A cosine adaptive particle swarm optimization based long-short term memory method for urban green area prediction

Introduction

Urban green spaces manifest as critical elements within the urban fabric, contributing significantly to both ecological sustainability and human well-being. These verdant areas offer a multitude of benefits, ranging from environmental amelioration to social and psychological advantages. Ecologically, urban greenery plays a pivotal role in enhancing biodiversity, providing habitats for various species, and maintaining ecological balance within urban ecosystems. This biodiversity is not only vital for ecosystem health but also contributes to the resilience of urban areas against environmental stressors. Rao et al. (2022) evaluates the accessibility of urban green spaces in 254 Chinese cities using the two-step floating catchment area method, revealing significant disparities in accessibility both within and between cities, especially within walking or cycling distances. It associates higher social status, reflected in housing prices, with greater access to green space services, highlighting pronounced inequities. The research offers targeted recommendations to improve the distribution and use of urban green spaces, aiming to enhance their ecological, social, and economic benefits in urban ecosystems.

From an environmental perspective, green spaces are instrumental in mitigating urban pollution. Ghahramani et al. (2021) undertook a novel application of Artificial Intelligence (AI) techniques in the realm of urban green spaces (UGS) assessment by implementing a unified topic modeling approach. Investment in park green space can improve the quality of life for urban residents. Paul et al. (2020) conducted a study on large cohorts in Ontario, Canada that found that increased exposure to urban green space is associated with a reduced risk of developing major neurological conditions, specifically dementia and stroke. Stuhlmacher, Kim & Kim (2022) conducted a study to analyze the relationship between the development of park and non-park green spaces and the likelihood of gentrification in Chicago, using satellite imagery and demographic data, revealing that green space investments’ impact on gentrification varies with time and neighborhood characteristics. Hu et al. (2022) investigates the relationship between residential green space, neighborhood walkability, and atherosclerosis in urban settings, analyzing data from 2021 adults with suspected coronary heart disease (CHD). Utilizing advanced statistical methods, the study evaluates the impact of green space and walkability on coronary artery calcium scores, highlighting the complex interplay between urban development, environmental factors, and cardiovascular health. Hogendorf et al. (2020) found that while increasing distance to green spaces slightly reduced leisure walking time and slightly increased walking for active commutes among Dutch adults, there was no significant association with cycling, indicating a limited impact of green space proximity on walking and cycling behaviors. Basu & Nagendra (2021) found that public spaces like parks exhibit significant gender and income inequalities, resulting in uneven access to green space. Jiang, Stickley & Ueda (2021) explored the relationship between green spaces and suicide mortality in Japan, revealing that the protective impact of greenery against suicide varies with urbanity and demographic factors, indicating the potential of green spaces in suicide prevention strategies.

Related work

The development of urban green spaces is influenced by various factors. Kwartnik-Pruc & Trembecka (2021) examines the involvement of local governments in shaping public green spaces, essential to sustainable urban development, with a specific focus on Poland. It analyzes municipal strategies in property acquisition, legal frameworks, and policy formulation for green space enhancement, using data from the Polish Central Statistical Office to track the progression of public green spaces in major Polish cities, notably Krakow. The study not only (Aly & Dimitrijevic, 2022) investigated the distribution and types of green spaces in Krakow but also assessed the municipality’s initiatives in expanding and safeguarding these areas against urban development, emphasizing the significance of green space accessibility for residents. The findings provide a comprehensive evaluation of Krakow’s management of green spaces and reflect broader trends in public green space development, offering valuable insights for future urban environmental policies and sustainable planning practices. Feltynowski & Kronenberg (2020) investigates the proportion and types of urban green spaces in five towns in Poland, comparing various data sources to reveal discrepancies in the perception and management of green spaces, especially highlighting the underestimation of green areas in smaller towns when relying solely on public statistics.

Long short-term memory (LSTM) models are used in traffic flow prediction (Bharti, Redhu & Kumar, 2023). In 2020, Stessens et al. (2020) developed a GIS-based model to infer the perceptions of naturalness, quietness, and spaciousness by users of public green spaces based on the attributes of the green spaces. Das (2022) examined factors contributing to the environmentally unjust development and management of organized green spaces in three Indian cities (Bhubaneswar, Cuttack, and Kolkata), and evaluated various strategies aimed at achieving environmental justice. Huang, Wu & Cheng (2021) demonstrated that a strategically designed ecological network effectively augments landscape connectivity and curtails fragmentation, consequently elevating the quality of the urban ecological environment and fostering the sustainability of urban green spaces.

