Machine Learning Transforms Landscape Architecture: A Comprehensive Review of Applications and Future Visions

The convergence of big data, artificial intelligence (AI), and digital technology has reshaped the landscape of traditional disciplines, and landscape architecture is no exception. As a core component of AI and a powerful tool for big data processing, machine learning has emerged as a transformative force in landscape architecture research and practice, addressing long-standing challenges in quantitative analysis, site assessment, and design generation that once relied on subjective experience and manual labor. A groundbreaking study published in the Journal of Beijing Forestry University systematically reviews the progress of machine learning applications in landscape architecture, dissects its practical value across key workflow stages, and outlines promising future research directions, shedding light on the digital and intelligent evolution of the discipline.

In recent years, the built environment industry—encompassing architecture, urban planning, and landscape architecture—has witnessed a mature ecosystem for the implementation of machine learning algorithms, supported by three foundational pillars: the explosion of multi-source data, advances in algorithmic models, and the popularization of open-source application platforms. The advent of big data, remote sensing, and 5G technologies has generated a wealth of “new data” for landscape architecture, with geographic information data standing out as a key resource that reflects the spatial attributes and dynamic changes of urban and rural sites. Unlike traditional research methods that struggle with large-scale data analysis through sampling and classified weighting, machine learning algorithms excel at mining latent patterns from massive, real-world datasets, enabling comprehensive and in-depth exploration of site evolution laws. Meanwhile, the rapid development of deep learning, a major branch of machine learning, has endowed computers with exceptional image processing and generation capabilities, laying the groundwork for automated design case generation. The availability of open-source libraries such as Scikit-Learn, PyTorch, and Tensorflow has further lowered the barrier to deploying machine learning algorithms, allowing landscape architecture researchers and practitioners to apply these technologies efficiently without extensive computer science expertise.

The intrinsic alignment between machine learning’s capabilities and the core demands of landscape architecture practice makes its application highly viable. Traditional computer technologies in landscape architecture require manual design of operational rules, while machine learning possesses robust rule-learning and implicit rule-capturing abilities. By training on large datasets, it can derive potential patterns and use them to predict and simulate unknown data—an ability that dovetails with the landscape architecture workflow, which involves analyzing site conditions, identifying change laws, and intervening in site development through planning and design. While machine learning encompasses a diverse array of algorithms with distinct functions and focuses, the cross-over between algorithm design goals in computer science and landscape architecture practice remains limited, meaning few algorithms can be directly applied to design tasks. Instead, the mainstream application approach involves decomposing planning and design processes into discrete steps and developing tailored algorithms for each step’s objectives.

Classic machine learning algorithms with proven or potential applications in landscape architecture can be categorized by their problem-solving focus. Classification algorithms, including Naïve Bayes, Support Vector Machines (SVM), decision trees, and random forests, deliver high-efficiency data classification and are widely used for landscape land use classification. Deep learning algorithms such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Generative Adversarial Networks (GAN) boast powerful image recognition capabilities, making them ideal for rapid information extraction from remote sensing images and street view data; GAN and similar models also offer strong image generation potential, driving innovation in generative design. Regression algorithms like Principal Component Analysis (PCA) and logistic regression excel at correlating data and fitting functions, uncovering intrinsic relationships within datasets, and are thus applied to research on the driving forces of landscape pattern evolution. Additionally, natural language processing (NLP) algorithms such as tf-idf, word2vec, BERT, CRF, and LSTM have found their way into landscape perception research, enabling the classification of thematic and emotional information from large volumes of online text data to explore public perception laws of landscape spaces.

Based on the key stages of landscape architecture planning and design—information extraction, analysis and evaluation, and design generation—machine learning applications in the field fall into three core domains: site information extraction, landscape analysis and evaluation, and deep learning-based scheme self-generating systems, each addressing specific practical challenges and delivering tangible value.

