Advancing Land Use Modeling with Rice Cropping Intensity: A Geospatial Study on the Shrinking Paddy Fields in Indonesia

Abstract

Indonesia faces significant challenges in meeting food security targets due to rapid agricultural land loss, with approximately 1.22 million hectares of rice fields converted between 1990 and 2022. Therefore, this study developed a prediction model for the loss of rice fields by 2030, incorporating land productivity attributes, specifically rice cropping intensity/RCI, using geospatial technology—a novel method with a resolution of approximately 10 m for quantifying ecosystem service (ES) impacts. Land use/land cover data from Landsat images (2013, 2020, 2024) were classified using the Random Forest algorithm on Google Earth Engine. The prediction model was developed using a Multi-Layer Perceptron Neural Network and Markov Cellular Automata (MLP-NN Markov-CA) algorithms. Additionally, time series Sentinel-1A satellite imagery was processed using K-means and a hierarchical clustering analysis to map rice fields and their RCI. The validation process confirmed high model robustness, with an MLP-NN Markov-CA accuracy and Kappa coefficient of 83.90% and 0.91, respectively. The present study, which was conducted in Indramayu Regency (West Java), predicted that 1602.73 hectares of paddy fields would be lost within 2020–2030, specifically 980.54 hectares (61.18%) and 622.19 hectares (38.82%) with 2 RCI and 1 RCI, respectively. This land conversion directly threatens ES, resulting in a projected loss of 83,697.95 tons of rice production, which indicates a critical degradation of service provisioning. The findings provide actionable insights for land use planning to reduce agricultural land conversion while outlining the urgency of safeguarding ES values. The adopted method is applicable to regions with similar characteristics.

1. Introduction

Global food security issues are closely related to Sustainable Development Goals (SDGs) target 2 and remain a concern due to various factors such as climate change, increasing population, and economic growth. In 2024, Indonesia, a tropical developing country with a population and growth rate of approximately 280 million (the fourth largest in the world) and 1.11%, respectively [1], also experienced similar issues. In 2024, its economic growth reached 5.0%, which is greater than the world’s average economic growth of 3.2% [2]. These demographic and economic pressures increase the demand for land used for housing, industrial purposes, and other infrastructure. The expansion process has led to the massive conversion of agricultural land to built-up areas. Consequently, the country’s rice cultivation area has decreased from approximately 8.4 million hectares to 7.18 million hectares from 1990 to 2022 (1.22 million hectares), leading to an annual decline of approximately 38,000 hectares [3].

The reduction in agricultural land area due to industrial expansion and urbanization was reviewed through Alonso’s Bid-Rent Theory (1964) [4] and Harvey’s Class Monopoly Rent analysis (1974) [5], which explained the economic mechanisms governing spatial competition for land resources. These theoretical frameworks were particularly evident in peri-urban areas where agricultural land values competed with urban and industrial demands. Empirical studies across diverse geographical contexts [6,7,8] consistently showed strong correlations among urbanization, industrialization, and agricultural land conversion, providing robust theoretical support for understanding this globally significant phenomenon.

Based on the description above, rice is a staple food in Indonesia that is mainly cultivated on irrigated fields, partly on dry and tidal lands. In 2022, the total area of these fields in Indonesia was 7.46 million ha, with the largest distribution in Java (3.4 million ha), Sumatra (1.9 million ha), and the Sulawesi Islands (0.8 million ha) [9]. However, rice fields in Java have declined sharply due to high population density and economic growth. This region comprised approximately 156.93 million people, constituting over 55% of Indonesia’s population [10], and economic growth of 4.18% was recorded in 2024 [11]. These two factors, alongside others, caused the drastic decline of rice fields in Java, resulting in the conversion to built-up areas. Specific examples associated with the causes of declining agricultural land area in Java include (a) factory construction on former rice fields in the Pengkadan Sub-district, Cirebon (−6.806226°, 108.67718°); (b) rice fields encroached upon by settlements in the Babakan Sub-district, Cirebon (−6.87786°, 108.72523°); (c) toll road development for Patimban Seaport access in the Pamanukan Sub-district, Subang (−6.313074°, 107.82139°), specifically the development project of Patimban International Seaport in Subang Regency, West Java Province, which requires 334 hectares of mainland [12]; and (d) the West Java Petrochemical Complex industry expansion site in the Balongan Sub-district, Indramayu (−6.38996°, 108.38727°), which is owned by Pertamina in Indramayu Regency, West Java Province, and covers 331.92 hectares [13]. Most of the affected agricultural land consists of rice fields and gardens with relatively small average sizes of approximately 0.6 ha per farming household [12]. Figure 1 shows photographs with respective locations plotted on Google Earth, depicting the dynamics of the conversion of rice fields to non-agricultural uses in West Java Province.

Paddy fields play an important role in sustaining ecosystem services (ES), and aside from their main function as food production sources, they provide adequate biodiversity habitats and environmental regulation, which are regarded as essential components of Green Infrastructure (GI) [14]. Concerning their significance, paddy fields also face existential threats from consolidation issues and neglect. The assessment of ES values (ESV) is essential, as it shows the ecological and economic benefits that support sustainable decision-making and effective management strategies in response to climate change and urbanization challenges. Related studies conducted by Huang, et al. [15], in China, focused on the use of remote sensing data to explore and estimate ESV from Paddy Field Ecosystems (PFE). Meanwhile, Myeong and Yi [16] studied the ESV of paddy wetlands in Korea using contingent valuation methods to measure public willingness to contribute to their preservation.

The Indonesian government has issued a policy through Law No. 41/2009 on the Protection of Sustainable Food Agricultural Land (PSFAG) and the derivative regulations to address the conversion of agricultural land in Indonesia, specifically in Java. Despite these efforts, the conversion of agricultural land is still massive. In 2023, the government, through the Ministry of Agrarian and Spatial Planning/National Land Agency (Kementerian Agraria dan Tata Ruang/Badan Pertanahan Nasional, ATR/BPN), established Protected Rice Fields (PRF). The policy was legalized in respect to presidential regulation number 59 of 2019 on the Control of Conversion of Paddy Fields [17]. This was aimed at hastening the establishment of maps to protect and maintain the availability of rice fields, including controlling the rapid conversion, as well as supporting national food needs. The PRF data is an input for revising Indonesia’s regional spatial planning map, and its determination has consequences associated with both the government’s and community’s rights and obligations. However, the effectiveness of land conversion control in the future still needs to be proven.

