Introduction

Settlement consolidation plays a critical role in offsetting cropland loss. Historical and expected cropland losses to human settlements undermine the achievement of Sustainable Development Goals (Bren d'Amour et al., 2017; van Vliet, 2019), including Zero Hunger (SDG 2) and Sustainable Cities and Communities (SDG 11). However, rural depopulation worldwide suggests a window of opportunity to counteract this loss by implementing settlement consolidation, a process that typically involves reclaiming low-density settlements into cropland (Janus and Markuszewska, 2019; Daskalova and Kamp, 2023).

Settlement consolidation has been intensively adopted by many countries (e.g., China, Germany, Poland). Empirical studies demonstrate that such initiatives not only increase cropland area but also facilitate large-scale farming operations and reduce the risks associated with food deficit (Jürgenson, 2016; Janus and Markuszewska, 2019; Dong et al., 2022). Among these, China is the first large economy to formulate a nationwide planning for settlement consolidation, known as the “National Land Consolidation Planning (2014–2020)”, to address rising food demand and diminishing cropland quantities (Liu et al., 2014). This strategy is further reinforced by the “National Land Planning of China (2016–2030)”, which envisions the continuation of long-term consolidation efforts (NLPC, 2017). Current consolidation practices reveal that the costs and benefits of settlement consolidation exhibit considerable spatial heterogeneity (Jin et al., 2017; Zang et al., 2021), influenced by a wide range of biophysical and social-economic factors (e.g., elevation, temperature, and population density). Therefore, to guarantee effective consolidation, there is an urgent need to delineate the potential for settlement consolidation at large scales that allow for interventions tailored to local contexts.

To effectively upscale local consolidation solutions to address regional challenges (Lambin et al., 2020), a fundamental step is to assess the distribution of consolidation potential, despite challenges posed by current assessment methods. While scholars have employed indoor interviews or multi-criteria evaluation models to assess consolidation potential (Gao et al., 2022a; Tao et al., 2024), existing methods meet significant challenges when applied at larger spatial scales. Specifically, interviews are typically confined to small-scale studies or individual projects due to the time and financial investments required (Cay and Uyan, 2013; Colombo and Perujo-Villanueva, 2019; Pan et al., 2023). In contrast, multi-criteria evaluation models, which rely on the subjective weighting of various assessment dimensions, are constrained by their inherent dependence on expert knowledge, thereby limiting their broader applicability (Lin et al., 2020; Zhao et al., 2024a). These limitations underscore the need for robust, scalable methodologies that can objectively and efficiently evaluate consolidation potential across diverse spatial contexts.

Machine learning methods, particularly supervised classification approaches, have been widely applied to map consolidation potential (Xu et al., 2019; Yang et al., 2024). However, these models fall short when it comes to feeding them with merely single-labeled references, as conventional classification models require both positive and negative samples. In our case, references to suitable areas for consolidation (positive samples) are sufficient, but definitive references for areas unsuitable for consolidation (negative samples) are absent. A potential solution is to treat all unconsolidated areas (unlabeled samples) as negative samples during model training. However, this approach introduces significant biases into outputs due to the hidden presence of settlements with consolidation potential within the unlabeled dataset. This leads to class confusion—where positive samples are misclassified as negative—and feature misrepresentation, reducing the model’s predictive accuracy. To address this challenge, an alternative approach is to build models with bootstrapped training datasets by combining all positive samples and a down-sampled subset of unlabeled ones with replacement (Bepler et al., 2019; McDonald et al., 2021). This method, namely Positive-Unlabeled (PU) bagging, has proved effective in dealing with the aforementioned issue in fields such as remote sensing and biological sciences (Lu and Wang, 2021; Liu et al., 2024). By leveraging PU bagging, it becomes possible to produce robust and reliable evaluations of consolidation potential, particularly when combined with comprehensive datasets of consolidated settlement samples from diverse physical and socioeconomic contexts.