Site information extraction is the foundational step of landscape architecture practice, and machine learning has revolutionized this process by efficiently processing fine-grained, diverse multi-source data collected via remote sensing and UAV technologies. Current research in this domain centers on three aspects: land use identification and classification, extraction of external site environmental information, and the integration of information extraction with other analytical workflows. Land use identification and classification based on remote sensing and machine learning has become highly mature, largely replacing traditional visual interpretation. Researchers leverage different algorithms to identify and classify land use types such as forestland, construction land, farmland, wetlands, and water areas at the macro scale for landscape pattern research, and recognize urban and rural infrastructure including roads, buildings, and water canals at the meso and micro scales to support ecological optimization. The technology also enables the identification of land feature boundaries such as urban contours and coastlines, facilitating studies on the dynamic changes of landscape patterns. Beyond land use types, machine learning can extract specialized environmental information by analyzing remote sensing spectral data or combining it with professional calculation formulas for inversion, including soil properties (moisture content, fertility, heavy metal content), agricultural properties (crop productivity, biomass), vegetation properties (forest stock volume, chlorophyll content), water body properties (turbidity, temperature), biological properties (habitat quality, invasive species), and climatic properties (land surface temperature). This rich information extraction capability has been widely applied in complex planning projects such as wetland planning, mine restoration, water ecological restoration, and protected area planning, enriching the cognitive dimensions of sites and providing a basis for in-depth site analysis. Importantly, in landscape architecture, site information extraction is not an end in itself but a precursor to further analysis; the technology is often integrated with landscape pattern analysis and site ecological optimization research to provide scientific guidance for planning and design, addressing practical site issues.

Landscape analysis and evaluation is a critical stage that links site information to design decisions, and machine learning has introduced quantitative and scientific methods to this process, moving beyond the limitations of traditional subjective evaluation. Its applications here are divided into three key areas: landscape pattern analysis based on remote sensing data, street view evaluation based on street view images, and post-occupancy perception evaluation based on text data, each focusing on different scales and research objectives. Landscape pattern analysis using machine learning primarily addresses three tasks: dynamic change analysis, driving force mechanism research, and future simulation and prediction. In driving force analysis, traditional statistical methods and basic machine learning algorithms are constrained by data types and linear relationship assumptions, struggling to handle the complex nonlinear relationships between multiple spatial variables. Advanced algorithms such as Boosted Regression Trees (BRT) and Artificial Neural Networks (ANN) can process diverse data types, handle missing data, and quantify the contribution rates of different driving factors to land use conversion across various thresholds, enabling in-depth exploration of landscape pattern change mechanisms. For example, researchers have used BRT to analyze the impact of natural, social, and economic factors on urban expansion in Yangzhou over different periods, revealing the dynamic relationship between driving factors and urban expansion at varying thresholds. In landscape pattern simulation and prediction, machine learning is often combined with artificial life algorithms such as Cellular Automata (CA) and multi-agent systems: machine learning mines historical change laws, while artificial life algorithms perform simulation generation based on these laws. The classic CA-Markov model, which combines Markov chain for statistical land change probabilities with CA for simulation, has been upgraded by integrating SVM, ANN, and other machine learning methods to incorporate social, economic, and natural driving factors as constraints, resulting in more realistic prediction outcomes with accuracy rates exceeding 80% in multiple case studies, such as the 2030 landscape pattern prediction for Ezhou City using the SVM-CA model.

Street view evaluation, driven by the popularization of street view big data from Google, Tencent, and other platforms, has become a vibrant research area with machine learning at its core. The technology supports three levels of analysis: street view spatial information analysis, street view perception evaluation information analysis, and street view perception evaluation law research. Spatial information analysis leverages image segmentation tools like Segnet to extract scene elements from street view images, calculate the percentage of landscape elements, and construct objective evaluation models, with applications in green view index (GVI) analysis and urban color research. Researchers have extracted green pixels from street view images to quantify urban street GVI values and mapped urban color distributions to support urban color control planning. To create more comprehensive evaluation models, machine learning is combined with subjective human scoring to establish integrated evaluation systems: researchers build labeled datasets through expert scoring or crowdsourcing, train machine learning models to summarize scoring rules, and develop urban space perception models. The Place Pulse dataset, evolving from version 1.0 (4,000 street view images, three perception dimensions) to version 2.0 (110,000 images across 56 global cities, six perception dimensions), has become a cornerstone for such research, enabling the development of models to predict street safety, vitality, and beauty. Going a step further, machine learning is used to explore the relationship between site elements and evaluation data, quantifying the impact of urban elements on landscape evaluation and providing scientific data support for planning and design. For instance, studies have found that increasing street windows and greenery can improve urban safety evaluations, validating Jane Jacobs’ “eyes on the street” theory, while other research has quantified the weight of over 150 elements such as bridges and waterways on six perception indicators including safety and vitality.