In line with the description above, geospatial technology development has been widely applied in various fields [18,19], including land use planning. This development process applies artificial intelligence combined with geospatial data, science, and technology to hasten the understanding of the real world [20]. Geospatial technology is fast, precise, accurate, and increasingly cost-effective. An example of a geospatial study based on GEE cloud computing for agricultural land monitoring was conducted in Ethiopia by Eisfelder et al. [21]. Additionally, the GEE platform was used to develop high-resolution (10 m) cropland data layers (CDL) across the United States. This significantly improved classification precision and facilitated evidence-based decision-making among agricultural stakeholders, showing applicability for effective crop health monitoring through sophisticated image processing and machine learning methods [22].

Asgarian and Soffianian [23] and Ahmadi and Ghamary Asl [24] reported that geospatial technology has proven effective in land use prediction modeling. This necessitated historical land use data, an analysis of changes, model creation, and validation. In addition, positive validation results were used to project future land use. The most widely applied methods to simulate land use include combined Multi Linear Perceptron Neural Network (MLP-NN), Markov Chain, and Cellular Automata (Markov-CA) algorithms [25,26]. This combination was precisely used to model urban development locations by revealing the non-linear relationships between the expansion process and its drivers [27]. It also automatically generates a large number of parameter values with less training data [28].

Several agricultural land use modeling studies have integrated this combined model with other methods. Pradhan et al. [29] integrated MLP-NN Markov-CA with multi-criteria decision analysis in a study conducted in India. Their study forecasted that Boro rice cultivation was expected to dominate land cover by 2028 and 2038. Liang et al. [30] integrated MLP-NN Markov-CA with a land expansion analysis strategy based on multi-type random patch seeds. Additionally, this was referred to as a patch-generating land use simulation (PLUS) in the study conducted in China.

Rice cropping intensity (RCI), which is a fundamental parameter for evaluating agricultural sustainability, is widely explored using cutting-edge remote sensing tools. This measurement is used to describe how often agricultural land is planted with rice within one year of the growing season. Minh, et al. [31] researched land use patterns in An Giang Province, Vietnam, using Synthetic Aperture Radar (SAR) data from the Sentinel-1A satellite and reported that triple-rice cropping systems occupied 46.6% of the area. Meanwhile, double- and single-cropping systems covered 870.4 km² (24.7%) and 258.5 km² (7.3%), respectively, with the remaining 21.4% allocated to other land uses. SAR data provides distinct benefits for RCI evaluation due to its ability to capture images in all weather conditions, both day and night [32]. He et al. [33] combined time series data from Sentinel-1 (radar) and Sentinel-2 (optical) imagery to map rice distribution and RCI in South China and reported that single-cropping systems were dominant (88% of the area), while double cropping accounted for 12% of the area, achieving an overall mapping accuracy of 81%.

Aside from advances in land use modeling, there is a gap in integrating geospatial methods to develop simulation models. These are aimed at mapping rice fields with their RCI attributes and quantifying ES—particularly food production—using multi-source remote sensing data at a 10 m spatial resolution. The novelty of this study focuses on combining high-resolution (10 m) RCI-based rice field mapping with a customized land use simulation model that explicitly incorporates related dynamics and ES evaluation. The main objective is to formulate a comprehensive model that predicts rice field losses and evaluates the impact of urban expansion on regional food production. This study uses advanced geospatial technologies that enable an accessible, rapid, precise, and cost-effective analysis of agricultural land conversion for urban uses, focusing on Indramayu Regency in West Java Province as a representative case study for Java Island.

Building upon the reviewed literature, this study operationalized three testable hypotheses: (1) Urban expansion in Indramayu Regency has hastened rice field conversion to built-up areas, threatening regional food security. (2) RCI critically influences production loss magnitude, with double-cropping areas representing disproportionately high value loss. (3) High-resolution geospatial technology that integrates multi-temporal Sentinel-1 SAR data can accurately quantify paddy field loss and its food provisioning impact. The proposed hypotheses are based on observed trends and proven method capabilities, directly informing Indonesia’s sustainable agricultural policies (PSFAG Law No. 41/2009 and PRF policy). This is realized by providing actionable insights to reduce urbanization-driven food security risks.

2. Materials and Methods

2.1. Study Area

Indramayu Regency is located on the northern coast of West Java Province, as shown in Figure 2. The regency faces significant urbanization pressure despite its predominantly agrarian characteristics, particularly after being designated as part of the Rebana Area in 2021 [34]. According to this designation, the Rebana Area is located in the northern and eastern parts of West Java Province and is designed as a metropolitan region that connects three main strategic points. These include Cirebon as the city center, Patimban Port, and West Java International Airport Kertajati in Subang and Majalengka. In 2019, Indramayu Regency had a population of 1,855,458 people, with a relatively high density of 883.80 people/km², which reportedly increased at an annual rate of 1.31% between 2019 and 2024 [35]. Landsat 7 and Sentinel-2 satellite imagery analysis showed that residential areas in Indramayu increased from 16.2% to 17.0% between 2005 and 2021 at the expense of productive rice fields [36]. Additionally, urban areas have mainly been developed along major transportation corridors and coastal zones [37].

Indramayu Regency, a significant rice-producing region in West Java, occupies an area of 204,011 hectares, with 56.8% of this area used as irrigated or rain-fed paddy fields. The flat topography (0–3% slope) and tropical climate (1500–2500 mm annual rainfall) tended to support intensive rice cultivation [38]. In addition, land cover in the district included settlements (15.2%), ponds (12.4%), and mangrove forests along the coast (8.1%). Agricultural practices in the area are supported by horticultural and plantation crops. Considering this perspective, horticultural crops such as vegetables (chili, onion, and tomato) and fruits (melon and watermelon) are planted alternately with rice, depending on the season, water availability, and market demand [38].

The annual rice planting calendar is broadly divided into two seasons. Planting Season (PS) I occurs during the rainy season from October to March, while PS II is observed during the dry season, from April to September [39]. Rice plants were routinely cultivated in PS I from October to November. During this period, land preparation, tillage, and seedling planting were performed, with the harvest carried out from February to March. For PS II, rice was planted from April to May, while the harvest time fell from August to September. The elevation of the area ranges from 0 to 266 m above sea level, with an average temperature of 25–32 °C. The alluvial plain-dominated region made Indramayu a fertile area suitable for agricultural purposes, specifically rice cultivation.

