regarding model interpretability, computational requirements, and institutional capacity limitations affecting widespread adoption. Deep learning approaches, while achieving superior prediction performance with 99.18% overall accuracy and a 0.88 Kappa coefficient, often function as complex systems requiring substantial computational resources and technical expertise [14]. Convolutional LSTM models demonstrate higher complexity due to the number of elementary operations and require larger data volumes as deep learning models, making them more demanding in terms of data volume compared to cellular automata–Markov approaches [61].

Data standardization and integration challenges significantly constrain machine learning adoption in planning contexts, requiring comprehensive datasets encompassing multiple temporal periods and diverse driving factors. Successful margin-based random forest implementations require systematic analysis of core driving factors while maintaining stability across different similarity measures, demanding extensive data collection and preprocessing capabilities [64]. Large-scale applications processing parcel-level land use changes demonstrate both analytical potential and computational demands, requiring substantial data infrastructure and processing capabilities for effective implementation [36].

Institutional integration requirements encompass technical capacity development, workflow adaptation, and stakeholder engagement frameworks necessary for effective machine learning implementation in planning practice. Successful translation requires addressing the balance between prediction accuracy and computational intensity, with mixed-cell cellular automata models offering practical solutions for large-scale applications while achieving balance between simulation performance and computational intensity [74]. Scalability assessment suggests that cellular automata–Markov-based methods show great potential for urban growth forecasting, especially for short-term forecasting when computational resources are limited, while sophisticated approaches require substantial infrastructure investment for effective implementation. The translation from research innovations to practical applications necessitates comprehensive capacity building programs, standardized implementation protocols, and institutional frameworks supporting cross-disciplinary collaboration between technical specialists and planning practitioners.

4. Discussion

4.1. Implementation Challenges and Performance Trade-Offs

Contemporary machine learning applications face fundamental trade-offs between algorithmic sophistication and practical implementation feasibility, with computational complexity presenting significant adoption barriers across diverse institutional contexts. The Beijing–Tianjin–Hebei application demonstrates this paradox, where convolutional-recurrent integration achieved 99.18% overall accuracy and projected urban expansion to 8736–9155 km² during 2039–2043, yet demands substantial computational resources often exceeding institutional capabilities [14]. Comparative analysis reveals critical performance-complexity relationships: Casablanca applications showed ConvLSTM achieving 94.50% Kappa coefficient while requiring larger data volumes compared to CA-Markov approaches reaching 97.20% Kappa with greater computational efficiency [61], illustrating how sophisticated methods may underperform simpler approaches in resource-constrained environments where computational demands exceed institutional capabilities.

Institutional and socio-political constraints compound these technical barriers through decision-making requirements that conflict with algorithmic characteristics. Deep learning architectures achieve exceptional performance with an overall accuracy of 99.18% and a Kappa coefficient of 0.88, but their “black box” nature undermines transparency requirements essential for planning decision-making contexts [14]. Democratic planning processes require explainable AI approaches that balance predictive performance with decision support transparency [59,62], creating institutional resistance where stakeholder engagement frameworks struggle to incorporate algorithmic predictions into participatory planning processes. Implementation complexity varies substantially across organizational capabilities, where Kolkata applications demonstrated multilayer perceptron performance (Kappa: 0.9025) significantly outperforming CA-MARKOV (0.6941) and STCHOICE (0.5392) [13]; successful deployment requires technical expertise often unavailable in planning organizations. Professional capacity limitations significantly constrain effective translation from research innovations to planning practice, where practitioners may lack the technical background necessary to properly implement sophisticated analytical tools that require professional development and institutional adaptation that many agencies are unprepared to undertake [50].

Cross-regional transferability remains limited despite claims of model generalizability, with performance variations reflecting underlying differences in geographic and institutional contexts. The Temporal-VCA framework achieved precision improvements up to 22.12% through comprehensive utilization of fine spatial and temporal information [31], but such detailed datasets are rarely available outside specific institutional contexts. European implementations achieve 88–95% accuracy through standardized data protocols and mature planning frameworks [40,75], while developing economies demonstrate different performance characteristics requiring region-specific adaptation strategies [57]. Data quality and integration challenges significantly constrain machine learning adoption in planning contexts, requiring comprehensive datasets encompassing multiple temporal periods and diverse driving factors that demand extensive data collection and preprocessing capabilities often exceeding available technical resources [64]. Multi-source integration challenges, temporal consistency issues, and validation costs significantly complicate model development while creating barriers to widespread adoption in resource-constrained environments.

4.2. Future Development and Application Pathways

Machine learning applications demonstrate substantial potential for market-level decision support and economic development optimization, extending beyond traditional planning applications to address comprehensive urban challenges. Real-world implementations demonstrate this transformation: Shanghai’s multi-scenario planning framework exemplifies policy-relevant applications, where Random Forest–CA integration achieved 94.88% overall accuracy while enabling systematic comparison between planned development scenarios (projecting 242.10–400.72 km² for 2021–2051) versus unplanned scenarios (242.91–403.98 km²) [17,73]. These quantitative scenario comparisons directly inform municipal land use policy decisions and development control strategies. Beijing–Tianjin–Hebei applications demonstrate large-scale regional planning capabilities, where CNN–RNN integration projected urban expansion to 8736–9155 km² during 2039–2043, representing 10.99–16.31% increases that directly support infrastructure investment planning and resource allocation decisions across the mega-region [14].

