Land Use and Land Cover Change Assessment and Predictions in Flood Detention Areas of Yangtze River Basin Based on AIF-HOM-PLUS Model

Abstract

As global urbanization accelerates and economic development progresses rapidly, a series of ecological and environmental challenges have emerged. In certain countries, particularly in developing nations such as China, India, and Bangladesh, flood detention areas (FDAs) have been increasingly encroached upon by urbanization, resulting in growing conflicts between flood control functions and economic development. Therefore, accurately predicting urban expansion trends in these regions is considered essential for providing scientific guidance for sustainable regional development. In this study, the PLUS model was selected as the baseline based on comparative experiments. On this foundation, a novel AIF-HOM-PLUS framework was developed. In this framework, a new method, Adjacent Image Fusion (AIF), was proposed to reduce local temporal noise by utilizing adjacent multi-temporal data. Subsequently, Higher-Order Markov chains (HOM) were incorporated to capture complex temporal dependencies and long-term transition patterns. The Middle-Reach Yangtze River urban agglomeration (MRYRUA), including FDAs in the Yangtze River Basin (YRB), was selected as the study area, and LULCCs in 2035 and 2050 were predicted. The results showed the following: (1) among the basic models, the PLUS model exhibited the best performance, while the AIF method significantly improved its overall accuracy (OA) by 2%; (2) the area of impervious surfaces within the FDAs of the YRB will increase at an average annual rate of 1.29%, which pertains to the conflict between the United Nations Sustainable Development Goals (SDGs) 9.1 and SDG 11.a, which has become a critical issue that needs urgent attention; (3) the area of impervious surfaces in the MRYRUA will increase at an average annual rate of 1.3%, primarily at the expense of cropland and water bodies.

1. Introduction

As essential flood control infrastructure and significant residential areas, flood detention areas (FDAs) are vital to flood control in river basins and regional economic development [1,2]. However, with population growth and accelerated urbanization, substantial shifts have occurred in the land use and land cover change (LULCC) patterns within FDAs [3]. Among these changes, the most notable is the gradual encroachment of urban expansion into FDAs, particularly evident in flood-prone developing countries [4]. In recent decades, FDAs in regions such as the Yangtze River Basin (YRB) [5] and Huaihe River Basin [6] in China, the Chennai Basin [7] and Poisar River Basin [8] in India, and the Buriganga River Basin in Bangladesh [9] have been severely impacted by urban expansion. This phenomenon has not only transformed the natural and social landscape but also introduced new challenges to regional environmental safety and ecological balance [10].

The impacts of urban expansion on FDAs are multifaceted. First, urbanization-induced LULCC directly affects hydrological cycles and flood management processes [11,12]. Additionally, land previously used for natural water retention and flood attenuation has been converted to impervious surface, reducing the area’s natural flood storage capacity [13]. This, thus, increases flood risk and intensifies conflicts between flood management and economic development in FDAs [3]. Non-point source pollution generated by urbanization also significantly impacts water quality in these areas [14].

Furthermore, LULCC has far-reaching effects on local ecosystems. Habitat destruction and biodiversity loss are direct consequences of urban expansion [15,16]. Additionally, the degradation of ecological functions in FDAs—such as air and water purification and habitat provision—further intensifies regional environmental vulnerability [17]. Therefore, accurate prediction and understanding of LULCC in FDAs under urbanization are essential to developing effective land management strategies, mitigating environmental risks, protecting ecosystems, and promoting sustainable regional development.

In recent decades, research on land use change prediction has surged [8,18,19,20]. Among traditional methods, models such as the Cellular Automata (CA) model, CLUE-S model, multi-criteria decision models, Land Change Modeler (LCM), Future Land Use Simulation (FLUS) model, a patch-generating land use simulation (PLUS) model, and agent-based models are widely used in LULCC prediction [21]. Of these, the CA model is the most extensively employed, leveraging local spatial interactions to represent dynamic land use changes [22,23]. The core feature of the CA model is its rule-based system, where each cell’s state is determined by its neighboring cells, making CA particularly suitable for simulating spatially dependent processes like urban expansion, agricultural development, and forestland degradation [24]. Due to its simplicity and ease of interpretation, the CA model is applied across areas such as urban development, environmental conservation, and ecological studies [25]. In contrast, the CLUE-S model integrates socioeconomic factors, environmental constraints, and land use policies, providing a comprehensive understanding of LULCCs on a regional scale. This makes it particularly useful for evaluating the potential impact of policy changes on land use and land cover (LULC) [26]. The multi-level analysis capabilities of the CLUE-S model make it valuable for LULC planning and environmental protection strategies [27]. The FLUS model is an emerging model that combines CA with multi-criteria decision-making methods. By simulating various land use scenarios, the FLUS model aids in understanding and predicting future LULC trends [28]. The PLUS model, an advanced iteration of the CA model, utilizes machine learning (ML) to explore LULCC potential and incorporates traditional Markov chain models to improve prediction accuracy [29]. The PLUS model can constrain the quantity structure of LULC from the top down while introducing a novel threshold-dependent multi-type random patch seeding approach, making it highly applicable in LULCC prediction studies [30,31,32].

Hybrid LULCC models based on traditional methods are increasingly employed, combining the strengths of various models to achieve more accurate LULCC predictions [33,34,35,36,37,38]. Examples include the CA-Markov model [39,40,41], SD (System Dynamics)-CLUE-S model [26], CLUE-S-Markov model [27], and MOP-Dyna-CLUE model [42]. As climate change is a major driver of LULCCs, hydrological models such as SWAT are also applied in LULCC prediction [43,44].

With the advent of ML, applications in remote sensing are increasingly focused on tasks such as image processing and classification [45,46,47]. ML is gradually being explored for its potential in LULCC prediction [36], although complex LULCC modeling remains in its early stages [37]. Given the complexity of the driving factors behind LULCC under environmental and socioeconomic influences, further research on developing ML-based LULCC models to enhance predictive capabilities presents a promising area for exploration [38].

For FDAs, most studies have focused on flood control functions [1,48,49], while ignoring LULCC, which leads to a decrease in their flood control capacity. In fact, with economic development in the middle Yangtze River region, FDAs have been significantly encroached on [5]. Since the release of the Several Opinions of the State Council on Promoting the Rise of the Central Region in 2012, urban expansion in the middle-reach Yangtze River urban agglomeration (MRYRUA) has accelerated, resulting in a significant reduction in forestland [50]. Additionally, some FDAs have been encroached upon by urban development, presenting new challenges to regional flood control capacity [5].

The main land use types within FDAs include cropland, forestland, water bodies, and impervious surfaces. Among these, cropland is the dominant category and serves as a key pillar of regional economic development. Forestland is sparsely distributed, while water bodies—such as rivers, lakes, and ponds—play an essential role in flood storage [2]. In contrast, impervious surfaces such as buildings and roads have expanded rapidly with urbanization, significantly reducing the flood retention capacity of FDAs [1].

However, due to the complex spatial heterogeneity, varying proximity to urban centers, and specific policy constraints affecting FDAs, existing LULCC models often fail to capture the unique dynamics of these areas [51].

To address this research gap, taking the FDAs in the YRB as the major focal study area, with the middle-reach Yangtze River urban agglomeration (MRYRUA) serving as a comparative region. This study evaluates the performance of mainstream LULCC models and propose a fused prediction method that integrates continuous time-series data to reduce localized noise and enhance spatial-temporal accuracy. The hybrid model is then applied to simulate LULCC in 2035 and 2050.