2.2. Materials

The main data used were Sentinel-1 and Sentinel-2 satellite images sourced from the European Space Agency (ESA). This included Landsat images obtained from the United States Geological Survey (USGS). All main data were accessed through the Google Earth Engine (GEE) platform.

The Sentinel-1 satellite produces SAR imagery recorded by instruments operating in the C-band (at a frequency of 5.405 GHz). Despite the availability of multiple Sentinel-1 satellites, this study exclusively used data from the Sentinel-1A satellite based on its availability. The SAR data used consisted of time-series images covering a one-year rice cultivation season in the study area, from October 2020 to September 2021. Specifically, the Sentinel-1A satellite imagery had Ground Range Detected (GRD) characteristics, operated in Interferometric Wide (IW) mode with Vertical Horizontal (VH) polarization, and offered a spatial resolution of 10 m. Sentinel-1A data was also available for the same location every 12 days, and the GRD images available through GEE were subjected to pre-processing. This included thermal noise removal, radiometric calibration, and terrain correction using the Shuttle Radar Topography Mission (SRTM) 30 digital elevation model or the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) for regions with latitudes greater than 60 degrees (where SRTM is not available). The final terrain-corrected values were converted to decibels through logarithmic scaling (10 × log₁₀(x)) [40].

For Sentinel-2 imagery, only the green, red, and near-infrared (NIR) channels from the 2A and 2B satellites were used. These channels were adopted to generate annual Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) layers, with all channels possessing a spatial resolution of 10 m. Similar to the Sentinel-1A data, the images were acquired from October 2020 to September 2021. The two Sentinel-2 satellites operated simultaneously in the same orbit, resulting in a shorter revisit time of approximately five days in equatorial regions. Meanwhile, the Landsat optical imagery used included Landsat 7 and 8 from 2013, 2020, and 2024. In this context, only the visible (VIS), NIR, and shortwave infrared (SWIR) channels of Landsat images were used, all of which provide a spatial resolution of 30 m. Landsat data was obtained every 16 days for the same location and satellite. Both Sentinel-2 and Landsat images used were surface reflectance (SR) products.

Field survey data obtained between 2019 and 2023, as well as historical and freely available very high-spatial resolution imagery on the Google Earth Pro desktop tool, served as a reference for the land use/land cover classification of Landsat satellite images. Reference data on the growth phase of rice from Statistics Indonesia (Badan Pusat Statistik, BPS) were used for the analysis of Sentinel-1A imagery. These were generated through a nationwide rice growth phase survey conducted by BPS surveyors at the end of each month using the Area Sampling Frame (ASF) method. The dataset consisted of 1944 georeferenced points per month, particularly in Indramayu Regency. Additional supporting information included administrative boundary data and base map layers, sourced from the 1:25,000 scale Rupa Bumi Indonesia digital vector map (available at: https://tanahair.indonesia.go.id/ and accessed on 6 August 2024) published by the Geospatial Information Agency. (Badan Informasi Geospasial, BIG).

2.3. Methods

Figure 3 shows an overview of the study workflow, which integrates land use modeling and RCI mapping, alongside relevant data collection steps from modeling to ES valuation. This early presentation helps frame the subsequent detailed methodological explanations.

2.3.1. Land Use/Land Cover Prediction Map Creation

This subsection explains the steps adopted to generate a 2030 land use prediction map.

Landsat Image Classification in GEE and Accuracy Assessment

The classification of Landsat imagery to produce land use/land cover (LUC) maps was conducted in GEE. The images were three sets of SR products for 2013, 2020, and 2024. Each dataset consisted of visible, near infrared, and shortwave infrared channels, with a spatial resolution of 30 m. For each set of Landsat images in GEE, the function of masking clouds based on the Quality Assessment (QA) band was run, after which the median filter was applied. Both the mask clouds and median filter were used to minimize the influence of cloud cover on each image pixel.

Reference data in the form of geo-coordinated points evenly distributed randomly across the study area were prepared. Each point was labeled with a LUC category. Seventy percent (70%) of these points served as training samples for classifying Landsat images from 2013, 2020, and 2024, while the remaining 30% were used to validate the output. The 1000 reference data points were sourced from Sentinel-2 10 m Land Use/Land Cover (https://livingatlas.arcgis.com/landcover/ and accessed on 9 September 2024) and field surveys in 2019 and 2023. Before usage, the data was visually checked using historically available very high-resolution satellite imagery (2013, 2020, and 2024) on the Google Earth Pro desktop application.

Landsat imagery was classified to generate LUC maps for 2013, 2020, and 2024 using the Random Forest (RF) machine learning algorithm in GEE. The RF algorithm adopted an ensemble of decision trees, using the collective computational power to achieve superior prediction accuracy while reducing overfitting tendencies, including in the satellite imagery classification function [41]. It was selected due to its robustness in handling multispectral data and proven high accuracy in LUC classification [42,43]. In GEE, a majority filter was applied to each RF classification result to reduce noise. Additionally, post-classification refinement was implemented through visual interpretation to improve accuracy and generate the final LUC map. This refinement process was performed in QGIS desktop 3.34.11-Prizren software using the Thematic Raster Editor (ThRasE) plugin version 25.6 (https://plugins.qgis.org/plugins/ThRasE/ and accessed on 9 September 2024). The final LUC map consisted of five categories, namely, (1) rivers/lakes, (2) forests/plantations, (3) cropland, (4) built-up areas (settlements, industries, roads, and other artificial land uses/cover), and (5) ponds.

An accuracy assessment was conducted separately on the LUC maps of 2013, 2020, and 2024. This process concentrated on the remaining 30% of reference data that were not used in training samples. Furthermore, the confusion matrix method was adopted to generate overall accuracy and Kappa statistic values for each of the 2013, 2020, and 2024 LUC maps.