Ground-sourced data is crucial for conducting independent and bottom-up assessments of settlement consolidation potential. Although such data can be manually collected for individual villages through visual interpretation (Liu et al., 2023; Zhao et al., 2024b), its large-scale availability is severely hindered by time and labor constraints. Furthermore, existing time-series datasets on land use are not suitable for acquiring reference data, as they are based on the assumption that human settlements are permanent and cannot be legally or practically converted into other land uses (Yang and Huang, 2021). As a result, these datasets fail to illustrate the nuanced conversion of settlements into cropland. The absence of reliable ground-sourced data impedes the analysis of the empirical relationship between the spatial distribution of consolidated settlements and explanatory variables, hindering the identification of settlements with consolidation potential. However, recent policy initiatives in China, particularly in Zhejiang Province, offer a promising solution. Since 2018, the local government has allocated substantial budgets to settlement consolidation and synthesized the spatial details of completed consolidation projects. By integrating these records with explanatory variables, we can now estimate consolidation potential at larger spatial scales with unprecedented precision, which represents a significant advancement in data-driven, scalable, and actionable insights into settlement consolidation.

This study conducts a spatially explicit assessment of settlement consolidation potential in the case of Zhejiang Province, China. By leveraging a machine learning method, we address challenges inherent in the consolidation potential assessment. To develop a robust PU bagging-based modeling framework, we use spatial data from 524 consolidation projects collected from local authorities as references. Model performance was optimized through hyperparameter calibration and the selection of an appropriate base classifier, validated using a 5-fold cross-validation process to minimize uncertainties. By informing location-specific consolidation practices, this study contributes to more efficient and equitable land management strategies, addressing critical challenges in food security.

Theoretical framework and methodology

Conceptual framework and research design

We developed a conceptual framework for settlement consolidation based on the Driving Force-Land Change (DF-C) model (Hersperger et al., 2010). This theoretical model suggests that underlying drivers are closely related to land use changes, particularly in response to external factors such as land conflicts or policy interventions (Meyfroidt, 2016; Pendrill et al., 2024). Our refined framework enhances this theoretical model by specifically focusing on the complexity of land conflicts between settlements and cropland (Fig. 1).

Fig. 1
figure 1

Conceptual illustration of settlement consolidation.

Full size image

In our model, settlement consolidation potential is determined by multiple variables. Based on a brief literature review (Supplementary Table 1), we categorized these variables into three groups: biophysical, social-economic, and landscape-associated. Biophysical factors, including elevation, slope, soil properties, and climate, are critical as they largely determine the quality of reclaimed cropland (Jin et al., 2017; Du et al., 2018). Social-economic factors reflect dynamics such as population decline, deteriorating building conditions, and limited transport accessibility, all of which have been shown to significantly increase consolidation potential (Xu et al., 2019; Lin et al., 2020). Landscape-associated factors, including metrics such as the aggregation index, edge density, and patch density of cropland and settlements, as well as a cropland reclamation priority index proposed by Wang et al. (2021), captures the spatial configuration of land uses. These factors are highly relevant to settlement consolidation since isolated or fragmented land uses are more prone to convert into surrounding land uses, as observed in historical land-use changes (Zhou et al., 2017; Tao et al., 2020).

The primary principle in evaluating consolidation potential is to identify settlements with characteristics similar to those settlements that have successfully undergone consolidation. Following this principle, we employed a PU bagging approach to uncover the empirical relationship between the spatial distribution of settlements and explanatory variables. Specifically, historically consolidated settlements were designated as positive samples, while the remaining settlements were interpreted as unlabeled samples. Random subsamples selected from the unlabeled samples are integrated with positive samples to train PU bagging models, ultimately yielding a map of consolidation potential (see section “Assessment of settlement consolidation potential” for technical details). Moreover, this study assessed settlement consolidation potential in rural land systems of Zhejiang Province, which was delineated following an established land system classification framework (see Supplementary Materials).

Settlement consolidation offers not only a pathway to address food security challenges through increased cropland but also holds significant implications for environmental sustainability. In the discussion section of this paper, we explored these multifaceted impacts by briefly discussing (1) the potential crop production gains from newly reclaimed land, (2) the economic costs of consolidation, and (3) the broader environmental implications of settlement consolidation practices.