The specific objectives of this study are as follows:

(1)

To evaluate commonly used LULCC models quantitatively;

(2)

To propose a fused LULCC prediction model based on continuous time-series data to improve the predictive accuracy of the selected optimal model, and then apply this hybrid model to forecast LULCC;

(3)

To analyze the future development trends and impacts of the FDAs and the MRYRUA based on the prediction results, and to propose corresponding policy optimization measures based the findings.

2. Materials and Methods

2.1. Study Area

The study area includes the MRYBUA and FDAs in the YRB, spanning an area of around 317,000 km². Over 90% of this area consists of cropland and forestland, and more than 90% of the FDAs in the YRB are situated within the MRYBUA. Figure 1 illustrates this region.

2.2. Data

LULCC arises from the dynamic interaction between natural forces and human actions, shaped by a wide range of natural, social, and economic factors. Considering the geographic features of the study area, the availability of relevant data, and insights from the established literature [33,52], 18 indicators were selected across three main domains: natural environment, built environment, and socioeconomic factors.

To ensure spatial consistency in model input, all raster-based data layers were resampled to a spatial resolution of 100 m using bilinear interpolation for continuous variables (e.g., elevation, precipitation) and nearest neighbor resampling for categorical variables (e.g., land use type, soil type). The natural environment indicators include slope, elevation, distance to water bodies, aspect, soil type, distance to forestland, precipitation, and temperature. These factors reflect topographic constraints and climatic conditions that influence land suitability and vegetation patterns [36]. Built environment indicators include distance to primary roads, highways, railways, national roads, train stations, settlements, city centers, land use type, and distance to major urban centers. These variables represent urban development potential and spatial accessibility [32]. Socioeconomic indicators include population density, regional GDP, and land value, which reflect development intensity and market-driven pressures that often shape land transformation in rapidly growing regions [53]. A summary of these driving factors and their data sources is provided in

Table 1. Data source.

Driving Factors			Data Type	Source (Accessed on 1 April 2024)	Year	Resolution
Natural Environment	Topography	slope	DEM	https://search.earthdata.nasa.gov/search?q=ASTER	2019	30 m
		elevation	DEM	https://search.earthdata.nasa.gov/search?q=ASTER	2019	30 m
		distance to water	land use	https://essd.copernicus.org/articles/13/3907/2021/ [54]	1990–2019	30 m
		aspect	DEM	https://search.earthdata.nasa.gov/search?q=ASTER	2019	30 m
		soil type	soil	https://www.resdc.cn/data.aspx?DATAID=260		1 km
	Amenity	distance to forest	land use	https://essd.copernicus.org/articles/13/3907/2021/	1990–2019	30 m
	Climate	precipitation	precipitation	https://www.resdc.cn/DOI/DOI.aspx?DOIID=96	2020	1 km
		temperature	temperature	https://www.resdc.cn/DOI/DOI.aspx?DOIID=96	2020	1 km
Built Environment	Transportation	distance to major roads	Road	https://www.openstreetmap.org/	2014–2020
		distance to railways	Road	https://www.openstreetmap.org/	2014–2020
		distance to expressway	Road	https://www.openstreetmap.org/	2014–2020
		distance to railway station	POI	https://www.openstreetmap.org/	2014–2020
	Land Use	distance to settlement	POI	https://www.openstreetmap.org/	2014–2020
		distance to city center (CBD)	POI	https://www.openstreetmap.org/	2014–2020
		land use (land cover)	land use	https://essd.copernicus.org/articles/13/3907/2021/	1990–2019	30 m
		distance to (big) city	POI	https://www.openstreetmap.org/	2014–2020
Socio-Economy	Population	population density	population	https://www.resdc.cn/DOI/DOI.aspx?DOIID=32	1995–2019	1 km
	Economy	GDP	GDP	https://doi.org/10.6084/m9.figshare.17004523.v1 [55]	1992–2019	1 km

, with the details of each driving factor shown in Figure 2.

2.3. Methods

To identify a model better suited for predicting LULCCs in the MRYRUA, three commonly used models were selected for experimentation: the CA-Markov model, the LCM model, and the PLUS model. Prediction in these models follows similar principles: LULCCs are first extracted from two preceding time periods, followed by the generation of LULCC potential maps using ML or expert evaluation. Finally, future LULCCs are forecasted using CA or other techniques. A three-period dataset is typically used, where the first two periods (1990 and 2005) are applied for model training, and the third period (2020) serves as a benchmark for assessing and comparing model performance.

After the most suitable predictive model was identified, it was optimized by incorporating the AIF and HOM modules to enhance the accuracy of LULCC predictions. As both the AIF and HOM models require multi-temporal data, input data spanning from 1990 to 2020 was utilized. A quantitative analysis was performed to evaluate the effectiveness of these modules. Finally, the optimized AIF-HOM-PLUS model was applied to predict LULCC in the study area for the years 2035 to 2050. The experimental flowchart is shown in Figure 3.

Table 2. Main land use change prediction methods.

Method	Platform	Main Steps	Limitations [36]
CA-Markov	Idrisi Selva/TerrSet	Land use change extracted using Markov chain, with land use change prediction via CA-Markov simulation.	Requires the creation of suitability maps, with manually set contributions from driving factors, introducing high subjectivity.
Land Change Modeler (LCM)	TerrSet	Land use change prediction using the LCM module, with main steps including change analysis (extracting land use change), development potential analysis (driver analysis and suitability maps), and change prediction.	Mainstream methods are available for each step, though significant subjectivity remains.
RFC-CA-Markov	PLUS	Similar to LCM: change analysis (extracting land use change), development potential analysis (using RFC to automatically derive relationships between driving factors and land use change, extracting only changed land), and change prediction (based on CA with random seeding).	Uses fixed methods for prediction; only parameter adjustments are possible, with no ability to modify the model itself.
Logistic–Markov–Spatial Allocation Algorithms	CLUE-S	Similar to LCM: change analysis (extracting land use change), development potential analysis (using logistic regression to derive relationships between driving factors and land use change), and change prediction.	Includes only spatial analysis, requiring additional simple mathematical models (e.g., Markov chain) for change analysis. Development potential analysis uses logistic regression, with ROC testing for validation of accuracy.

presents a comparison of the main land use change prediction methods.

2.3.1. The CA-Markov Model

In recent years, the CA-Markov model has been widely used to simulate the spatial and temporal dynamics of LULCC [30,56]. The CA model effectively captures spatial changes, as well as the complexity and unpredictability inherent in urbanization scenarios. However, due to its limitations, the CA model does not support time-series prediction. To address this, the Markov model has been integrated to quantitatively model land use change derived from multi-year land use transition matrices, thereby compensating for the limitations of CA model.

The core of the Markov chain lies in the transition probability matrix P, which represents the probability of transition from the current state to the subsequent state. For land use change, the transition matrix P is defined as follows:

$$P=\left[\begin{array}{cccc}{p}_{11}& p_{12}& \cdots & p_{1n}\\ p_{21}& p_{22}& \cdots & p_{2n}\\ ⋮& ⋮& \ddots & ⋮\\ p_{n1}& p_{n2}& \cdots & p_{nn}\end{array}\right]$$

(1)

In this context, $p_{ij}$ represents the probability of land use type i transitioning to type j, with the sum of elements in each row of the matrix equals 1.

The CA model operates as a dynamic system with discrete spatial units and states, where the state of each cell is updated simultaneously based on specified transition rules. The state of each cell at any given time depends on its state from the previous time step, expressed by the following formula:

$$S_i\left(t+1\right)=f(S_i\left(t\right),N_i\left(t\right),P)$$

(2)

$S_i\left(t\right)$ denotes the state of cell i at time t, $N_i\left(t\right)$ represents the set of neighboring states for cell i, P is the Markov transition probability matrix, and f represents the rule function for updating cell states, typically incorporating neighborhood rules, transition probabilities, and spatial constraints.