Predicting Land Use in 2030

The prediction of land use change in 2030 was conducted by adopting Multi-Layer Perceptron Neural Network (MLP-NN) and Markov Chain Cellular Automata (Markov-CA) algorithms, which were integrated with the Land Change Modeler (LCM) module in TerrSet 18.31 software. The MLP-NN is a supervised learning algorithm inspired by the decision-making processes of biological neural networks. The algorithm identified complex non-linear patterns in large-scale datasets, such as land use modeling, which was achieved through historical related maps and driver factors. These enabled the detection of spatiotemporal trends [25]. Moreover, the model was used to predict future land use changes, including the generation of a transition potential map, which is responsible for quantifying the likelihood of change for each spatial unit (e.g., pixel or grid cell) by training historical data [44].

The Markov-CA model combined the Markov Chain (for transition probabilities) and Cellular Automata (CA) (for spatial allocation). A Markov Chain is a mathematical model used to describe a system that transitions between various states (in this case, the states are land use categories such as cropland, built-up areas, and water bodies) based on fixed probabilities [45]. CA refers to a discrete model that includes a grid of cells that evolves over distinct time steps based on predefined rules. This dictates how each cell’s state changes according to the respective state of the neighbors [46]. The Markov Chain component calculates the magnitude of land use change by deriving transition probabilities from historical land use transitions.

The transition probabilities were organized into a transition matrix, which estimated the total number of pixels expected to shift between land use classes over a specified timeframe. The CA component then spatially allocated the predicted changes. Therefore, MLP-NN Markov-CA incorporated two critical inputs: (1) the transition potential maps from MLP-NN, which are responsible for identifying where changes occurred, and (2) the spatial neighborhood rules (e.g., the influence of adjacent pixels). By synthesizing these inputs, the CA ensured that changes were distributed realistically across the landscape.

The land use change modeling process adopted historical land use data from 2013 and 2020 as baseline information to develop the prediction model. The modeling considered five driving factors, namely, proximity to settlements, water bodies (lakes and rivers), road networks, elevation, and slope, as shown in Figure 4. The initial modeling stage started with identifying significant land use transition patterns between 2013 and 2020. Based on land use change analysis, the main sub-model was the main focus of the modeling, including the transition from cropland to built-up areas. The sub-model was selected because it represented the dominant pattern of built-up area expansion in the study area.

The results of MLP-NN modeling in LCM were presented in a report, including the accuracy value of the transition submodel and the influence of each driving factor on the model. Following Eastman [47], the present study established a minimum accuracy threshold of 80%. If this threshold was not achieved, the modeling process was repeated with different input parameter settings until the target was reached. However, after attaining satisfactory accuracy, the suitability map, which featured values ranging from zero to one, where higher values showed a greater likelihood of land conversion, was generated. Based on the suitability map and a transition probability matrix for 2020–2024, as shown in

Table 1. Markov Chain transition probability matrix for 2020–2024 (calculated from 2013–2020 change trends).

	Probability Change to:
		Rivers/Lakes	Forest/Plantations	Cropland	Built-Up	Ponds
	Rivers/Lakes	0.7256	0.0093	0.2557	0.0094	0.0000
	Forest/Plantations	0.0008	0.6105	0.3624	0.0242	0.0021
Given:	Cropland	0.0023	0.0055	0.9757	0.0122	0.0044
	Built-up	0.0009	0.0033	0.0000	0.9909	0.0049
	Ponds	0.0000	0.0046	0.0181	0.0033	0.974

, a simulated map of LUC 2024 was generated.

The suitability map was exclusively derived from 2013–2020 land change dynamics, representing spatial transition potentials driven by static geographic factors (e.g., proximity to settlements, roads, water bodies, elevation, and slope). These factors exhibit minimal short-term variability, enabling the map to guide change allocation beyond the calibration period. The Markov Chain adjusts change quantities based on annualized transition probabilities for projections to 2024 (4 years) and 2030 (10 years), while CA preserves the suitability-based spatial rules. This separation of quantity (temporal) and location (spatial) dynamics enables scalability across time horizons.

Model validation was then conducted by comparing the 2024 land use simulation results with the actual data using the Kappa statistic (κ). This calculation is frequently adopted to test the reliability or similarity between variables [48], specifically the concordance of land use categories in pixels between the predicted map and the corresponding pixels on the actual map. A Kappa value of 0.75 or higher is considered excellent [48]. Meanwhile, achieving a Kappa value greater than this threshold shows that the model has reliable predictive capabilities for use in long-term simulations. The final stage entails the generation of a land use prediction map for 2030, following validation of the model.

2.3.2. RCI Creation

The RCI distribution map was obtained from the analysis of the time series of Sentinel-1A and Sentinel-2A/B satellite imagery. The Sentinel-1A time series data consisted of 29 images collected during a one-year rice growing season (from October 2020 to September 2021). To complement this dataset, annual NDVI and NDWI layers were incorporated. The Sentinel-2 data shared the same acquisition period as Sentinel-1A. Before integration with Sentinel-1 data, normalization was performed on both NDVI and NDWI data. This normalization was based on the minimum and maximum values of the Sentinel-1A data. Subsequently, the K-means clustering and hierarchical clustering analysis (HCA) methods were sequentially applied to the dataset. K-means, which is an unsupervised classification method, operates by selecting K random centroids from the dataset. This entails the grouping of data based on proximity to centroids, updating the positions using the average location within clusters, as well as repeating the process until convergence is achieved [49]. K-means was selected due to its easy implementation for land cover classification, relatively fast processing time, and high accuracy (91.42%) [50]. Meanwhile, HCA is a clustering method that builds a hierarchical data structure by progressively merging or separating clusters in the form of a dendrogram based on similarity measures between objects [51]. The method was selected because of its ability to cluster complex spatial data without determining the number of clusters beforehand. Its flexibility in handling various data types and its capability to show clustering results in the form of a dendrogram facilitate the interpretation and analysis of inter-cluster relationship patterns [52].

In this context, the RCI mapping process started with the implementation of K-means classification in GEE on the combined Sentinel-1A time series data with NDVI and NDWI. This process generated a single image containing 60 unlabeled clusters. Subsequently, the K-means results were integrated with Sentinel-1A time-series data, NDVI, and NDWI. From this four-data combination, 10,000 random sample points representing each cluster were extracted. Each sample point contained information on the cluster ID, backscatter values from each Sentinel-1A image layer, NDVI, and NDWI. The results of the extraction process were saved in comma-separated values (CSV) format.

