Sustainable Land Use and Land Cover Management Model for Flood Mitigation in Krueng Baro Watershed, Aceh, Indonesia

Changes in land use and land cover (LULC) are among the key drivers of global environmental shifts, often leading to negative impacts due to population growth and socio-economic expansion [1]. Understanding LULC dynamics and their consequences is essential for planning future interventions in affected areas [2]. These transformations directly influence community well-being by altering environmental conditions, such as increasing land degradation risk and flooding [3].

Water resources management and hydrological risks are of significant concern in our society. A contributing part of hydrological hazards worldwide, floods, erosion and sedimentation are among the most frequent and environmentally damaging natural disasters. Globally, between 1994 and 2013, floods accounted for 43% of recorded natural disasters, claiming nearly 2.5 billion lives [4]. Flooding is considered the most devastating natural disaster in the world's major cities, resulting in ever-increasing economic losses to the community [5]. Flooding is a complex and dynamic process influenced by the interaction between watershed management and various hydro-meteorological, hydrogeological, and geomorphological factors [6, 7].

The Krueng Baro watershed in Pidie Regency is the longest of the six watersheds in Pidie Regency. Damage to the Krueng Baro watershed causes LULC to be increasingly vulnerable to water flow appropriately and quickly for the survival of the majority of the population of Pidie Regency. The Krueng Baro watershed is an area frequently affected by floods, occurring on average three times a year. Based on an analysis of land use, rainfall, land slope, and soil type, this area is highly susceptible to flooding [8]. It has led to several disastrous impacts on the community. The watershed is also vulnerable to changes in LULC, with the primary factor being deforestation [9].

It is necessary to address this damage by finding a treatment model suitable for the Krueng Baro watershed landscape. The development of logistic regression models to relate LULC to water resources and hydrological risks has yet to be used explicitly in solving the problems that occur, such as flooding, erosion, and sedimentation [10-12]. This model has been widely used in modelling urban and regional growth, though less so than the Cellular Automata (CA)-Markov model. CA-Markov does not use variable drivers, only temporally based on the probability of change from previous data [13, 14].

LULC change has the most significant effect on increasing the surface flow coefficient value, which impacts increasing peak discharge due to high runoff [15]. LULC changes in the Krueng Baro watershed indicate changes in the ecosystem that can threaten the function of the area. One way to do this is by utilizing remote sensing technology [16, 17].

The Sendai Framework for Disaster Risk Reduction 2015-2030 highlights the crucial role of land use planning and policy in addressing the root causes of disaster risk, such as rapid and unregulated urbanization, poor land management, and the lack of regulations and incentives for private investment in disaster risk reduction [18]. Although global efforts to integrate flood risk management into urban land use planning have increased, the practical implementation of these strategies still faces considerable challenges [19].

Properly designed land-use strategies can effectively reduce flood risks by restricting development in vulnerable areas, enforcing building regulations to minimize runoff, and allocating designated routes and open spaces to improve response and recovery efforts [20]. A best-fit model that spatially links LULC change to discharge and sedimentation for developing an area and appropriate management system has yet to be found. So, this research aims to build a LULC management model to mitigate flooding in the Krueng Baro watershed.

The Krueng Baro watershed is situated in Pidie Regency, Aceh, Indonesia, covering an area of approximately 210.75 km². Geographically, it lies between 96°0’0” and 96°21’20” eastern longitude (EL) and 5°3’30” and 5°21’20” northern latitude (NL) (Figure 1). Based on the official boundaries defined by the Ministry of Home Affairs in 2022, the majority of the Krueng Baro watershed is located within Pidie Regency. However, a small section of its forested area extends into Aceh Besar Regency.

In the Krueng Baro watershed, there is the Krueng Baro River, which is 29.405 km long, up to the outlet of the Keumala Dam. The downstream section of the Krueng Baro river is situated in Blang Asan Village, within Sigli City, while the upstream originates in Geuni Village, located in Keumala District. The Krueng Baro River is the primary source of irrigation water needs in the Krueng Baro area and PDAM Tirta Mon Krueng Baro, so it is invaluable in meeting water needs, especially in Pidie Regency.