Post-occupancy perception evaluation based on text data addresses the challenge of analyzing large volumes of perceptual information from traditional questionnaires, in-depth interviews, and emerging online text big data. Traditional manual analysis methods are prone to errors due to sampling limitations, while machine learning can process tens of millions of online text data points to extract public perception laws. Its applications in this area include analyzing the emotional polarity of text to determine public satisfaction with site design and extracting keywords from large text datasets to generate keyword knowledge graphs based on co-occurrence relationships, which are then clustered to explore evaluation laws. Research is categorized by data source: questionnaire-based analysis, which allows for targeted questionnaire design to collect evaluation data and extract patterns using machine learning; online comment big data-based analysis, which crawls travel notes and comments from online platforms, extracts high-frequency keywords and emotional tendencies, and conducts evaluation from perspectives such as travel routes and supporting services; and social media image-based analysis, which leverages the rich annotation labels of social media images for algorithm training, extracting keywords and analyzing destination image perception. For example, researchers have used the deep learning-based text analysis tool DeepSentiBank to extract keywords from Beijing images on Flickr, conducting a tourism image perception evaluation of the city.

Deep learning-based scheme self-generating systems represent the cutting-edge application of machine learning in landscape architecture design, leveraging the strong image recognition and generation capabilities of deep learning to create design schemes based on learning from historical 2D images and 3D data. Due to the complexity of constraint conditions in planning projects, which increase with scale, these systems exhibit distinct application effects in large-scale planning and small-scale design, and research is thus divided into planning case generation and design case generation. Large-scale planning case generation currently focuses on single-element design such as road network planning, building functional layout, and building form planning, as multi-element comprehensive planning requires balancing complex real-world conditions—a challenge that remains unaddressed by current technology. Comprehensive planning schemes are typically created by designers overlaying multiple computer-generated single-element schemes. A classic example is the urban texture restoration experiment for the area surrounding Rome’s Termini Station, where researchers collected building data from the OpenStreetMap (OSM) database, trained a deep learning system, delimited research areas and set planning goals, and used deep learning to automatically divide building lots, match similar cases from the case library, and iterate to optimize outcomes that meet preset goals such as open space ratio and green space rate. To address the uncontrollability of deep learning-generated planning outcomes, researchers have integrated urban evaluation systems with generative design, as demonstrated in the planning of the northern extension of Wenzhou’s Central Green Axis. Here, a scheme self-generating system was trained on datasets from six cities, generating multiple schemes for road networks, spatial form, and building functional layout; an evaluation system with 13 indicators was then used to screen optimal schemes, which were overlaid to produce the final comprehensive plan.

Small-scale design case generation benefits from fewer constraint conditions and more flexible design content, making it a more active research area, though the randomness of outcomes remains a barrier to practical application and scientific research—compounded by the difficulty of quantifying abstract design goals. Plant configuration is a typical application scenario: researchers have developed self-generating systems using artificial neural networks, with deep learning responsible for plant species selection and positioning, and designers for data collection, model training, and output goal constraints. Combining evaluation systems and Ecotect environmental simulation technology, these systems generate plant configuration schemes by simulating site climate, calculating environmental factor indicators, and iteratively optimizing outcomes based on quantitative plant configuration theories and evaluation systems to determine the “optimal solution” for plant layout and species matching. Other researchers have adopted a different approach, treating randomly generated cases from self-generating systems as a scheme selection pool for secondary creation by designers. For example, GAN has been used to train 135,192 topographic data samples to build an automatic terrain generation model, producing terrain with natural mountain and river spatial characteristics; CycleGAN has been applied to the rapid generation of colored landscape architecture plans, enabling the conversion between layout color block diagrams and colored plans by training on 2,725 landscape layout schemes and 325 rendered images. Beyond academic research, commercial digital design companies have combined machine learning with other digital technologies to develop intelligent design systems that generate landscape plans in seconds based on input site conditions and automatically produce multiple building schemes with matching facades and optimized profit intensity based on design indicators, marking the initial commercialization of machine learning in landscape architecture design.

Despite the significant progress in machine learning applications in landscape architecture, the field is still in a stage of exploration and development, with several key challenges and limitations that need to be addressed. The application of machine learning in landscape architecture has evolved through three stages: phenomenon description, phenomenon analysis, and case generation, each with its own shortcomings. Phenomenon description, initially focused on remote sensing data interpretation for land use mapping, has matured into precise site information extraction, but its integration with subsequent landscape analysis workflows remains to be deepened. Phenomenon analysis, applied in landscape pattern analysis and landscape evaluation, has advanced from simple quantitative analysis to in-depth mechanism exploration, but landscape pattern simulation and prediction still lack close integration with driving force analysis; landscape evaluation, meanwhile, is hindered by issues such as insufficient available data, single data sources, incomplete evaluation systems, and the need for extensive manual data annotation and cleaning due to the multi-source and heterogeneous nature of online big data. Case generation, based on deep learning’s image generation technology, offers transformative potential for design innovation but is plagued by the “black box” mechanism—its uncontrollable and unexplainable operation process—leading to unpredictable planning and design outcomes that limit practical application. Additionally, the current AI landscape is characterized by “artificial input for intelligent output,” with large amounts of manual work such as data annotation required to address issues like data noise and missing values, turning AI, which aims to replace repetitive labor, into a source of additional manual work for designers.