Using the HCA method with Euclidean distance and Ward D2 component available in the factoextra version 1.0.7 package in R software version 4.5.1, the 60 unlabeled clusters in the CSV file were consolidated into ten final clusters based on respective similarity levels. Ward D2 is an agglomerative clustering method used to minimize the total variance within clusters. Meanwhile, Euclidean distance measures the shortest path between two points in multidimensional space. These clusters were then labeled with respective land use/land cover based on reference data from field surveys, very high-resolution Google Earth Pro imagery, and the ASF survey. The labeling process identified five clusters as rice fields with different cropping patterns (rice cropping zone, RZ), and the other five clusters were classified as non-rice areas (including built-up areas, trees, water bodies, and other non-rice lands). The HCA results and labeling scheme are presented in the dendrogram in Figure 5.

To improve the accuracy of classification results, the RF machine learning algorithm in GEE was applied to a combination of image data from HCA, Sentinel-1A time series, NDVI, and NDWI. The RF classification was performed using 10,000 randomly distributed training points extracted from the data. A majority filter algorithm was applied to the RF output image to reduce noise. In order to ensure that only rice fields were analyzed, five clusters of units from the RF results were overlaid with the Rice Field Baseline Map, and non-rice clusters were excluded from subsequent processing. The image containing five rice unit clusters was then merged with Sentinel-1A time series data without including the NDVI and NDWI layers. In this study, both layers were not required to identify rice growth phases.

Following the description above, 10,000 randomly distributed points representing each rice unit cluster were extracted from the data. Each extracted point contained information on the rice unit cluster ID label and Sentinel-1A time series backscatter values. Based on the perspective, the extracted data was stored in CSV format. In R, data from each cluster was divided into four sub-clusters (four quartiles). The dataset for each quartile in the respective cluster was aggregated using the median function to produce a single representative data point for each sub-cluster. These data were plotted in the form of rice growth phase graphs and then analyzed to determine the RCI.

Sentinel-1 SAR backscatter values have a strong correlation with rice growth phases [53]. During land preparation, backscatter values in the initial planting phase reached their lowest point. This was because flooded rice fields produced specular reflection that directed most radar energy away from the sensor, causing fields to appear as dark areas in SAR imagery [54]. In the early vegetative phase of the rice plants, backscatter values started to increase as the seedlings sprouted above the water surface, producing volume scattering mechanisms and double-bounce reflection between plants and the water surface [55]. However, during the late vegetative phase, approximately 30–45 days after planting, backscatter values reached the first peak when the plants had been fully developed and water started to diminish in fields, reducing the double-bounce reflection effect [56].

During the reproductive phase, backscatter values continued to increase until they reached the highest value, as the formation of panicles and rice grains created complex structures responsible for generating strong volume scattering [57]. In the ripening phase, there was a slight decrease in the SAR backscatter signal as the conditions became drier before harvest, reducing the water content in the tissues [58]. After harvest, the values varied depending on whether the fields were left to fallow, cultivated with commercial crops, or prepared for the next rice planting period [59].

Assuming that each growth stage proceeded normally for one month, the phases followed the sequence of the land preparation and planting (T) and early (V-I) and late vegetative (V-II) phases, alongside the reproductive (R) and mature (M) phases. The land preparation and planting phases were identified after the dB value had reached the lowest point, followed by V-I, V-II, R, and M stages in the subsequent months [59]. This was followed by the fallow (F) or commercial crop planting (C) phases before the next rice planting period. An ideal representation of the relationship between rice growth phases and SAR backscatter response was shown in Figure 6. Additionally, further identification was conducted for the five rice planting unit clusters by carefully examining each backscatter profile representing these clusters and comparing it with reference data on rice growth phases from the ASF survey data. By identifying the number of lowest points in the backscatter profile of an area (rice planting unit), it is possible to determine how many times the land unit was planted with rice in one growing season. Assuming that each planting ended with a harvest, the number of rice plantings represented the number of harvests. Therefore, each rice planting unit cluster was attributed to the number of harvests in one growing year.

2.3.3. Determining the Potential Rice Loss by 2030

In order to determine the potential paddy fields and corresponding production loss in 2030 compared to 2020, three main stages were implemented. The first stage included overlaying the predicted land use map for 2030 (using only the built-up category and ignoring the other categories) with the actual map of 2020 (using only the ‘Agricultural Land’ category and ignoring the other categories). In this context, the overlay generated a prediction map of agricultural land loss by 2030.

The second stage consisted of overlaying the results from the first stage with the RCI map. The result of this overlay was a potential distribution map of paddy fields that would be lost by 2030, containing RCI attributes.

The third stage entailed multiplying the land area by the average rice productivity rate per harvest. This enabled the estimation of the ESV, particularly the direct use value regarded as a provider of food resources for the community, predicted to be lost by 2030. As paddy fields decreased gradually annually (assuming a constant rate of decrease), the total potential rice production loss in 2030 was calculated cumulatively from the land lost each year. The formula for annual rice loss is stated in Formula (1), while the total rice production loss from 2020 to 2030, or over 10 years, was determined using Formula (2). The variable D represented the proportion of double-cropping land, with the proportion of single-cropping land = 1 − D. L represents the total land lost over 10 years (2020–2030) in hectares. W shows rice productivity for each harvest in tons per hectare. Additionally, the average rice yield in Indramayu Regency was 7.2 tons per hectare [60].

$$P_{annual}=\left(D\times {\displaystyle \frac{L}{10}} \times 2W\right)+\left(\left(1-D\right)\times {\displaystyle \frac{L}{10}} \times W\right)$$

(1)

$$P_{total}={\sum }_{t=1}^{10}{P}_{annual}\times (10-t)$$

(2)

Since all components except $(10-t)$ were constant, and the sum of the series 0 + 1 + 2 + … + 9 = 45, Formula (2) was simplified to Formula (3):

$$P_{total}=4.5\times L\times W\times (D+1)$$

(3)

3. Results

3.1. Classification and Accuracy Assessment of LUC Maps

Figure 7 shows the LUC maps generated from Landsat image classification using the RF algorithm in GEE for 2013, 2020, and 2024. A comparative bar graph is also provided to present the percentage distribution of each land use category—water, forest/plantations, crops, built-up, and ponds—across those years, highlighting temporal trends and the relative magnitude of land use change. The overall accuracy and Kappa value for each LUC map are shown in

Table 2. Accuracy assessment results of LUC maps for 2013, 2020, and 2024.