Case study

Zhejiang Province of China was selected as the study area for two primary reasons (Fig. 2). On the one hand, settlement consolidation in Zhejiang Province is a pioneering initiative, serving as a flagship model in China (Liu et al., 2023; Yu et al., 2023). This province faces crucial land conflicts between crop production and settlement development, driven by economic development and population growth (Yue et al., 2022). To address these challenges, local authorities launched an innovative policy, “Three-Year Action Plan for Comprehensive Land Consolidation and Ecological Restoration”, in 2018 (Bryan et al., 2018). Investigating settlement consolidation potential in Zhejiang is crucial for providing scientifically grounded insights that can optimize land management practices and offer global lessons for regions facing similar conflicts. On the other hand, the availability of detailed spatial data on past consolidation projects in Zhejiang Province allows for a rigorous data-driven assessment of consolidation potential. The comprehensiveness of this dataset ensures the robustness of our research design and allows for more reliable predictions of future consolidation opportunities.

Fig. 2: Study area.
figure 2

The blue polygons are implemented settlement consolidation projects (2018–2022). Credit: World Ocean Base by ESRI.

Full size image

Historical settlement consolidation across Zhejiang Province showed diverse spatial patterns between 2018 and 2022 (Fig. 2). During this period, 5754 hectares (ha) of settlements were consolidated, with Jiaxing (2107 ha), Huzhou (1071 ha), and Wenzhou (618 ha) leading consolidation efforts, accounting for 67% of the province’s total. In contrast, Jinhua (331 ha), Lishui (182 ha), and Shaoxing (63 ha) exhibited lower levels of settlement consolidation. Notably, settlement consolidation practices were more spatially aggregated in Huzhou and Jiaxing compared with Quzhou and Taizhou.

Assessment of settlement consolidation potential

Assessing the consolidation potential involves distinguishing settlements that exhibit such potential from those that do not. The ideal modeling process would rely on a training dataset including settlements that possess and lack consolidation potential, respectively (Bepler et al., 2019; Geng et al., 2024). However, at large spatial scales, it is challenging to effectively and efficiently determine which settlements cannot be consolidated. As a result, this modeling task represents a classic Positive-Unlabeled learning (PU learning) problem, characterized by merely having positive and unlabeled samples that are contaminated by hidden positive examples (Lee and Liu, 2003; Liu et al., 2003).

Both positive and negative samples are typical modeling inputs for traditional supervised learning methods. While we can adapt supervised learning methods to address the PU learning problem by treating all unlabeled samples as negative samples, this simplification potentially introduces biases in predictions and cannot effectively address PU learning issues (Geng et al., 2024; Gu et al., 2024). In the context of evaluating settlement consolidation potential, unlabeled settlements simultaneously include settlement possess and lack consolidation potential. Regarding unlabeled settlements with consolidation potential as negative samples can lead to class confusion and feature misrepresentation, substantially diminishing prediction accuracy (Lu and Wang, 2021; Stupp et al., 2021; McDonald et al., 2021).

PU bagging is a pervasive PU learning algorithm to discriminate positive samples from unlabeled ones and has been widely applied in land use and remote sensing studies (Lei et al., 2021; Liu et al., 2024). The underlying mechanism of PU bagging involves building multiple datasets by combining all positive samples and a down-sampled subset of unlabeled ones with replacement (Fig. 1). Base classifiers (e.g., random forest or decision tree) are then trained on bootstrapped datasets, and predictions are recorded for the out-of-bag (OOB) samples—those not included in the random sample. The final predictions are obtained by repeating the above steps multiple times and averaging model outputs (Elkan and Noto, 2008; Zhao et al., 2022). Existing studies demonstrate that PU bagging significantly reduces interference from hidden positive samples on prediction through bootstrap procedures, thereby performing better than traditional supervised learning methods (Mordelet and Vert, 2014; Bepler et al., 2019; McDonald et al., 2021).

Therefore, we adopt PU bagging to assess settlement consolidation potential. We chose the decision tree and random forest as the base classifier for PU bagging. The decision tree models were set to have a maximum of 100 trees, meanwhile the subsampling and feature sampling of random forest were set as 0.5. In this study, pixels representing historically consolidated settlements were designated as positive samples for PU bagging modeling, whereas other settlements in rural areas were considered unlabeled samples.

PU bagging training would subsequently determine hyperparameter settings, such as bagging number, down-sampling ratio, and the probability threshold for classifying settlements possessing consolidation potential, through 5-fold cross-validation. Specifically, we evaluated classifier performance across varying numbers of bags (1, 10, 50, and 100), each comprising all positive samples and a random subset of unlabeled samples. The number of unlabeled samples in each bag was controlled by the down-sampling ratio, which ranged from the number of positive samples up to five times that number.