The CA-Markov model combines the transition probabilities of the Markov chain with the spatial rules of CA, enabling spatiotemporal simulation of land use change:

$$S_i\left(t+1\right)=\left\{\begin{array}{c}j, if p_{ij}·w\left(N_i\right)\ge Threshold \\ S_i\left(t\right), otherwise\end{array}\right.$$

(3)

In this formula, $p_{ij}$ represents the transition probability of a cell from state i to state j, $w\left(N_i\right)$ represents the weighting function of the neighborhood state, reflecting the influence of neighboring cells on the cell’s state transition, and Threshold is the conversion threshold that determines whether a cell undergoes a change.

Through the above formula, the CA-Markov model integrates temporal transition probabilities with spatial proximity, enabling precise predictions of land use change. In this study, the model was configured with a 5 × 5 cellular neighborhood and 15 iterations. This parameter setting was selected based on multiple experimental trials, where it consistently yielded the highest prediction accuracy. Similarly, for the other LULCC models used in this study, key parameters were also determined through repeated experiments to ensure optimal model performance under comparable conditions.

2.3.2. The Land Change Modeler (LCM)

The LCM is part of the Clark Labs software package (Terrset 2020) [57], with land transformation primarily structured into three core modules: (1) Change Analysis: extracting and analyzing past LULCC, creating change maps; (2) Transition Potentials: simulating the potential for land transitions—this module has six methodologies: a Multi-Layer Perceptron (MLP) neural network, Decision Forest (an implementation of the Random Forest ML approach), Logistic Regression, WNL (a Weighted Normalized Likelihood approach), SVM (Support Vector Machine), and Sim Weight (Similarity-Weighted Instance-based Machine Learning; (3) Change Prediction: projecting the course of change into the future. This module has two methodologies: Markov Chain and External Model.

To distinguish it from the machine learning approach utilized in the PLUS model, a statistical method, specifically logistic regression, was employed for the extraction of potential maps in the LCM model. During this process, land use categories with fewer than 40,000 cells were excluded from the potential analysis to avoid instability caused by insufficient sample size.

2.3.3. The PLUS Model

The PLUS model (PLUS v1.4.0) integrates a rule-mining framework according to a Land Expansion Analysis Strategy (LEAS) and a CA based on multi-type random patch seeds (CARS) [30].

The LEAS demands two dates of land use data for its transition analysis. After overlaying the land use data from two periods, only the changes observed in the latter period are extracted. This approach differs from the CA-Markov model, which requires recording both pre- and post-change data. By only extracting the updated land use data, the process of deriving change potential is streamlined, thereby improving computational efficiency. Subsequently, the random forest method is employed to extract conversion rules for each land use type, based on the driving factors and land use change maps.

The CARS module is a CA model that includes a patch-generation mechanism employing a range of random land use seeds. As a scenario-based land use simulation tool, this model combines both ’top-down’ influences (e.g., global land use needs) and ’bottom-up’ dynamics (e.g., local land use competition). During the simulation, land use demands modulate local competition with an adaptive coefficient, guiding land use changes to satisfy future requirements.

The parameters for both the LEAS and CARS modules were determined through comprehensive sensitivity analyses to ensure optimal performance and model stability.

For the LEAS module, the key parameters included the input sampling rate, mTry, and the number of decision trees. Specifically, the input sampling rate was tested in 20 experiments ranging from 0.001 to 0.9, with 0.01 identified as the optimal value. mTry was set to 18, corresponding to the total number of driving factors. The number of decision trees was selected based on nine experiments with values ranging from 10 to 100; a setting of 20 trees was adopted to balance model accuracy and computational efficiency.

For the CARS module, the parameters evaluated included the neighborhood size, the patch generation threshold, the expansion coefficient, and the percentage of seeds. The neighborhood size was tested using values of 3, 5, and 7, with the best performance achieved at a size of 3. The patch generation threshold was tested across nine values between 0.1 and 0.9, with 0.7 selected as optimal. The expansion coefficient and the percentage of seeds were each evaluated over 20 experiments covering the range 0.01 to 0.9, and values of 0.2 and 0.1 were selected, respectively.

These selected parameters were consistently applied in all subsequent simulations to ensure comparability across different scenarios and time periods. Additionally, the neighborhood weights and annual land demand were derived from historical land use data. As shown in

Table 3. Model accuracy under the selected parameter-setting parameter configuration.

#	Number of Regression Tree	Sampling Rate	Patch Generation Threshold	Expansion Coefficient	Percentage of Seeds	Kappa Coefficient	Overall Accuracy
1	20	0.01	0.7	0.2	0.1	0.811886	0.890456
R-1	20	0.01	0.7	0.2	0.1	0.811846	0.89038
R-2	20	0.01	0.7	0.2	0.1	0.811866	0.890438
R-3	20	0.01	0.7	0.2	0.1	0.811364	0.890095
R-4	20	0.01	0.7	0.2	0.1	0.811503	0.890236

Note: row 1 corresponds to the selected parameter configuration. Rows R-1 to R-4 represent repeated simulations under the same setting to assess model stability and convergence.

, four repeated simulations were performed under the selected configuration. The results showed a Kappa coefficient variance of 1.03 × 10⁻⁶ and an overall accuracy (OA) variance of 3.33 × 10⁻⁷, indicating that the model produced highly consistent outputs and demonstrated robust convergence.

2.3.4. The AIF-HOM-PLUS Model

The AIF-HOM-PLUS model is a three-stage integrated framework developed to improve the accuracy and robustness of long-term LULCC simulations. It combines the AIF, HOM, and PLUS modules in sequence.

First, the AIF method is used to integrate adjacent temporal land use maps. This approach enhances temporal consistency and reduces classification noise by prioritizing the most stable land use types across time, using frequency-based rules and global probability weighting. The result is a temporally smoothed and more reliable land use dataset, suitable for robust modeling.

Second, the HOM model is employed to predict future land use change quantities. Unlike traditional first-order Markov chains, HOM considers multiple historical land use states simultaneously to capture complex temporal dependencies and long-term transition trends. This enables a more accurate estimation of future land use quantities, especially under conditions involving spatial heterogeneity and long-term dynamics.

Finally, the predicted land use change quantities from the HOM model are input into the PLUS model, which performs spatial allocation based on patch-level simulation, neighborhood effects, and driving factors. This step generates spatially explicit and realistic land use maps for future target years.

By linking these three components, the AIF-HOM-PLUS model forms a cohesive, data-driven framework that integrates temporal smoothing, high-order trend modeling, and spatial simulation to support more credible LULCC forecasting.

The core principle of the AIF method lies in the land use type selection rules. The fundamental idea is to prioritize the selection of the dominant land use type based on historical data to simplify decision-making. When historical records do not indicate a dominant type, the method incorporates a global probability distribution to randomly select among candidate land use types. By integrating global information, this approach ensures a more rational selection while reducing the influence of local noise or occasional factors. Assuming three periods of land use data D1, D2, D3, the selection rule for the final land use result $D_{final}$ for a pixel can be expressed as follows:

$$D_{final}\left(i,j\right)=\left\{\begin{array}{c}{argmax}_{t\in \mathrm{T}}freq\left(t, X_{i,j}\right), if {max}_{t\in \mathrm{T}}freq\left(t, X_{i,j}\right)\ge 2, \\ t_k~P_t^{\prime }, if freq\left(t, X_{i,j}\right)=1 for all t\in X_{i,j},\end{array}\right.$$

(4)

In this formula, $freq\left(t, X_{i,j}\right)$ represents the frequency of pixel (i, j) belonging to a specific land use type t across the three periods. The notation $t_k~P_t^{\prime }$ indicates that $t_k$ is randomly selected from $X_{i,j}$ according to the probability distribution $P_t^{\prime }$.