	LUC 2013	LUC 2020	LUC 2024
Overall accuracy	91.67%	95.67%	90.65%
Kappa statistic	0.75	0.87	0.72

.

3.2. MLP-NN Evaluation Results and 2030 Land Use Prediction Model

The results of the land change model evaluation obtained with the MLP-NN Markov-CA and LCM TerrSet methods, alongside the five driving variables shown in Figure 5, exhibited an accuracy rate of 83.90%. These results also outlined the influence of each variable on the model. The smallest order (first) showed the variable that exerted the most influence on the model, while the largest order (fifth) represented the least influence. The results showed that proximity to settlements was the most influential variable compared to the others, and slope was the least influential factor in the model. The following factors, including proximity to rivers and lakes, main roads, and elevation, ranked second, third, and fourth, respectively, in terms of influencing the model.

The use of the 2024 potential transition map in conjunction with the Markov Chain probability matrix aided in generating the 2024 LUC prediction map. Validation against actual 2024 data yielded a Kappa statistic of 0.91, confirming the model’s capacity to project near-term changes using suitability rules derived from 2013–2020 observations. This agreement showed that static geographic drivers remained effective for spatial allocation, even beyond the calibration window, supporting the model’s application for 2030 projections.

Table 3. Area of each land use category in 2013, 2020, and 2030, as well as the difference between 2013–2020 and 2020–2030.

No.	Category	2013		2020		2030		Δ2013–2020		Δ2020–2030
		(Hectare)	(%)	(Hectare)	(%)	(Hectare)	(%)	(Hectare)	(%)	(Hectare)	(%)
1.	Rivers/Lakes	1921.86	0.92	1720.62	0.82	1720.62	0.82	−201.24	−0.10	0	0.00
2.	Forest/Plantations	6922.26	3.32	4665.33	2.24	4665.33	2.24	−2256.93	−1.08	0	0.00
3.	Cropland	159,536.34	76.48	158,012.19	75.75	153,271.8	73.48	−1524.15	−0.73	−4740.39	−2.27
4.	Built-up	19,933.56	9.56	23,412.96	11.22	28,334.16	13.58	3479.4	1.67	4921.2	2.36
5.	Ponds	20,280.6	9.72	20,783.52	9.96	20,602.71	9.88	502.92	0.24	−180.81	−0.09
	Total	208,594.6	100.0	208,594.6	100.0	208,594.6	100.0

shows land use change dynamics within three periods, namely, 2013, 2020, and 2030. The study area totaled 208,594.6 ha, with cropland dominating land use within all periods. In 2013, cropland covered 159,536.34 ha (76.48%) of the total area, while built-up regions constituted only 19,933.56 ha (9.56%).

The analysis of land use change for the 2013–2020 period showed a significant decrease in agricultural land area of 1524.15 ha (0.73%), which was inversely proportional to the increase in built-up region of 3479.4 ha (1.67%). This expansion of the built-up area showed an intensive process of land conversion, specifically from agricultural land to built-up areas. In addition, this period witnessed a considerable decrease in forest/plantation area of 2256.93 ha (1.08%), with pond area experiencing a slight increase of 502.92 ha (0.24%).

Projections of land use change for the period of 2020–2030 showed an intensifying trend of agricultural land conversion. It was predicted that there would be a significant reduction in agricultural land compared to the previous period, reaching 4740.39 ha (2.27%). However, the built-up area was predicted to experience a significant expansion of 4921.2 ha (2.36%), increasing its proportion from 11.22% to 13.58% of the total area. This finding showed that the process of urbanization and development would continue at a higher rate, with most of the built-up area expansion occurring through the conversion of productive agricultural land. An illustration of cropland conversion to built-up areas during the period of 2013–2020 and its projected continuation from 2020 to 2030 is presented in Figure 8.

3.3. Results of RCI Mapping Analysis with Sentinel-1A Imagery

Figure 9 shows five rice cropping zones in Indramayu Regency during the planting year from October 2020 to September 2021. Rice cropping zones (CZ)-5 and CZ-1 dominated with 27.26% and 23.75% of the total fields, covering 27,767.04 ha and 24,191.51 ha, respectively. The other CZs comprised the following areas and proportions: CZ-2, 11,627.47 ha (11.41%); CZ-3, 15,296.43 ha (15.01%); and CZ-4, 22,993.64 ha (22.57%). Therefore, the total rice fields mapped in the Indramayu Regency comprised 101,876.09 ha.

The spatial distribution of CZs in this region exhibited varied patterns. CZ-1 (light blue/cyan color) was concentrated in the central and eastern parts of the region, with significant presence around Arahan, Lohbener, Lelea, and Bangodua, as well as in the Jatibarang, Sliyeg, and coastal Indramayu areas, with a relatively dispersed pattern. CZ-2 (dark blue color) was dominant in the northern region, specifically Cantigi, Sindang, and Pasekan, forming significant clusters in Bongas and Gabuswetan. It was found in small groups in Lelea, alongside a more clustered distribution compared to CZ-1 in the southern part.

CZ-3 (yellow color) was highly dominant in the northwestern part, particularly Sukra, Patrol, and Anjatan, forming a large, homogeneous block in Kandanghaur. It was also found in the eastern edge of the Krankeng region, with a distribution that tended to be grouped in large and compact blocks. Meanwhile, the CZ-4 (magenta/pink color) had a widespread distribution throughout the map area, with high concentrations around Sukra, Anjatan, Losarang, and Tukdana, forming significant patterns in the Sukagumiwang, Kertasemaya, and Balongan regions, with a mixed distribution pattern between clustered and dispersed.

In this context, CZ-5 (green color) was highly dominant in the southwestern part, specifically in the Gantar, Haurgeulis, and Kroya regions. It formed a large and continuous block throughout the southwestern region and was found in small groups characterized by the most homogeneous and consolidated distribution compared to other zones in the central and eastern parts. Generally, the distribution pattern showed that the northwestern and southwestern parts were dominated by CZ-3 and CZ-5. The northern part had a concentration of CZ-2, the central and eastern parts had a mixture of CZ-1 and CZ-4, alongside small portions of CZ-2 and CZ-5, with CZ-4 as the zone that was most evenly distributed throughout the region.