Predicting the area of urban green spaces is critically important for several interconnected reasons. Accurate green space predictions are vital for public health and well-being, as these areas provide essential spaces for recreation, relaxation, and socializing, thereby enhancing the mental and physical health of city residents. Long short-term memory models are often used for predicting tasks like landslide displacement (Duan, Su & Fu, 2023) and air pollutant prediction (Luo & Gong, 2023). Zhou, Zuo & Zhao (2022) proposed a large-scale urban land subsidence prediction method. Usharani (2022) introduced an improved loss function within an LSTM neural network to enhance the accuracy of sea surface temperature predictions at various time horizons around India. Luo & Gong (2023) developed an innovative ARIMA-WOA-LSTM model, aimed at enhancing the precision in predicting air pollutants, thereby optimizing air pollution management. By employing ARIMA to process the linear components of pollution data and utilizing a whale optimization algorithm-enhanced LSTM (WOA-LSTM) for predicting non-linear elements, the research significantly advances the model’s efficacy. The study further substantiates the superiority of the ARIMA-WOA-LSTM model over various counterparts in terms of pollutant prediction accuracy, model prediction precision, and stability, through comparative analyses with several other models. Du et al. (2022) proposed a Particle Swarm Optimization (PSO) based LSTM model for urban water demand forecasting.

This has also catalyzed research on enhanced particle swarm algorithms. Metaheuristics methods (Aranha et al., 2022; Sörensen, 2015) based on a diverse range of natural, artificial, and social behaviors or patterns. These methods are often used to solve optimization problems in various fields (Velasco, Guerrero & Hospitaler, 2022, 2024). For instance, Guo et al. (2024) proposed a novel dynamic learning method for breast cancer image classification. In 2022, Guo et al. (2022), Tian et al. (2022) discussed novel strategies for swarm intelligence. In 2023 and 2024, novel metaheuristic methods (Guo et al., 2023a, 2023b; Zhou et al., 2024) are proposed for single-objective optimization problems.

To increase the accuracy of prediction, this article presents a novel cosine adaptive particle swarm optimization-based long short-term memory model for urban green spaces prediction. The major contribution of this article can be summarized as follows:

1. A new cosine adaptive strategy applied to the evolution of particle swarm algorithms enhances the search capability of traditional particle swarm algorithms. The cosine adaptive strategy enhances the particle swarm’s ability to escape local optima and increases the diversity of the particle swarm.

2. The Cosine Adaptive Particle Swarm Optimization (CAPSO)-LSTM is used in urban green area prediction, and experimental results demonstrate that the proposed CAPSO-LSTM can accurately predict the area of urban green spaces, providing significant assistance in urban construction planning.

Materials and Methods

Long short-term memory cosine adaptive particle swarm optimization

Long short-term memory (LSTM) networks, a pivotal subset of recurrent neural networks (RNNs), are designed to overcome the limitations of traditional RNNs in learning long-term dependencies. Distinct for their unique architecture, LSTMs are equipped with specialized units called gates: the forget gate, input gate, and output gate. These gates effectively regulate the flow of information, allowing the network to retain or discard data over intervals of varying lengths. This capability is particularly beneficial in tasks involving sequential data, such as natural language processing, time series analysis, and speech recognition. By addressing the issue of vanishing gradients, a common problem in standard RNNs, LSTMs facilitate more efficient and robust learning of dependencies across extensive time lags, making them a fundamental tool in the field of deep learning. The architecture diagram of LSTM is shown in Fig. 1.

In the domain of LSTM neural networks, it has been demonstrated that the accuracy of prediction outcomes can be enhanced through the modification of the learning rate and the neuronal count. The CAPSO-LSTM algorithm is introduced, representing an amalgamation of LSTM with the CAPSO algorithm. This integration is achieved by employing CAPSO to optimize both the learning rate and the neuron quantity within the LSTM framework, thereby augmenting the precision of the prediction model. The algorithm’s methodology encompasses the following specific steps:

• Step 1: Normalization is performed by Eq. (2), which ranges from 1 to −1. Normalization eliminates the undesirable effects caused by odd sample data and improves the convergence of the training network.

  (2) $${\mathrm{X}}^{\mathrm{\prime }}=2\ast (\mathrm{X}-\mathrm{m}\mathrm{i}\mathrm{n}(\mathrm{x}))/(\mathrm{m}\mathrm{a}\mathrm{x}(\mathrm{x})-\mathrm{m}\mathrm{i}\mathrm{n}(\mathrm{x}))-1$$where X is the input data, min(x) is the minimum value of input data, max(x) is the maximum value of input data, X′ is the result after normalization.

• Step 2: In the CAPSO algorithm part, the optimization objects are the initial learning rate and the number of neurons of LSTM, and the fitness function is a mean square error (MSE). The fitness function of the CAPSO is shown in Eq. (3):

  (3) $$MSE=\frac{1}{n}\sum _{i=1}^n{(Y_i-{Y_i}^{\mathrm{\prime }})}^2$$

  where n is the number of samples, The $Y_i$ is the true value and ${Y_i}^{\mathrm{\prime }}$ is the predicted value.

• Step 3: In the LSTM training process, the optimal number of neurons and the initial learning rate obtained by the CAPSO algorithm will be used as parameters in the LSTM.

• Step 4: Start model prediction and calculate model prediction error.