Hyperparameter settings of PU bagging would be optimized to control model complexity, reduce standard deviation, and maximize the mean of the modified F1 score across folds. The modified F1 score can address evaluation challenges in PU learning, accommodating the absence of negative samples and mitigating biases in precision or F1 score calculations (Lee and Liu, 2003; Norlin and Paulsrud, 2017; Geng et al., 2024). The modified F1 score is calculated using positive and unlabeled samples, defined as the square of recall divided by detection prevalence (Bekker and Davis, 2020; McDonald et al., 2021). This metric, which is proportional to the F1 score, ensures balanced consideration of both type 1 (False Positive) and type 2 (False Negative) errors (Jain et al., 2017; McDonald et al., 2021).

The model incorporated a wide range of explanatory variables for settlement consolidation, selected based on empirical evidence and theoretical relevance (Fig. 1 and Supplementary Table 1). These biophysical, socioeconomic, and landscape-associated variables were integrated into the PU bagging model to analyze their empirical relationships to the locations of completed consolidation projects and capture the unique influence of each variable. By focusing on how these factors have historically shaped the spatial distribution of consolidation practice, the model effectively informs future site selection. To ensure robustness, we aligned the temporal scope of these variables to the year 2018, ensuring their relevance to recent consolidation activities.

Data source

Table 1 provides an overview of the datasets used in this study. Spatial details of completed consolidation projects were acquired from the local government (https://www.zj.gov.cn/). Land use data, derived from the China land cover dataset (Yang and Huang, 2021), were used to identify rural land systems in Zhejiang Province, China (Supplementary material, and Supplementary Fig. 1). Additionally, land use data from 2018 was used to calculate the cropland reclamation priority index (Wang et al., 2021), and to derive landscape metrics of cropland and settlements, including aggregation index, patch density and edge density-factors that serve as explanatory variables of settlement consolidation.

Table 1 Datasets used for our analysis. All relevant spatial datasets were resampled to 30-m raster images, which are in the Krasovsky coordinate system with a standard local Albers projection.
Full size table

Spatial layers, including building height (Wu et al., 2023), Euclidean distance to railways and roads, population density (WorldPop, 2018), Euclidean distance to IUCN strict natural reserves and wildness areas, were resampled into 30-m raster images and used as additional explanatory variables of settlement consolidation. Euclidean distance to IUCN's strict natural reserves and wildness areas was produced in 2017 while building height data was generated in 2020. Given the minimal time differences between these datasets and the study’s focal year of 2018, we believe their impact on the main conclusions is marginal. To delineate the possible impacts of population dynamics on settlement consolidation potential, WorldPop data were employed to calculate population change for the periods 2000–2010 and 2010–2018 (WorldPop, 2018).

Terrain data, soil properties data, and climate data were derived from ASTER DEMv3, SoilGrids (Poggio et al., 2021), and WorldClim (Fick and Hijmans, 2017), respectively. Considering the relative stability of soil properties over short periods, we applied this data to investigate the relationship between soil properties and settlement consolidation, as minimal temporal variations are unlikely to significantly affect the analysis. Annual mean temperature, annual precipitation, and solar radiation (1970–2000) were chosen to incorporate a long-term historical climate condition into the assessment of consolidation potential.

Results

Spatial distribution of settlement consolidation potential

Our assessment found 29,784 ha of settlements possessing consolidation potential in Zhejiang Province (Fig. 3a). A large proportion of this potential is concentrated in Jiaxing (12,961 ha), Huzhou (8987 ha), and Wenzhou (3340 ha), representing 44%, 30%, and 11% of the total, respectively. Furthermore, these areas are more spatially contiguous in Jiaxing compared with Wenzhou.

Fig. 3: Consolidation potential of settlements.
figure 3

a Represents the potential area for settlement consolidation; b displays the summarized potential areas for settlement consolidation at a 30-m resolution, aggregated to each land system grid (0.9 × 0.9 km²). Credit: World Ocean Base by ESRI.