The temporal patterns of S–1A VH backscatter across five CZ in Indramayu Regency showed distinct variations in rice planting schedules and cropping intensity, as in Figure 10. CZ-1 exhibited the lowest backscatter point of approximately −24 dB in early January 2022, signifying the flooding period for field preparation and initial planting. Based on the 110- to 120-day growth cycle, the first harvest occurred from late April to early May 2022. The second decrease in late June showed preparation for the second planting process, which yielded a second harvest around October–November. This pattern confirmed a double-cropping rice cultivation system with minimal fallow periods between the two growing seasons.

CZ-2 showed the lowest backscatter point of approximately −25 dB from late December to early January, with the rice growth cycle leading to a first harvest in late April. A second sharp decrease in late June showed preparation for the second planting, which was harvested in October. Therefore, CZ-2 also showed a double cropping pattern alongside planting periods concurrent with CZ-1, which occurred due to variations in water management practices.

Building upon the description above, CZ-3 had the lowest backscatter point in February 2022, implying a later planting time compared to CZ-1 and CZ-2. With respect to normal rice growth duration, the first harvest was estimated to occur within June–July 2022, precisely when the graph showed a second decrease in backscatter values. This implied that CZ-3 also followed a double cropping system, but with a schedule shifted to approximately one month compared to CZ-1 and CZ-2. The subsequent harvest was estimated to occur beyond the presented data timeframe from November to December.

CZ-4 exhibited the lowest point from late November to early December 2021, leading to the initial harvest around March–April 2022. The second low point in May 2022 marked preparation for the second planting, which yielded the subsequent harvest around September. This pattern showed that CZ-4 also implemented a double-cropping system, but with an earlier start to the growing season compared to other zones.

Aside from the description above, CZ-5 exhibited the lowest backscatter point in mid-January 2022 with a smoother pattern, leading to harvest around May–June 2022. The absence of a significant second sharp decrease within the presented data period shows variations in agricultural practices in the next growing season, such as secondary crop cultivation after rice or longer fallow periods. As a result of these characteristics, CZ-5 was categorized as single rice cropping.

3.4. Estimation Results of Rice Field Area and Rice Production Loss in 2030

Figure 11 shows the distribution of paddy fields, supplemented with information on rice planting intensity (RCI) within a year. This was predicted to be lost because of the development of built-up areas in 2030, compared to 2020. Approximately 1602.73 hectares of paddy fields, comprising 980.54 hectares (61.18%) and 622.19 hectares (38.82%), with double- (2 RCI) and single-cropping (1 RCI), respectively, were projected to be lost in Indramayu by 2030. The average rice productivity for both double- and single-cropping fields was assumed to be the same, at 7.2 tons per hectare per harvest. Therefore, the total rice production predicted to be lost during the 2020–2030 period, based on calculations using Formula (3), amounted to 83,697.95 tons.

4. Discussion

The findings, which projected the loss of 1602.73 hectares and 83,697.95 tons of paddy fields and rice, respectively, by 2030, supported Alonso’s Bid-Rent Theory (1964) and Harvey’s Class Monopoly Rent analysis (1974). Both theoretical frameworks explained the economic–spatial mechanisms through which urban and industrial activities with higher land purchasing power dominated strategic locations, outcompeting agricultural economic values. This phenomenon reflected an unbalanced spatial competition between agricultural and more economically profitable non-agricultural sectors. Empirical evidence from Southeast Asia reinforced the validity of these findings [61]. For example, in Ho Chi Minh City, Vietnam, agricultural land occupied the largest area but decreased by approximately 10% from 2000 to 2020 [62], while in Central Luzon, Philippines, it reduced by 5% between 1970 and 2019 [63].

We integrated multi-source geospatial data with advanced modeling methods, including MLP-NN Markov-CA and Sentinel-1A SAR time series analysis. This combination provided a robust framework for predicting land use changes. This was evidenced by the MLP-NN Markov-CA model’s accuracy and Kappa value of 83.90% and 0.91, respectively. These findings are in line with previous studies, including the analysis conducted by Tahir et al. [64] in the Lahore District (Pakistan), with 93.6% accuracy and a Kappa value of 0.92, and Hussain et al. [65] in Punjab (Pakistan), with 89.08% accuracy and a Kappa value of 0.81.

The five rice zones (RZ) in Indramayu Regency generally implemented a double-cropping system (2 RCI). However, the zones showed significant differences in terms of planting start times and intervals between two rice cultivation cycles. This reflected adaptation to local conditions, such as water availability and agricultural practices prevalent in each zone, known as Pranata Mangsa (PM) [66,67]. In this context, PM is a traditional knowledge system adopted by farmers in Indramayu, which is related to planting times based on weather patterns and natural phenomena. The system was passed down to subsequent generations and is associated with the occurrence of natural events. The PM further served as a guide for determining the appropriate planting and harvesting times, with the expectation of optimizing agricultural yields. Although PM is traditional, it is consistently practiced, with some modifications and adaptations incorporating scientific data, enabling local communities to have better resilience in adapting to climate change and extreme hydrological events caused by global warming [68].

Aside from the natural river networks, the two large reservoirs, Jatiluhur (in Purwakarta Regency) and Jatigede (in Majalengka Regency), influenced rice planting schedules, both in Indramayu and the surrounding areas. These reservoirs irrigated most of the paddy fields in the study area through irrigation channels, with both ensuring water availability for rice cultivation throughout the year. Irrigation channels from the reservoirs located south of Indramayu were classified as primary, secondary, tertiary, and quaternary [69,70]. This classification was based on the hierarchy of channels in the irrigation system, starting from the largest (primary) channel, which takes water directly from main sources, such as dams or rivers, to the smallest (quaternary) channel, which is responsible for irrigating farmers’ fields. Paddy fields close to the primary irrigation channels were irrigated first, followed gradually by those near the quaternary channels. Therefore, logically, paddy fields in the southern (upstream) areas with respective irrigation systems were initially planted with rice first, followed by those in the middle, and then the northern (downstream) regions. Paddy fields that were not reached by irrigation channels depended on rain (rain-fed fields), and these were only planted with rice annually (single cropping/1 RCI) [71]. After the harvest, there was a decrease in rainfall during the second planting season, where fields were cultivated with commercial crops (cash crops such as corn, chili, cassava, peanuts, melons, and watermelons) [72] or left to fallow [73].