Looking ahead, the future development of machine learning in landscape architecture will focus on four key directions, driven by technological advancement and practical demand, moving toward greater integration, intelligence, and practicality. First, the construction of landscape evaluation models based on multi-source data will address the current limitations of incomplete evaluation systems caused by data issues. With the development of multi-modal machine learning, transfer learning, and other advanced technologies, integrating multi-source data and combining diverse methods to build a comprehensive landscape evaluation system based on online big data will become a core research direction. Additionally, the combination of street view image big data and machine learning will evolve beyond single street view evaluation to form a comprehensive construction evaluation system including pedestrian flow analysis, green space management and maintenance assessment, and park construction evaluation, supported by mature big data and 5G technologies and full coverage of ground monitoring stations enabling short-interval, multi-temporal image data analysis.

Second, planning guidance based on intelligent landscape analysis technology will strengthen the integration of landscape pattern driving force mechanism research and simulation prediction. The current combination of machine learning (for site information interpretation and law calculation) and artificial life algorithms (for simulation prediction) provides a viable framework for an integrated AI planning analysis workflow. Future research will focus on two key points: integrating the complex impact mechanisms behind landscape patterns with prediction and simulation to improve the scientificity and accuracy of simulations, and building decision-making systems based on simulation and prediction to reverse-derive specific green space planning schemes, linking analytical results directly to practical design decisions.

Third, the construction of fully AI-driven planning and design methods will address the “black box” issue of deep learning and improve the practical value of generative design. Current deep learning-based generative design relies heavily on computer algorithms with limited designer control, resulting in highly experimental schemes with little practical significance. Future research will integrate existing analytical ideas with AI to enable computers to adapt to the complex constraint conditions of landscape architecture, developing more practical fully AI-driven landscape planning and design methods that balance algorithmic generation with practical design requirements.

Fourth, clarifying the role positioning of designers and AI will be crucial to the sustainable development of machine learning in landscape architecture. At present, machine learning has initially replaced some simple repetitive tasks in landscape analysis and evaluation with its efficient data processing and implicit rule-capturing capabilities, while deep learning-based generative design has further unleashed AI’s creative potential. However, data cleaning and selection, and artificial intervention in algorithms remain core bottlenecks. As technology advances and more advanced algorithms and data sources emerge, data-related challenges will be gradually resolved, allowing designers to shift their energy back to the core of planning and design—creative thinking and value judgment. For abstract design problems that are difficult for computers to quantify and judge, researchers have made initial explorations, such as quantifying abstract plant configuration theories for algorithmic constraints, using evaluation systems to guide scheme selection for abstract urban planning goals, and combining subjective evaluation systems with objective street view information analysis to manually revise machine learning perception evaluation results. These efforts represent important steps in addressing the limitations of AI in handling abstract design tasks. In the future, it will be essential to distinguish between repetitive, simple tasks that AI can efficiently perform and tasks requiring designer experience and subjective judgment, ensuring that AI technology serves as a tool to assist design rather than a substitute for human creativity. This balance is particularly important for landscape architecture, a discipline highly intertwined with human subjective aesthetics, to avoid the homogenization and standardization of planning and design outcomes caused by over-reliance on AI.

The application of machine learning in landscape architecture is not just a technological upgrade but a fundamental shift in the discipline’s research and practice paradigm, moving from subjective experience-driven to data-driven, quantitative, and scientific design. As multi-source data integration deepens, algorithms advance, and the collaboration between designers and AI matures, machine learning will play an increasingly important role in addressing complex ecological, social, and aesthetic challenges in landscape architecture, driving the discipline toward a more intelligent, sustainable, and human-centered future. The integration of machine learning with landscape architecture also highlights the broader trend of interdisciplinary convergence in the built environment industry, where the combination of engineering, computer science, ecology, and the arts will unlock new possibilities for creating high-quality, resilient, and livable urban and rural landscapes.

Author Information: Zhao Jing, Chen Ran, Hao Huichao, Shao Zhuang; School of Landscape Architecture, Beijing Forestry University, Beijing 100083, China Journal: Journal of Beijing Forestry University, Vol. 43, No. 11, November 2021 DOI: 10.12171/j.1000−1522.20200313

(Word count: 4896)