Full size image

The average settlement area among rural land systems is 5.44 ha, while the average potential consolidation area is 0.67 ha (Fig. 3b). Regions like Lishui, Shaoxing, Zhoushan, and Ningbo show limited consolidation potential, with less than 5% of land systems exceeding 1 ha of potential consolidation area. In contrast, rural systems in Quzhou, Huzhou, and Jiaxing exhibit significant consolidation opportunities, with 2%, 14%, and 26% of land systems containing more than 5 ha of settlements with consolidation potential, respectively.

Rural land systems with smaller settlement sizes exhibit high consolidation potential (Fig. 4). Across Zhejiang Province, 71% of rural land systems exhibit limited consolidation potential, with less than 1% of the total consolidation potential. Nonetheless, over 18% of rural land systems still offer significant consolidation opportunities, with over half of their settlements suitable for consolidation. A clear trend is observed within the 0–25 ha settlement area range: as settlement size increases, the mean proportion of settlement with consolidation potential decreases. For instance, settlements between 0–5 ha exhibit a mean consolidation proportion of 24%, which drops sharply to 9% for those in the 15–20 ha range.

Fig. 4: Relative frequency distribution of settlement pixels as the joint occurrence of settlement sizes and the proportion of settlements with consolidation potential in rural land systems, shown for Zhejiang Province and its three geomorphic divisions: Plains, Hills, and Mountains.
figure 4

The delineation of geomorphic divisions is shown in Supplementary Fig. 2.

Full size image

Our results present a distinct pattern of settlement consolidation potential across plains, hills, and mountains (Fig. 4). Plains account for almost 80% of the potential area for settlement consolidation in rural areas, followed by mountains (17%) and hills (3%). For land systems with over 75% of settlements exhibiting consolidation potential, rural land systems in plains have the highest average settlement size, at 6.18 ha, which is substantially larger than the 0.78 ha in mountains and 0.54 ha in hills. Despite these contrasts, a common trend emerges across all geomorphic divisions: rural land systems with high consolidation potential—where over 90% of settlements are identified as potential consolidation areas—are predominantly associated with settlement sizes smaller than 2 ha.

The settlement consolidation potential is influenced by various biophysical, socioeconomic, and landscape-associated factors, each exhibiting different levels of influence (Fig. 5). Specifically, factors such as distance to conservation areas, annual mean temperature, and annual precipitation are more influential than others. In contrast, most landscape-associated factors, except for the aggregation index of settlement, have relatively limited influences. Solar radiation was excluded in the final prediction due to low feature importance.

Fig. 5: Mean feature importance of explanatory variables for settlement consolidation across bags.
figure 5

Feature importance results cannot offer any information about the directionality of each variable’s ability to accurately identify settlements with consolidation potential.

Full size image

Model performance and uncertainties

We assess the settlement consolidation potential at a 30 m resolution using ensemble models through 5-fold cross-validation (Fig. 6). Our results indicate that the PU bagging model with decision tree as the base classifier generally achieves higher recall scores compared to random forest models across most parameter settings. However, random forest models consistently exhibit higher detection prevalence scores. Considering the modified F1 scores, the choice of decision tree models as base classifiers outperform random forest models, demonstrating superior performance in identifying settlements possessing consolidation potential. Therefore, the decision tree was selected as a base classifier for the PU bagging model. Furthermore, the decision tree-based PU bagging models demonstrate relatively stable modified F1 scores across different down-sampling ratios, with a slight improvement as the number of bagging iterations increased. Accordingly, the optimized down-sampling ratio and bagging numbers are determined to be 1 and 100, respectively.

Fig. 6: Model performance in recall, detection prevalence, and modified F1 score of cross-folder validation.
figure 6

The panels are organized horizontally by the number of bags employed, with performance metrics displayed vertically. The x-axis represents increasing down-sampling ratios, and the y-axis indicates the mean value of each performance metric across 5-fold cross-validation, with error bars representing ±1 standard deviation. Each model variation is evaluated using the optimized threshold that maximizes the modified F1 score.

Full size image

To reduce uncertainties in the classification of settlement with consolidation potential, we optimized probability thresholds within the PU bagging model, focusing on maximizing the modified F1 score. As the model outputs probabilities indicating the likelihood of consolidation, selecting an appropriate threshold is critical to determining which settlement cells are classified as having consolidation potential. Through a comprehensive test of thresholds ranging from 0.01 to 1 during ensemble modeling, we identified an optimal threshold of 0.52 for the final prediction model. This threshold yielded a modified F1 score of 4.05, a recall of 0.75, and a detection prevalence of 0.14 across 5-fold cross-validation (Fig. 7a).