In the first condition ${max}_{t\in \mathrm{T}}freq\left(t, X_{i,j}\right)\ge 2$, the land use type exhibiting the highest occurrence is directly selected. This reflects greater temporal stability and reduces uncertainty while aligning with the inherent dynamics of land use change. Moreover, it avoids adding unnecessary complexity or stochasticity.

In the second condition $freq\left(t, X_{i,j}\right)=1 for all t\in X_{i,j}$, when each type occurs only once in the three periods and no dominant type exists, the global probability distribution $P_t^{\prime }$ is employed to ensure sufficient utilization of global information.

The probability $P_t^{\prime }$ is defined as follows:

$$P_t^{\prime }={\displaystyle \frac{P_t}{{\sum }_{t\in X_{i,j}}{P}_t}}$$

(5)

In this context, $P_t$ represents the globally weighted and normalized probability. It is derived from the global frequency statistics of land use types and ensures that each land use type in $X_{i,j}$ is proportionally distributed.

After generating accuracy-optimized land use data through the aforementioned land use type selection rules, HOM is employed to predict future land use quantities. Unlike traditional Markov chains, which rely solely on the state information of a single time step, HOM incorporates multiple historical states, capturing the complex dynamic trends and spatial heterogeneity of LULCCs. This approach significantly improves the predictive accuracy of future LULC states. The prediction process mainly involves the following key steps:

Construction of Higher-Order State Sequences

Historical land use states of each pixel are combined to form a higher-order state sequence:

$$S=\left\{X_t,X_{t-1},\cdots ,X_{t-n+1}\right\}$$

(6)

In this formula, S represents the higher-order state sequence. The construction of S depends on the historical state of each pixel, and can be seen as an extension of the first-order model (which depends only on $X_t$) in the temporal dimension.

Statistic of Transition Probabilities

Calculate the probabilities of each higher-order state transition to the next sequence state $X_{t+1}$ based on historical data. The definition of transition probabilities is as shown:

$$P(X_{t+1}|S)={\displaystyle \frac{count(S\to X_{t+1})}{\sum _{X^{\prime }\in \mathrm{T}}count(S\to X^{\prime })}}$$

(7)

In this context, $\mathrm{P}\left(X_{t+1}\right|\mathrm{S})$ represents the transition probabilities from higher-order state sequence to target state, $count\left(S\to X_{t+1}\right)$ denotes the transitions from state S to state $X_{t+1}$ in historical data, $\sum _{X^{\prime }\in \mathrm{T}}count(S\to X^{\prime })$ represents the aggregate of transitions from state S to all possible states, and T represents the land use type set and include all possible type.

Construction of Higher-Order Transition Matrices

Transition probabilities from all possible higher-order state sequences S to target state $X_{t+1}$ are organized into a higher-order transition matrix. The matrix has a dimension of (kⁿ, k), where kⁿ represents all possible higher-order state combinations, and k corresponds to the overall count of land use type. In practical applications, due to the exponential growth of kⁿ, sparse matrix storage is often employed for optimization.

Prediction of Future State Distributions

Using the state distribution $\mathrm{P}\left(X_t\right)$ of the current time step and the higher-order transition probability matrix, the future state distribution is calculated iteratively as follows:

$$P\left(X_{t+1}\right)=\sum _{S}P\left(X_{t+1}|S\right)·P\left(\mathrm{S}\right)$$

(8)

In this formula, $\mathrm{P}\left(X_{t+1}\right)$ represents the state distribution at the target time step t+1, and $\mathrm{P}\left(\mathrm{S}\right)$ is the distribution of higher-order states at the current time step. By iterative prediction, state distributions for multiple future time steps can be obtained, making this predictive approach particularly suitable for long-term dynamic simulations.

Generation of Prediction Results

The quantities of transitions in each land use type are calculated based on the predicted state distributions.

Combined with models such as PLUS, this method not only generates the quantities of transitions for each land use type but also produces spatial distribution maps of land use changes based on the simulation results.

2.3.5. Validation

Overall accuracy (OA) and the kappa coefficient (KC) are used to evaluate the classification results of future land use changes [58], while user accuracy and producer accuracy assess the accuracy of each land use type. OA is defined as the percentage of correctly classified pixels out of the total classified pixels, calculated as follows:

$$OA={\displaystyle \frac{{\sum }_{i=1}^k{N}_{ii}}{N}}$$

(9)

In this context, OA represents overall accuracy, N denotes the overall number of samples, and $N_{ii}$ refers the diagonal elements of the confusion matrix, representing the count of correctly classified samples. k indicates to the number of land use data categories.

The KC measures the agreement between the classification results and those expected by random chance, adjusting for consistency due to random factors. Unlike OA, the kappa coefficient considers not only predictive accuracy but also the probability of correct classification due to chance, offering a more realistic assessment of model performance. The kappa coefficient approaches 0 when classification results are nearly random, and reaches 1 when classification results perfectly align with actual values. The formula is as follows:

$$K={\displaystyle \frac{P_0-P_e}{1-P_e}}$$

(10)

$P_0$ represents the observed overall accuracy (OA), while $P_e$ denotes the expected accuracy due to random chance, calculated as follows:

$$P_e={\displaystyle \frac{{\sum }_{i=1}^k{(N}_{i+}\ast N_{+i})}{N^2}}$$

(11)

$N_{i+}$ denotes the actual sample count for class i, $N_{+i}$ represents the predicted sample count for class i, and $N^2$ indicates the square of the total sample size.

3. Results

3.1. Comparisons of Results Among Three Basic Models

Figure 4 shows a comparison between the 2020 land use change prediction results from three models and the actual outcomes for the study area. Seven land use types are represented, with cropland and forestland covering more than 90% of the area. Among the three basic models, the results from the PLUS model in urban areas align noticeably more closely with the actual land use distribution.

The results shown in Figure 4 reveal certain limitations in the predictions made by all three models for the study area. A common shortcoming lies in their underestimation of urban expansion rates, failing to capture the rapid pace of actual urban growth. This discrepancy may be attributed to the significantly higher urbanization rate observed during 2005–2020 compared to 1990–2005. Additionally, the projected results from the CA-Markov model show substantial deviations from the actual data in the northern part of the study area, likely due to the inherent boundary effects of the CA model. In contrast, the PLUS model, also based on the CA framework, mitigates these boundary effects through the integration of patch mechanisms and dynamic updating rules [30]. These enhancements effectively reduce boundary discontinuities, significantly lowering boundary effects and improving the accuracy of the predictions. However, all three models underestimated the actual rate of urban expansion. Even the most accurate model, the PLUS model, predicted only 56% of the actual expansion rate.

Table 4. Comparison of OA and kappa coefficient among three models in MRYRUA.

Model	Kappa Coefficient	Overall Accuracy
CA-Markov	0.798372	0.883471
LCM	0.751077	0.854963
PLUS	0.811808	0.890409

presents a comparison of the overall accuracy across the three models, using the KC and overall OA as evaluation metrics. From these metrics, it was observed that the PLUS model outperformed the other two models, yielding higher KC and OA values. Specifically, the KC of the PLUS model exceeded that of the CA-Markov and LCM models by 1.7% and 8.1%, respectively, while the OA was higher by 0.7% and 4.1%, respectively.

Figure 5 shows a comparison between the predicted results from three models and the actual outcomes for the FDAs in the YRB in 2020. The main land use types in the FDAs of the YRB were cropland, water bodies and impervious surfaces due to the region’s flat terrain. The Dongxihu FDA near Wuhan was zoomed in on, and the prediction results from the PLUS model were significantly more accurate in comparison to the other two models, particularly in terms of urban land expansion accuracy.