The flowchart of CAPSO-LSTM is shown on Fig. 2.

Experiments

Results and discussion

Prediction results are shown in Table 4. In the year 2017, the LSTM model exhibited the least prediction error at 0.22%, but this error margin escalated progressively over the years, culminating at 16.92% in 2021. This trend may indicate a degradation in the LSTM model’s long-term prediction capability or its inability to adapt effectively to evolving trends in the data.

The PSO-LSTM model shows an initial error of 0.63% in 2017, increasing to 11.57% by 2021. Despite its fluctuating error rates across the observed years, the PSO-LSTM generally outperforms the standalone LSTM model, suggesting that Particle Swarm Optimization (PSO) contributes to a more refined parameter selection for the LSTM, thus improving its predictive accuracy.

Overall, the CAPSO-LSTM model exhibits consistently stable and superior predictive performance, with error rates decreasing from 2.09% in 2017 to 1.02% in 2021. This improvement signifies the success of the CAPSO in optimizing the hyperparameters of the LSTM, thereby enhancing the model’s ability to capture the trends in urban green space area changes. Notably, in the year 2021, the CAPSO-LSTM model’s prediction error is significantly lower than that of the other models, highlighting its potential and practicality in handling such data.

Results of mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) are shown in Table 5. The performance of the LSTM, PSO-LSTM, and CAPSO-LSTM models was comprehensively evaluated. It was observed that the CAPSO-LSTM model consistently outperformed the other models, as evidenced by its significantly lower MAE (2,242.83), RMSE (2,431.38), and MAPE (0.03), indicating a superior predictive accuracy. The improvements in forecasting precision can be attributed to the optimization of hyperparameters using the CAPSO algorithm, which has been demonstrated to enhance the LSTM model’s capability to predict urban green space areas more reliably. Conversely, the LSTM model, devoid of optimization techniques, was found to have the highest error rates, suggesting a lesser degree of reliability in its predictions. The PSO-LSTM model, which utilized Particle Swarm Optimization, showed a moderate performance improvement over the standard LSTM model, as reflected by its intermediate values of MAE and RMSE, and a lower MAPE than that of the LSTM model. Train and prediction results of LSTM, PSO-LSTM, and CAPSO-LSTM are shown in Figs. 3–5.

Compared to traditional metaheuristic algorithms, CAPSO is specifically designed for the parameter training of recurrent neural networks, utilizing a cosine oscillation strategy to enhance the particle swarm algorithm’s ability to escape local optima.

Moreover, the dataset for urban green space is collected on an annual basis, leading to a scarcity of training and testing data. When trained with such limited datasets, the CAPSO-LSTM model exhibits a distinct advantage over comparative algorithms. However, it is imperative to acknowledge that a deficiency in training data may precipitate a decline in model accuracy, accompanied by an escalation in both MAE and RMSE. This presents a formidable challenge to the CAPSO-LSTM framework, underscoring the importance of devising strategies for efficient training leveraging short-term data as a critical future research avenue.

In summary, the CAPSO-LSTM model significantly improves the accuracy of urban green space area predictions by optimizing the hyperparameters of the LSTM. This has important practical applications in urban planning and green space management, as accurate predictions can assist urban planners in making better-informed decisions to maintain or increase urban greenery.

Conclusion

The imperative of accurately forecasting urban green space areas is acknowledged for its critical role in environmental conservation, urban planning, and ensuring societal well-being in urban areas. In this study, a new predictive model termed CAPSO-LSTM has been proposed, utilizing the CAPSO algorithm to fine-tune the hyperparameters of an LSTM network, thereby augmenting the model’s performance.

In the conducted simulation experiments, the CAPSO-LSTM model’s performance was found to be superior to that of both the PSO-LSTM and the traditional LSTM models. The specific experimental outcomes revealed that the CAPSO-LSTM model achieved an MAE of 2,242.83, an RMSE of 2,431.38, and a MAPE of 0.03.

However, it is recognized that the scarcity of training data, due to the annual collection cycle of urban green space-related data, has resulted in elevated MAE and RMSE among the algorithms tested. Based on these experimental outcomes, the development of predictive models capable of achieving high accuracy with limited training samples is identified as a critical area for future research.

In conclusion, the introduction of the CAPSO-LSTM model marks a significant step forward in the predictive modeling of urban green spaces. The employment of the CAPSO algorithm for hyperparameter optimization within the LSTM framework has proven effective, as evidenced by the simulation results. Future work should aim to address the identified limitations by exploring more computationally efficient algorithms that maintain the model’s predictive accuracy while minimizing resource consumption. Additionally, further studies could investigate the model’s adaptability to various urban contexts and the integration of additional predictive variables, such as climate patterns or urbanization rates, to enhance the model’s robustness and applicability to real-world scenarios. This future research will be critical for advancing the practical application of the CAPSO-LSTM model in sustainable urban development and green space management.

Supplemental Information