Fig. 7: Model uncertainty assessments.
figure 7

a Represents the number of settlement pixels that the final model classifies as positive (settlement with consolidation potential), broken apart by the initial training dataset label (positive or unlabeled), where the positive label means this settlement has been consolidated during 2018–2022. b Illustrates the coefficient of variation (CV) of settlement consolidation probability as a function of its values. The CV is calculated as the ratio of the standard deviation to the mean settlement consolidation probability for each cell across 50 runs. A wide range suggests significant variation among models, while a narrow range reflects greater agreement, implying a more reliable estimate. The shaded areas indicate interquartile ranges (25%–75%), and the lines represent the ensemble median values. The optimized cutoff threshold for determining consolidation potential is presented as a dotted vertical line.

Full size image

We estimate the coefficient of variation (CV) of settlement consolidation probability in repeated 50 model runs to quantify the uncertainties of the PU bagging model. The CV of consolidation probability exhibits an inverse-U pattern along with the increase of consolidation probability (Fig. 7b and Supplementary Fig. 5). The uncertainty in Zhejiang Province is specifically found most in low consolidation probability (around 0.3–0.5), whereas the CV would meet a rapid decline as consolidation probability goes beyond around 0.5 (Fig. 7b and Supplementary Fig. 5). Besides, the inverse-U dynamic pattern of CV is observed across three geomorphic divisions. Although the highest value of CV in plains is larger than in Hills and Mountains, modeling uncertainties in plains would meet a more significant reduction in consolidation probability around 0.4 to 0.7.

Discussion and conclusion

Settlement consolidation as a solution for food insecurity and beyond

Settlement consolidation offers a promising solution to offset cropland loss and boost food production, particularly in the context of global food security. Our results explicitly illustrate the spatial heterogeneity in settlement consolidation potential. Here, we identify 29,784 ha of potential areas for settlement consolidation in Zhejiang Province, predominantly in Jiaxing (12,961 ha) and Huzhou (8987 ha). From the perspective of geomorphic divisions, plains stand out, accounting for almost 80% potential area for settlement consolidation (Fig. 4), followed by mountains (17%) and hills (3%). Moreover, implementing consolidation in these areas can increase local food production by 0.91–0.97% (53,306–55,558 tons) compared to the 2022 levels (Supplementary Fig. 3). While this growth may seem modest, it holds strategic importance in offsetting ongoing declines in crop production and improving food security.

In addition to promoting food production, the consolidation potential map can explicitly support the integration of settlement consolidation into multi-objective planning (Strassburg et al., 2020; Bateman et al., 2024). Food security, biodiversity conservation, and climate change mitigation are all integral elements of the Sustainable Development Goals (Salerno et al., 2024). Therefore, stakeholders need decision-making support to achieve these multiple objectives and reconcile potential trade-offs between policy responses (Bai et al., 2018). For instance, cropland expansion or intensification is widely employed to meet challenges on the global food system from climate change, population dynamics and dietary change (Willett et al., 2019), while these measures are proven to potentially intensify global land competition between cropland and natural habitat, causing habitat loss and degradation (van Vliet, 2019; Williams et al., 2020). Previous studies indicate that settlement consolidation can contribute to biodiversity conservation and carbon emissions reduction by optimizing rural landscapes (Yue et al., 2023; Francini et al., 2024). Our study further demonstrates that settlement consolidation can serve as an alternative pathway to address significantly increasing food demand without compromising conservation objectives by mitigating the trade-offs between food security and ecosystem conservation.