In 2020, impervious surfaces accounted for 22% of the Dongxihu FDA, significantly exceeding the average of 4.4% across all FDAs. This disparity highlights substantial regional differences in terms of economic development within the FDAs.

A comparison of LULCC prediction results within the FDAs was conducted, as shown in

Table 5. Comparison of OA and kappa coefficient among the three models in the FDAs.

Model	Kappa Coefficient	Overall Accuracy
CA-Markov	0.693723	0.88208
LCM	0.683946	0.881763
PLUS	0.713626	0.888473

. The PLUS model continues to exhibit a higher KC and OA than the other two models. However, due to the scattered distribution and less distinct spatial characteristics of the FDAs, prediction accuracy in these areas is lower than that for the MRYRUA. For example, the KC of the PLUS model’s prediction within FDAs is 14% lower than that within the MRYRUA.

The PLUS model uses a random forest decision tree to capture the relationships between driving factors and land use change maps, enabling it to handle complex nonlinear relationships [30]. As a result, it achieved the best performance in this experiment, which involved 18 driving factors. In contrast, the CA-Markov model employs a Markov chain to capture land use transitions but does not directly incorporate driving factors, instead relying on suitability maps created for each land use type based on expert knowledge. Consequently, its performance was slightly lower than that of the PLUS model in this study. The LCM model employed logistic regression, a statistical method, to analyze the relationship between driving factors and land transitions. While logistic regression is effective for linear relationships, significant nonlinear spatial relationships between driving factors and land transitions led to lower prediction accuracy for this model [18].

In this study, the PLUS model outperformed all other models, and was therefore chosen for the subsequent phase of the experiment. A fusion model was then developed based on the PLUS model to enhance its predictive accuracy.

3.2. Comparisons of Results About the AIF-HOM-PLUS Model

Figure 6 presents the misclassified land use types under four different prediction scenarios in 2020. The scenarios were predicted using four models: Markov-PLUS, HOM-PLUS, AIF-Markov-PLUS and AIF-HOM-PLUS, respectively. From the magnified regions in Figure 6a,b, it is apparent that the PLUS model demonstrates higher classification accuracy for impervious surfaces and forest land when combined with the HOM method compared to the Markov chain. Furthermore, a comparison of the upper-right corners of the magnified regions in Figure 6b,d revealed a slight improvement in the prediction accuracy for forest land. This improvement is achieved after optimizing the land use data with the AIF method before applying the PLUS model. However, since the PLUS model already has a high baseline accuracy, some differences were not particularly noticeable after the improvement methods were applied in the figures. The accuracy comparisons are analyzed in detail in the Section 4.

The misclassification results of land use types in the FDAs were not considered significant enough to be presented here. However, additional details and results can be found in the Section 4.

3.3. Future Prediction Results

The LULC in the study area was predicted using the AIF-HOM-PLUS model with the results presented in Figure 7. From 2005 to 2020, the process of urban expansion was highly evident, particularly around the three central city clusters in the MRYRUA. In the predicted results, the trend of urban expansion is projected to persist, especially from 2020 to 2035. However, the pattern of urban expansion is projected to shift from the suburbs of metropolitan areas to smaller cities. Meanwhile, as the base of impervious surfaces increases, its growth becomes less pronounced.

The MRYRUA comprises 31 cities [32], with the three provincial capitals exhibiting the highest levels of economic development, while the other cities show relatively weaker economic growth. Figure 8 presents the urban expansion rates for each city in the agglomeration from 1990 to 2050. The cities are ranked based on their total GDP in 2020, from highest to lowest.

Between 1990 and 2005, urban expansion was mainly concentrated in the three core cities of Wuhan, Changsha, and Nanchang, which together accounted for 22.2% of the total urban growth in the MRYRUA. Among these cities, Changsha experienced the fastest growth speed, with its impervious surfaces increasing by 1.4 times over 15 years. From 2005 to 2020, the expansion rates of the cities became more balanced. However, between 2020 and 2035, urban expansion began to slow down, and growth started to shift toward smaller cities. Notably, Pingxiang (City Index 27), Jingdezhen (City Index 28), and Xiantao (City Index 29) saw the fastest growth. From 2035 to 2050, the expansion rate significantly decreased across all cities.

Figure 9 presents the predicted results for the FDAs in the YRB. The future development of the FDAs will be primarily characterized by urban expansion. From 2020 to 2035, expansion is expected to follow the trend of previous development. However, between 2035 and 2050, the pace of urban expansion is projected to significantly slow, likely nearing saturation. This trend mirrors the development pattern observed in the MRYRUA.

There are 42FDAs in the YRB [59], several of which have become key areas for urban expansion. According to the gravity model theory in spatial economics [60], economic activities and the expansion of impervious surfaces in a region are inversely proportional to the distance from the central city. Therefore, we ranked the FDAs based on their distance from the city, from closest to farthest, and analyzed the land use changes in each FDA.

As shown in Figure 10, between 1990 and 2020, most FDAs experienced rapid growth in impervious surfaces, while between 2020 and 2050, the growth rate significantly slowed in most FDAs. The prediction results indicate that, from 2020 to 2035, the growth of impervious surfaces in some FDAs is expected to be negative, while these same areas show an unusually rapid growth rate from 2035 to 2050, which clearly does not align with actual expectations.

A likely explanation lies in the small size and baseline urban area of certain FDAs, where pixel-level uncertainty or edge effects can lead to disproportionate percentage changes. To address this, we implemented a post-processing correction: for FDAs exhibiting one period with a non-positive growth rate and another with a positive rate, we recalibrated both using a compound-growth-constrained adjustment. Specifically, the corrected values preserve the original total compound growth while bringing each stage closer to the respective median values (0.1463 for 2020–2035 and 0.1259 for 2035–2050).

The corrected growth trajectories, shown in Figure 10b, provide a more consistent and realistic representation of future urban expansion in the FDAs, especially those susceptible to localized prediction noise.

Figure 11 illustrates the projected changes in the impervious surface ratio across FDAs after applying the corrected growth rates. By 2035, the impervious surface ratios in the Baitanhu FDA (FDA Index 2) and Dongxihu FDA (FDA Index 7) are expected to exceed 25%, suggesting the necessity of reassessing their designation as FDAs. Furthermore, by 2050, the impervious surface ratios in the Junshan FDA (FDA Index 1), Wuhu FDA (FDA Index 11), and Quyuan FDA (FDA Index 28) are projected to surpass 10%, potentially imposing considerable pressure on local flood control capacity and ecological security.

4. Discussion

4.1. A Comparative Analysis About AIF-HOM-PLUS Model

Table 6. Comparison of OA and kappa coefficient for AIF-HOM-PLUS model in the MRYRUA.

Model	Kappa Coefficient	Overall Accuracy
Markov-PLUS	0.909141	0.946874
HOM-PLUS	0.911578	0.948354
AIF-Markov-PLUS	0.9271	0.95738
AIF-HOM-PLUS	0.927306	0.957532

shows the comparison of the OA across the four models, using the kappa coefficient and OA as evaluation metrics in the MRYRUA. It is obvious that the AIF-HOM-PLUS model shows the best performance with highest kappa coefficient and OA compared with other three models. Compared to the lowest-performing Markov-PLUS model, the AIF-HOM-PLUS model achieved improvements of 2% in the kappa coefficient and 1.1% in OA. However, when compared with the AIF-Markov-PLUS model, the AIF-HOM-PLUS model demonstrated marginal improvements of less than 0.1% in both the kappa coefficient and OA. This finding suggests that the improvement in predictive accuracy of the AIF-HOM-PLUS model is primarily due to the AIF component. Given that the primary role of the AIF method is to reduce localized noise in the imagery, it further implies that the original LULC data may contain substantial local noise.