The dominance of paddy fields with double-cropping (2 RCI) in harvest loss predictions outlined the vulnerability of highly productive agricultural land to urbanization. The areas, which contributed significantly to regional rice production, were often located near settlements and infrastructure, making them the primary targets for conversion [74]. While this study assumes stable climatic conditions, future global climate projections (e.g., increased drought frequency or altered monsoon patterns) could significantly affect water availability and cropping intensity. Integrating climate data would strengthen long-term predictions and is recommended for future work. According to FAO, the results obtained were in line with global FAO trends [75], where agricultural land is increasingly used for urban development, thereby intensifying food insecurity in rapidly developing regions. This study also proved that proximity to settlements was the most influential driving factor for land conversion. Compared with the present study, Purswani et al. [76] and Kamwi et al. [77] reported similar findings, namely, that an important variable affecting land conversion was the proximity to settlements. Therefore, strategic spatial planning should focus on protecting the environment and productive agricultural zones from uncontrolled urban encroachment. In Indonesia, it was considered part of the Strategic Environmental Assessment (Kajian Lingkungan Hidup Strategis, KLHS) mentioned in Law Number 32 of 2009 concerning Environmental Protection and Management. Several technical implementations were then regulated in Government Regulation Number 46 of 2016, which concerns Procedures for Implementing KLHS and provides detailed guidance on the enactment of KLHS in Indonesia. During its application, several obstacles were encountered, such as the quality of information or the limited availability of accurate and up-to-date environmental data to support KLHS analysis [78].

The advanced methods adopted in the present study, particularly the 10 m resolution Sentinel-1A data used for RCI mapping, addressed an important gap in KLHS. This model provided actionable insights for policymakers by capturing small-scale variations in planting intensity. For example, zones with high RCI values were prioritized for protection under Indonesia’s PSFAG policy or targeted for intensification programs to offset losses elsewhere. However, reliance on historical trends for future projections may underestimate sudden policy shifts or disruptions caused by natural disasters (such as floods and landslides), showing the need for an adaptive modeling framework.

The socio-economic consequences of paddy field loss are equally concerning. However, the agrarian economy in Indramayu Regency is highly dependent on rice cultivation, and the displacement of farmers can worsen rural poverty, resulting in migration to urban centers. This finding is consistent with the results of previous studies, which demonstrate that agricultural land conversion led to the loss of farmers’ livelihoods and increased unemployment. The farmers were mainly aged, with a poor educational backgrounds, making it difficult for them to change professions [79].

Aside from the benefits outlined, this study had several limitations. The model captured spatial patterns through time-invariant drivers that failed to explicitly integrate dynamic factors such as climate change impacts and abrupt policy shifts. A key limitation is the exclusion of climate change projections, which may alter paddy field productivity and conversion pressures through shifts in water availability or extreme weather frequency. Subsequent research should incorporate climate scenarios (e.g., Coupled Model Intercomparison Project Phase 6/CMIP6 models [80]) to evaluate these dynamics. The model’s assumption of a constant rate of change may further underestimate non-linear responses to such climatic stressors. These were indirectly reflected through calibration with recent trends (2013–2020), limiting long-term robustness. The use of Sentinel-1A time series data covered only one year of the growing season, while a longer time series of data would have been ideal. This approach was expected to map paddy fields more accurately, considering that (after overlaying the baseline paddy fields map with the RZ map) a small portion could not be detected with the acquired data and adopted methods. Additionally, during the planting season, paddy fields were not cultivated with rice, partly due to flooding [81], causing the time series Sentinel-1A SAR backscatter values representing the characteristic phases of its growth to go undetected.

The calculation of potential ESV applied to lost paddy ecosystems currently covers only part of the provisioning services category. Ideally, this calculation should also account for other categories, such as regulating, cultural, and supporting services. Therefore, future studies were recommended to use longer Sentinel-1A time series data to comprehensively calculate the ES for paddy fields across all four categories. This included the incorporation of dynamic variables (e.g., climate indices and population growth layers) within scenario-based frameworks to enhance adaptability.

The results of this study were expected to provide an overview of the importance of maintaining paddy fields. Furthermore, the methods used had the potential to be applied in other locations in Indonesia and places with similar characteristics to the study area. By maintaining Protected Paddy Fields, Indonesia strived to ensure food availability, reduce rice imports, and support farmers’ welfare. This certainly contributed to achieving the national Food Self-Sufficiency target and the achievement of SDGs by 2027 [82] and 2030, respectively, particularly objective number 2, Zero Hunger.

5. Conclusions

In conclusion, this study successfully used the latest geospatial technology to predict the loss of paddy fields in Indramayu Regency by 2030. This was realized by combining high-resolution land use modeling and RCI mapping. The results identified a major threat to agricultural land, with an estimated 1602.73 hectares of paddy fields projected to be converted to urban areas, thereby causing rice production losses of 83,697.95 tons between 2020 and 2030 (10 years). Most of these losses (61.18%) are projected to occur in highly productive double-cropping systems (2 RCI). This highlights the significance of protecting all relevant aspects to maintain the ES value, including ensuring food security at the local, regional, and global levels.

These findings highlighted the critical role of advanced geospatial tools (including the use of big geospatial data from multiple sources, machine learning algorithms, and cloud computing) in supporting land use policies, specifically in rapidly urbanizing regions. Meanwhile, this study offered high-resolution insights, with key limitations including (1) temporal scope, as Sentinel-1A data covered only one growing season (2020–2021), limiting the analysis of interannual climate variability impacts on RCI, and (2) ESV, as only provisioning services (food production) were quantified, excluding regulating and cultural services.

Apart from these constraints, the method’s key strengths—open-access satellite data (Sentinel-1/2, Landsat), reproducible cloud computing (Google Earth Engine), and scalable machine learning workflows (RF, MLP-NN Markov-CA)—enabled direct application to other agricultural regions facing urban pressures where urbanization has threatened food security. To reduce the impacts, proactive measures, namely, the implementation of stricter zoning regulations, agricultural intensification, and integrated spatial planning, were needed. The method developed in this study could be applied in other regions with similar challenges, thereby providing a scalable solution to balance urban growth and agricultural sustainability.

Future studies are recommended to use Sentinel-1A SAR image data with a longer time span and conduct a more comprehensive calculation of the ESV of paddy fields. This recommendation is aimed at exploring the dynamic interactions among policy interventions, climate change, and land use patterns in order to improve the prediction model and support long-term food security strategies.