Cost–benefit considerations for resource-constrained agencies reveal critical implementation priorities that balance technological sophistication with practical applicability. Cellular automata–Markov-based methods show great potential for urban growth forecasting, especially for short-term applications when computational resources are limited, while sophisticated approaches require substantial infrastructure investment for effective implementation. Mixed-cell cellular automata models offer practical solutions for large-scale applications while achieving a balance between simulation performance and computational intensity [74], providing cost-effective alternatives to GPU-accelerated implementations that enable 16,000× spatial processing acceleration but require substantial infrastructure investment [36]. Successful translation requires addressing the balance between prediction accuracy and computational intensity, where simpler approaches may achieve better practical outcomes than sophisticated but resource-intensive alternatives requiring specialized technical support.

Contemporary research necessitates standardized benchmarking frameworks addressing methodological fragmentation and performance comparability limitations across diverse planning contexts. Essential framework components should include: standardized reference datasets incorporating multi-resolution urban data across representative geographic contexts; unified performance metrics combining traditional measures with planning-specific indicators; computational efficiency benchmarks establishing processing time standards for different institutional capacity levels; cross-regional validation protocols requiring model testing across multiple geographic contexts; reproducibility standards mandating code availability and detailed parameter specifications. Implementation requirements encompass systematic professional development programs addressing technical knowledge gaps constraining widespread adoption, comprehensive data governance frameworks ensuring quality consistency, and institutional capacity building initiatives balancing methodological sophistication with practical applicability.

The translation from research innovations to practical applications necessitates comprehensive capacity-building programs, standardized implementation protocols, and institutional frameworks supporting cross-disciplinary collaboration between technical specialists and planning practitioners. Future applications demand integrated decision support platforms combining multiple analytical capabilities within unified operational frameworks that facilitate non-technical user engagement with sophisticated modeling systems. Critical implementation priorities include coordinated training frameworks integrating theoretical understanding with implementation experience through collaborative institutional partnerships, with development trajectories emphasizing integrated intelligent systems combining multiple analytical approaches, 3D modeling capabilities addressing vertical development complexity, and comprehensive decision support platforms. The research demonstrates that machine learning technologies can enable more responsive and adaptive urban management through real-time monitoring and scenario-based planning capabilities, but successful implementation depends critically on addressing institutional capacity, stakeholder engagement, and transparency requirements while maintaining democratic legitimacy in planning processes.

4.3. Limitations

This review has several limitations. First, we limited our search to English-language publications in Web of Science, which may have introduced language and database bias. Second, the rapid evolution of machine learning technologies means that some recent developments may not be fully captured. Third, the lack of standardized evaluation metrics across studies made direct comparisons challenging. Fourth, we did not conduct a formal quality assessment due to the absence of appropriate tools for this field.

5. Conclusions

This systematic review of 74 publications spanning 2020–2024 reveals that machine learning applications in land use prediction have achieved remarkable technological advancement while encountering persistent implementation barriers limiting widespread adoption in planning practice. Our analysis establishes clear methodological evolution across three categories: traditional machine learning approaches achieving consistent reliability with random forest algorithms reaching over 90% overall accuracy, deep learning architectures delivering exceptional performance with convolutional-recurrent integration achieving 99.18% overall accuracy, and hybrid ensemble methods consistently outperforming single-algorithm approaches through sophisticated algorithmic complementarity.

Previous reviews have focused primarily on algorithmic performance without systematically addressing implementation barriers constraining professional adoption in planning practice. Our research identifies significant disparities between technical capabilities and practical implementation requirements, with critical implications for planning policy development. While sophisticated algorithms demonstrate superior analytical performance, their complexity often exceeds institutional technical capacity, suggesting that planning agencies should prioritize graduated implementation strategies aligning technological sophistication with organizational capabilities. Cross-regional transferability remains limited despite claims of model generalizability, indicating that regulatory frameworks must establish model validation standards and region-specific adaptation protocols while maintaining transparency requirements essential for democratic planning processes. Data quality and availability constraints emerge as fundamental limiting factors, highlighting the need for systematic investment in data infrastructure and professional capacity building rather than advanced algorithmic acquisition.

Future development trajectories should emphasize integrated intelligent systems combining multiple analytical approaches, 3D modeling capabilities, and comprehensive decision support platforms. However, successful implementation requires policy frameworks that balance innovation adoption with accountability mechanisms, ensuring algorithmic predictions support rather than replace professional judgment. Budget allocation strategies should prioritize long-term capacity development over short-term technology acquisition, with cost–benefit analysis revealing that simpler, well-implemented approaches often provide better practical value than sophisticated systems requiring substantial ongoing technical support. Critical research priorities include standardized benchmarking frameworks and explainable artificial intelligence approaches for transparent planning decision-making. The research demonstrates that machine learning technologies can enable more responsive and adaptive urban management through real-time monitoring and scenario-based planning capabilities, but sustainable deployment depends critically on phased implementation strategies, cross-sector partnerships, and continuous evaluation mechanisms that maintain public trust while advancing technical capabilities in planning practice.