Figure 12a shows a comparison of producer accuracy among the four models for the MRYRUA. For forestland, water bodies and impervious surfaces, the AIF-Markov-PLUS model demonstrate higher producer accuracy, aligning well with the predominant land use types in the MRYRUA. In contrast, the AIF-HOM-PLUS model exhibits greater accuracy in identifying cropland, shrubland, and barren land.

Figure 12b presents a comparison of user accuracy among the four models for the MRYRUA. Among the four models, the HOM-PLUS model achieves the highest prediction accuracy for shrubland and impervious surfaces, while the AIF-Markov-PLUS model demonstrates the highest accuracy for cropland, forestland, water bodies, and barren land. The AIF-HOM-PLUS model achieves the highest accuracy only for grassland.

Table 7. Comparison of OA and kappa coefficient for AIF-HOM-PLUS model in the FDAs.

Model	Kappa Coefficient	Overall Accuracy
Markov-PLUS	0.84512	0.937123
HOM-PLUS	0.85139	0.939857
AIF-Markov-PLUS	0.863461	0.944768
AIF-HOM-PLUS	0.867047	0.946636

shows a comparison of the OA across the four models, using the kappa coefficient and OA as evaluation metrics in the FDAs. The AIF-HOM-PLUS model demonstrates the highest accuracy in the kappa coefficient and OA among the four models. However, because of the fragmented distribution and less distinct spatial features of the FDAs, the prediction accuracy in these areas is lower compared to that of the MRYRUA. While the model performs well in general, further improvements such as localized calibration or the inclusion of more detailed ancillary data may enhance its reliability in these more complex zones.

Figure 13a illustrates the contrast in producer accuracy among the four models for the FDAs. For water bodies and impervious surfaces, the AIF-Markov-PLUS model demonstrates the highest producer accuracy. However, the HOM-PLUS model achieves the highest accuracy in identifying forestland, while the AIF-HOM-PLUS model performs best in identifying cropland.

Contrary to accuracy, the user accuracy among the four models presents different results. As shown in Figure 13b, the HOM-PLUS model achieves the highest prediction accuracy for impervious surfaces, while the AIF-Markov-PLUS model demonstrates the highest accuracy for cropland. The AIF-HOM-PLUS model achieves the highest accuracy for forestland and water bodies.

The fusion model we propose—AIF-HOM-PLUS—not only improves the KC by 2% over the widely used PLUS model [30], but also introduces two methodological innovations. First, the AIF approach enhances input data quality by integrating temporal signals, which reduces noise and better captures spatial dynamics [61]. Second, the incorporation of HOM improves temporal dependency modeling [62], especially for rapidly changing urban areas. These methodological advances increase the reliability of LULCC simulations and hold promise for application in other flood-prone or fast-developing regions [63].

4.2. Analysis of Prediction Results

Before presenting the projected land use changes for 2035 and 2050, we evaluated the historical performance of different models to quantify systematic biases in urban expansion prediction.

Specifically, the baseline Markov-PLUS model reproduced only 56.00% of the actual urban expansion between 2005 and 2020, indicating a substantial underestimation of impervious surface area dynamics.

In contrast, our proposed AIF-HOM-PLUS model showed a marked improvement, reproducing approximately 64.89% of the actual urban expansion over the same period.

Based on this improved performance, we applied a correction factor of 1.54 (i.e., 1/0.6489) to the original predictions from the AIF-HOM-PLUS model in order to construct an adjusted upper-bound scenario.

As shown in

Table 8. Urban impervious surface area projections for 2020, 2035, and 2050: original vs. corrected estimates based on historical model bias (unit: km²).

Year	Observed Area	Model-Predicted Area	Bias-Corrected Area
2020	13,470.33	-	-
2035	-	13,780.83	13,821.81
2050	-	13,948.50	14,684.87

, the original model predicted impervious surface areas of 13,780.83 km² in 2035 and 13,821.81 km² in 2050. After correction, the adjusted estimates increased to 13,948.50 km² and 14,684.87 km², respectively.

It is important to emphasize that this correction was applied only to the total predicted urban area, without modifying the spatial allocation or land use transition patterns simulated by the model. All subsequent spatial and transition analyses are based on the unadjusted AIF-HOM-PLUS outputs.

Figure 14a illustrates land use change transitions within the MRYRUA from 1990 to 2020. Over recent decades, these regions have undergone significant transformations, with the most notable being the expansion of impervious surfaces. Between 1990 and 2005, the area of impervious surfaces increased by 3186.79 km², of which 2805.68 km² were converted primarily from cropland. Since 1990, numerous land management policies have been introduced, including the Basic Farmland Protection Regulation in 1994 and the Grain for Green reforestation program initiated in 1999. However, stronger protection of cropland only began in 2004 with the establishment of the “1.8 billion mu redline” for cropland preservation. Consequently, from 2005 to 2020, although 4472.54 km² of cropland was still converted to impervious surfaces, the region saw a considerable increase in the conversion of forestland to cropland, with 11,348.96 km²—nearly 10% of total forestland—converted to cropland. At the same time, 7201.41 km² of cropland was converted to forestland.

Figure 14b depicts projected land use change transitions in the MRYRUA from 2020 to 2050. Based on the prediction results, the expansion of impervious surfaces is expected to growth at a rate of 1.3% per year, while substantial reciprocal conversions between cropland and water bodies are decreasing. With the expected slowdown in national economic growth, the pace of economic development in the MRYRUA is also projected to decelerate, leading to a slower increase in impervious surfaces. Meanwhile, the area of cropland within the region is expected to continue decreasing, which may pose new challenges to local food security.

Figure 15a illustrates land use transitions in the FDAs of the YRB from 1990 to 2020. Between 1990 and 2005, the area of cropland in the FDAs decreased by 262.98 km², while water bodies and impervious surfaces increased by 138.01 km² and 136.39 km², respectively, largely at the expense of cropland. This trend suggests that FDAs were gradually developed during this period. From 2005 to 2020, the area of impervious surfaces within the FDAs expanded by 225.54 km², primarily at the expense of cropland. This indicates that cropland in the FDAs is gradually being encroached upon by impervious surfaces, which may affect the flood control capacity of these areas [5].

Figure 15b presents the projected land use change transitions in the FDAs of the YRB from 2020 to 2050. According to this projection, the area of impervious surfaces in the FDAs is expected to increase by 153.07 km² between 2020 and 2035, and by only 54.47 km² from 2035 to 2050. In this forecast, no additional interventions, such as policies or ecological protections, were applied to the land use change model. With a projected increase in impervious surfaces of less than 10%, the trend is beneficial for flood management policies, as the limited land use change implies minimal disruption to flood retention functions.

Over the next 30 years, the trend of urban expansion in the FDAs is expected to continue, following the pattern observed in the MRYRUA. From 2020 to 2035, the area of impervious surface within the FDAs is expected to increase by approximately 28.4%, while within the MRYRUA, the impervious surface area is projected to grow by 31.7%. However, the increase in impervious surface within the FDAs is primarily concentrated in FDAs surrounding the city. Figure 16 presents the changes in the proportion of impervious surface within the MRYRUA, the FDAs, and typical FDAs from 1990 to 2050. Among these results, the Dongxihu FDA and Wuhu FDA are located near the city, while the Honghuzhong FDA and Jianxin FDA are located further from the city. The Jingjiang FDA is most important FDA in the YRB. These spatial patterns suggest that FDAs surrounding urban areas are increasingly incompatible with their original flood control function, which presents significant new challenges for regional flood risk governance.