Towards effective evaluation of settlement consolidation potential

This study introduces a scalable and robust tool for evaluating settlement consolidation potential at large spatial scales. Traditional methods, such as interviews or multi-criteria evaluation models, are often confined to smaller spatial scales or lack robustness (Zhou et al., 2017; Pašakarnis et al., 2021). Similarly, employing conventional machine learning methods would introduce significant biases in consolidation potential evaluation. To address these challenges, we leverage the PU bagging approach specifically designed for scenarios where only positive and unlabeled samples are available. This method employs positive samples and random subsamples from unlabeled data to analyze the empirical relationship between training samples and underlying consolidation drivers, enabling the development of a generic model capable of identifying potential consolidation areas within unlabeled settlement datasets. To further enhance the model’s accuracy and reliability, our modeling framework incorporates ensemble modeling, hyperparameter tuning, and an optimization algorithm for consolidation probability thresholds. These methodologies significantly improve prediction accuracy by maximizing recall and the modified F1 score while minimizing the coefficient of variation, thereby reducing modeling uncertainty. By providing precise spatial assessments, our innovative evaluation framework delivers valuable insights to policymakers and supports informed decision-making, paving the way for more location-specific and effective land-use strategies.

Stakeholder decisions are pivotal in land use changes (Malek and Verburg, 2020). In the context of settlement consolidation, key actors—such as local governments, villagers, and other stakeholders—make decisions based on the aforementioned explanatory variables of settlement consolidation and decision preferences (Liu et al., 2020; Gao et al., 2022b). Traditionally, understanding these preferences required extensive in-person consultations, which are time-consuming and costly (Tian et al., 2018; Zhao et al., 2024b). In contrast, machine learning algorithms offer a more effective and reliable approach by analyzing the relationship between implemented consolidation projects and explanatory variables (Mordelet and Vert, 2014; Stupp et al., 2021). By learning from the characteristics of areas where consolidation has already occurred, these models can identify new areas with similar features that align with these established patterns. Since the location of implemented projects reflects stakeholder decision preference, we reasonably assume that the areas identified by the machine learning model would also align with these stakeholders’ decision preferences. Additionally, the model’s high recall score validates this assumption, indicating its robustness in identifying consolidation sites shaped by stakeholder decisions (Fig. 6 and Fig. 7). These results emphasize the model’s effectiveness in reflecting decision-making preference, ensuring alignment between consolidation location selection with local preferences.

Policy implications

Financial investment is critically relevant to settlement consolidation. To assess the economic feasibility of such an effort, we further evaluated the required financial investment by using empirical cost data from previous consolidation projects in Zhejiang Province. On average, settlement consolidation in Zhejiang Province costs 0.48 million USD per ha, encompassing financial compensation for displaced homeowners, infrastructure development for agriculture (e.g., roads and water supply), project planning, and design expenses (Supplementary Fig. 4). Moreover, consolidating all potential areas in Zhejiang Province would require a total investment of approximately 18.06 billion USD. The investment is expected to be unevenly distributed across space (Supplementary Fig. 4), which further demonstrates the necessity of spatial assessment of consolidation potential to carefully and strategically implement settlement consolidation to maximize benefits while minimizing cost.

Currently, substantial subsidies for consolidation projects are provided by local authorities, primarily funded through local tax revenues. However, to enhance economic incentives and ensure long-term financial sustainability, the Zhejiang provincial government has introduced a market-driven mechanism. Under the policy outlined in “Zhejiang Province’s Opinions on Implementing Comprehensive Land Consolidation and Ecological Restoration Projects” (ZPG, 2018), land quotas generated from settlement consolidation—analogous to transferable development rights—can now be traded in the market (Chen et al., 2023). This innovative approach not only addresses funding gaps but also fosters broader participation in settlement consolidation.

Settlement consolidation goes beyond demolishing old houses and reclaiming cropland, it has profound implications for local well-beings, livelihood, and cultural heritage of local communities. Housing is a critical factor in economic and social stability for these families, significantly influencing their willingness to participate in consolidation practices (Liu et al., 2020; Gao et al., 2023). Thus, regions with high consolidation potential are supposed to implement comprehensive welfare programs to compensate residents who lose their residences and ensure the maintenance of desired livelihoods after consolidation. Other than the economic dimension, some settlements hold significant cultural and historical value. Many old buildings in rural areas are important historical assets that deserve deliberate protection to meet cultural preservation goals and sustain the historical identity of these communities. Interviews reveal that for some residents, settlements serve as emotional anchors, sites of ancestral worships, and centers of family connection (Fang et al., 2016). Consequently, settlement consolidation should prioritize minimizing land use conflicts and creating synergies between settlement development, local well-being, and cultural protection.