In addition, our comparative analysis with historical land use data reveals that most conventional LULCC models significantly underestimate the pace of urban expansion [51]. Even with the enhanced AIF-HOM-PLUS model, the predicted rate of change remains lower than the actual observed rate. This underestimation implies that urban encroachment into FDAs may be more severe than previously projected, posing greater risks to regional flood control systems [64]. These findings highlight a critical insight: without proper adjustment, standard models may fail to detect early signs of urban pressure on flood-sensitive zones. However, our model still faces limitations in capturing small-scale spatial variability and abrupt land transformations, especially in fragmented FDAs [65]. Further model refinement is needed to address these challenges.

4.3. Impact of Land Use Change on Economic Development and Flood Control Policies

4.3.1. Impact of Land Use Change on Economic Development and Flood Control Policies in FDAs of YRB

FDAs in China exhibit a range of distinctive characteristics that set them apart from conventional land use zones [5]. Functionally, they serve a dual role—providing emergency flood storage during extreme events, while also supporting agriculture, residential settlements, and, in some cases, industrial activities during normal periods [1]. This dual-functionality results in inherent policy tensions between flood control objectives and socio-economic development goals [66]. In addition, FDAs are often spatially fragmented, institutionally regulated at multiple levels, and closely tied to major river systems, making their management highly complex [3]. These characteristics have profound implications for both land use planning and flood policy design.

Since the 1950s, 42 FDAs have been gradually established in the middle and lower reaches of the Yangtze River [59]. As a crucial component of flood control infrastructure, these FDAs have been repeatedly activated during major flood events. For instance, during the historically significant 1954 flood—the largest in nearly a century—the Jingjiang FDA was activated three times, ensuring the safety of the Jingjiang and southern embankments. However, over the past 70 years, the functional structure of these FDAs has undergone substantial changes.

The results for land use changes from 1990 to 2020 reveal significant cropland loss within the FDAs: from 1990 to 2005, 262.98 km² of cropland was lost, followed by an additional 209.58 km² between 2005 and 2020. These combined losses account for 5.4% of the total cropland area within the FDAs, signaling a profound shift in the industrial structure of these agricultural zones, as agricultural activities are gradually being replaced by impervious surfaces and urban expansion. This transformation has not only diminished the flood control capacity of the FDAs but has also altered the region’s economic foundation.

An accompanying challenge is the increased difficulty in activating these FDAs, particularly those that have been converted into economically developed regions. Activating such areas would entail significant economic losses and pose safety risks to the local population. For instance, the Dongxihu FDA near Wuhan has seen its impervious surfaces increase from 3% in 1990 to 21% in 2020 [67], and it is now a high-tech industrial park, housing numerous businesses and permanent residents. The economic and safety impacts of utilizing this area as an FDA would be substantial, underscoring the need for current policies to account for the emerging risks associated with land use changes in flood management decisions.

Nevertheless, FDAs in the YRB still play a significant role in flood control, particularly during major flood events, where they effectively alleviate flood pressures and reduce flood risks in the middle and lower reaches of the river. However, it has observed that the rate of urban expansion within FDAs is expected to slow down in the future, which will likely reduce the conflict between economic development and flood control functions within these areas. However, numerous challenges still exist within FDAs.

Firstly, among the 42 FDAs in the YRB, most of the flood control infrastructure remains underdeveloped [3]. For FDAs located farther from urban centers, their primary function is still flood control, and their flood retention capacity can be further enhanced. The “Room for the River” program in the Netherlands serves as a notable example [68,69]. This program improves flood discharge capacity by lowering the elevation of floodplain areas, creating water buffer zones, relocating dikes, deepening side channels, and expanding floodways. These measures have enabled the tributaries of the Rhine River to handle floodwaters of up to 16,000 cubic meters per second, while also improving the overall ecological environment of the riverine areas.

For FDAs closer to urban centers, some of these areas have already become important regions for urban expansion. Given the limited feasibility of reverting these areas to flood control facilities, we can draw on the UK’s Sustainable Drainage Systems (SuDS) approach [70,71]. The UK government has implemented several tidal FDAs in London [72], where tidal waters are directed into low-lying areas through sluice gates, thus reducing the flood threat to the city center. This approach is characterized by converting urban green spaces and wetland parks into adaptable FDAs, balancing ecological functions with flood control needs. Additionally, community involvement in co-developing flood management strategies with the government has helped raise public awareness and acceptance of FDAs.

Furthermore, the regions where FDAs are located often contain large areas of wetlands, many of which have been poorly managed during development, leading to resource waste and environmental degradation. In contrast, the United States has a more developed wetland protection policy, which employs market mechanisms such as the Wetland Mitigation Banking Program [70,71]. This program involves restoring, creating, or enhancing wetlands to offset unavoidable impacts on other wetlands, thereby promoting the protection and restoration of wetlands and FDAs. The wetland management strategies for FDAs in the YRB can draw on the US wetland mitigation banking policy to encourage the sustainable development of wetland areas.

4.3.2. Impact of Land Use Change on Economic Development and Flood Control Policies in MRYRUA

The MRYRUA, covering the provinces of Hubei, Hunan, and Jiangxi, has emerged as a crucial pillar of the Yangtze River Economic Belt since the issuance of the “Development Plan for the Middle Reaches Yangtze River Urban Agglomeration” by the National Development and Reform Commission in April 2015 [73]. In recent years, with the acceleration of urbanization, land use patterns have undergone significant changes. While this phenomenon has driven economic growth, it has also exerted a profound impact on ecosystems and the hydrological environment.

Through an analysis of land use changes in the MRYRUA from 1990 to 2020, we identified two distinct phases of development with this region. From 1990 and 2005, urbanization accelerated significantly, leading to a rapid increase in impervious surfaces, which doubled in area over 15 years. The majority of these impervious surfaces originated from the conversion of cropland. While this large-scale transformation contributed to swift economic growth in the region, it also disrupted the hydrological cycle by reducing rainwater infiltration and increasing surface runoff. Consequently, there was a heightened risk of urban flooding during periods of heavy rainfall, leading to both increased frequency and severity of flood disasters. Furthermore, the decline in cropland resulted in habitat fragmentation that exacerbated ecosystem degradation. The concomitant rise in industrial pollution and urban waste also poses substantial threats to water and air quality.

However, between 2005 and 2020, the pace of urbanization experienced a deceleration. While the growth of impervious surfaces primarily stemmed from agricultural land and forested areas, it is noteworthy that the total area of cropland actually increased during this period. This trend suggests that effective implementation and enforcement of cropland protection policies occurred within the MRYRUA. Between 1999 and 2001, China experienced three consecutive years of declining grain production, which failed to meet domestic demand. In response to this challenge, in 2005, the State Council issued the “Provincial Government Cropland Protection Responsibility Assessment Guidelines”, linking cropland protection policies directly to the performance evaluations of local officials. This mechanism facilitated the effective implementation of cropland protection policies. Although the expansion of cropland played a positive role in ensuring food security, it often resulted in significant conversion of forestland. The substantial reduction in forestland weakened the resilience of the regional ecosystem, diminished the ecological functions of forests in climate regulation, water conservation, and wind and sand prevention, and increased the risk of soil erosion. Consequently, despite rapid economic development in this region, both flood control capacity and ecological functions are now at significant risk of degradation.

Subsequently, through predictions of land use changes in the MRYRUA from 2020 to 2050, it is found that the area of impervious surfaces is expected to continue increasing. However, this growth will occur at a significantly slower rate compared to the past three decades. This trend suggests that the ongoing implementation of cropland protection policies has led to an increase in cropland that aligns with expectations, while the growth of impervious surfaces has kept pace with economic development. As China transitions from a phase of rapid economic growth to high-quality development, the pace of economic development is gradually decelerating, which suggests that the expansion of impervious surfaces will similarly slow down. Nevertheless, it is important to note that the continued growth in both cropland and impervious surfaces comes at a cost: reductions in forestland and water bodies persist. This situation poses challenges for ecosystem resilience and necessitates further enhancement of regional flood control capacities.

In response to the hydrological impacts caused by the expansion of urban impervious surfaces, the State Council of China issued the “Guidelines for Promoting Sponge City Construction” in 2015, signifying the comprehensive implementation of sponge city development. The initiative aims to enhance urban stormwater management and improve the urban water environment through natural infiltration, storage, purification, and the efficient use of rainwater [74,75]. Given the frequent flooding, rapid urban expansion, and ecosystem degradation in the MRYRUA, this region has become a critical area for the application of sponge city construction.

In the MRYRUA, cities such as Wuhan and Nanchang have been designated as pilot cities for sponge city development [76]. This initiative has catalyzed the implementation of various measures, including rain gardens, permeable pavements, and ecological green spaces. For instance, in Wuhan, sponge city construction has been carried out in the Qingshan pilot area, where green infrastructure such as lakes is utilized for stormwater retention, and the role of blue-green infrastructure, including source control facilities and natural lake systems, has been strengthened. Model evaluations indicate that the proportion of the sewer network capable of accommodating a rainfall event with a return period of three years has increased from the current 17% to approximately 50% [77]. In Nanchang, leveraging the Poyang Lake Ecological Zone, the city has effectively mitigated urban flooding and improved water resource efficiency through the renovation of the urban drainage system, the establishment of an ecological water network, and the promotion of sponge-type buildings and green spaces. Since being successfully selected as one of the “Second Batch of Systematic and Comprehensive Sponge City Construction Demonstration Cities” in May 2022, Nanchang has developed a total sponge city area of 159.96 square kilometers by the end of 2023, increasing the proportion of the impervious surfaces area meeting sponge city standards from 21.5% to 42.5% (Source: Nanchang City government, China, 1 February 2025).

Although the concept of sponge cities has been implemented in the MRYRUA, it continues to encounter numerous challenges, including high governance costs [78], difficulties in policy coordination [79], lack of public participation, and unreasonable spatial distribution of structural practices [80]. To advance a more scientific and systematic model for urban water management, international experiences are essential. Globally, countries adopt various approaches to urban flood management. Developed countries such as Germany, Australia, and Singapore possess extensive expertise in managing urban flooding.

In Germany, the concepts of compact cities and Low-Impact Development (LID) are actively promoted [81]. The focus of these concepts is on curbing urban sprawl, encouraging land-intensive use, and increasing urban green space. Additionally, ecological drainage is also employed, with measures such as green roofs, permeable roads, and sunken green spaces to reduce stormwater runoff. Furthermore, rainwater storage facilities are constructed to improve urban water resource recycling and alleviate flood control pressure.

Australia has adopted the Water-Sensitive City model in urban construction [82], which integrates water resource management with urban planning, as well as addressing social and environmental needs. This approach aims to tackle challenges such as urban flooding, drought, and water scarcity. The concept emphasizes the comprehensive management of urban water resources, optimizing water cycle systems, and using green infrastructure to cope with the water resource pressures brought about by urbanization. It has been widely implemented in cities such as Melbourne and Perth [83].

The urban development process of Singapore promotes the “ABC Waters Programme” (Active, Beautiful, Clean Waters) (Sources: PUB Singapore—ABC Waters Programme, 1 February 2025). The Programme integrates urban landscape design, ecological restoration, and water resource management [84]. The primary measures include establishing an urban water cycle system, optimizing rainwater collection, treatment, and reuse to improve water resource efficiency; creating waterfront spaces, enhancing shoreline landscapes, and achieving coordination between ecological protection and urban development; and implementing rainwater management zoning, developing differentiated strategies for rainwater management based on various land use types, and enhancing the adaptability of drainage systems.

The MRYRUA encounters various challenges, including land use changes, water resource management issues, and ecosystem degradation amid urban expansion and economic development. Drawing on successful international experiences such as the compact city model in Germany, the water-sensitive city model in Australia, and the ABC Waters Programme in Singapore, the MRYRUA can enhance its flood control capacity, optimize land use, and promote sustainable urban development. This can be achieved through compact city planning, green infrastructure, smart water resource management, cross-regional ecological governance, and policy innovation. In the future, through collaborative efforts among the government, markets, and society, the MRYRUA has the potential to transition toward a greener, smarter, and more resilient urban development path.

In addition to providing policy recommendations for the YRB and the MRYRUA, this study also contributes to the broader discourse on integrating land change science with floodplain governance [85]. By aligning empirical prediction models with real-world planning instruments, future research can help bridge the gap between environmental modeling and spatial policy implementation [86]. The experience gained from FDAs can be further extended to other transitional land systems—such as peri-urban wetlands and multifunctional floodplains—in developing countries [87]. Strengthening this link between spatial data analytics and adaptive land governance will be crucial for advancing interdisciplinary approaches in land system science and sustainable urbanization [88].

5. Conclusions

With the acceleration of urbanization, the conflict between economic development, flood control, and ecological protection in FDAs has become increasingly evident. Current research on FDAs has primarily focused on flood control capacity [49], while less attention has been given to their other aspects. Therefore, predicting the future LULCC of FDAs and assessing the impacts of changes in economic development, flood control capacity, and ecological protection on both the FDAs and surrounding urban areas are crucial. This study aims to predict the future development of FDAs using an appropriate LULCC model, providing data support future development planning. A hybrid prediction model was proposed in this study, based on existing models, and applied to the LULCC analysis of FDAs and nearby urban agglomerations. The main conclusions were as follows:

(1)

Model improvement: The PLUS model outperformed the CA-Markov and LCM models in prediction accuracy. With the integration of our proposed AIF method and the HOM approach, its OA improved by 2%.

(2)

Management for FDAs: FDAs in the YRB, especially those near urban centers, are increasingly occupied by impervious surfaces. This trend may compromise their flood control function. Therefore, differentiated strategies should be applied: rural FDAs may benefit from approaches like the Room for the River program, while urban FDAs could adopt solutions such as SuDS to balance development and flood management.

(3)

Urbanization trend in the MRYRUA: Impervious surfaces in the MRYRUA are projected to grow by 1.3% annually, mainly at the expense of cropland and water bodies, increasing flood and ecological risks. International approaches like LID, the Water-Sensitive Cities framework, and Singapore’s “ABC Waters Programme” can help improve sponge city strategies and support adaptive urban planning.

Although a suitable LULCC prediction model for MRYRUA was identified, and its OA was improved by 2% in our study, the reliance of the AIF method on global frequency may result in the oversight of local spatial characteristics, potentially leading to misclassification in heterogeneous areas such as FDAs. Furthermore, the model’s dependence on continuous time-series data may limit its applicability in areas with sparse data. Future research should focus on the enhancement of this method by incorporating local spatial information. Future research should aim to refine the method by integrating local spatial indicators and developing flexible mechanisms that accommodate variable data availability. In addition, increasing the number of decision trees in the LEAS module may further improve the stability and predictive accuracy of the model, which will also be explored in subsequent studies.

With the gradual improvement of the flood control system in the YRB [3,89], the flood retention capacity of the reservoirs in the upper and middle reaches has been steadily increasing, leading to reduced reliance on FDAs. In the future, the planning and development of FDAs in the YRB may require more comprehensive considerations. For other developing countries or regions with FDAs [7,8,9,90], the lessons from this study can provide guidance for sustainable land use planning. By advancing predictive models and spatial analytics tailored to FDAs, this research contributes to the broader effort of integrating flood resilience into urban development and land system science.