The degradation of the Krueng Baro River is generally caused by the massive extraction of excavation C for building materials, all of which come from within the river [8]. Based on data from the nearest rainfall stations to the Krueng Baro Watershed—namely Sarah Mane, Tangse, and Tiro—the average annual rainfall between 2012 and 2021 ranged from 1,586 to 1,907.3 mm/year at Sarah Mane station, 2,135.4 to 2,479.0 mm/year at Tangse station, and 2,908.4 mm/year at Tiro station.

The satellite imagery used in this study is from Sentinel-2A, which was launched in 2015 and provides the latest data for water area identification. Observations were conducted for the years 2016, 2020, and 2024. The use of time series or long-term analysis allows for detecting changes in LULC over time. With Sentinel-2’s high spatial resolution—10 meters for multispectral bands and 60 meters for the panchromatic band—this study can capture finer details of the Earth's surface, enabling a more precise identification of LULC within the watershed area..

These classifications were determined based on the watershed's landscape characteristics, including: (1) forest, (2) cropland, (3) swamp, (4) rice fields, (5) built-up areas, (6) dry bare land, (7) wet bare land, and (8) water body.

This research has provided an overview of LULC in the Krueng Baro watershed from 2016 to 2024. To analyze landscape transformations, LULC mapping was conducted using Google Earth Engine (GEE). Sentinel-2A imagery, processed through GEE, can be used to detect LULC changes [21]. As part of machine learning techniques, the RF algorithm enables unsupervised LULC classification for more accurate analysis.

The RF model uses spectral features (e.g., bands B2, B4, B5, B6) and indices, e.g., Normalized Difference Vegetation Index (NDVI) and Normalized Difference Built-up Index (NDBI), to classify land use groups, including urban areas, farmland, forests, and water bodies. To train the RF model, pseudo-labels are constructed using clustering techniques for classified. With ensemble learning, which combines the predictions of several decision trees to enhance classification or regression accuracy principles, RF's strength is its capacity to handle high-dimensional data while minimizing overfitting and guaranteeing correct results. The workflow is useful for identifying land use without needing labelled datasets because it includes data preparation, feature extraction, model training, and validation.

Leo Breiman first introduced RF introduced 2001 as a method that can improve accuracy by randomly generating attributes for each node. RF consists of a collection of decision trees that classify data into specific classes [22]. Decision trees are created by defining a root node and ending with several leaf nodes to obtain the final result. The process of forming decision trees in the RF method is similar to the Classification and Regression Tree (CART) process [23], especially in detecting the unsupervised classification for LULC. However, RF does not perform pruning.

The first stage creates the decision tree with a randomly selected subset from the training sample and a random selection offeatures. This approach introduces correlation almost exclusively between individual decision trees at these two levels of randomization, which typically improve the accuracy of the model andlessen the risk of overfitting (assessed by the overall accuracy). To assess the classification errors resulting from this method in comparison to a purely random classification, the Kappa test is applied, as outlined in the following equation [24].

where, N is total points, k is number of classes, R is test classes, and P is classified class.

Sentinel-2A satellite images and RF analysis were used in LULC analysis from 2016 to 2024. The results showed strong performance, with an average kappa value of 0.828 and an overall accuracy of 88.467%. The dependability of the data was confirmed by the land use classification accuracy, which continuously surpassed 85% in 2016, 2020, and 2024. The Mock Model is used to assess how streamflow dynamics are affected by changes in the landscape. Data pertaining to runoff provides important information, especially when high water flow causes flooding because of the river's restricted capacity.

Data generated from LULC analysis, such as the area covered by vegetation or the area that has been urbanized, can be integrated into the model parameters. For example, a decrease in vegetation cover can increase surface runoff and reduce infiltration, thus affecting the calculated river discharge. Thus, the combination of LULC analysis and the Mock Model provides an integrated approach to evaluate flood risk, water resource sustainability and the impact of development policies. The combination of LULC analysis and the Mock Model provides excellent benefits in managing watersheds sustainably amidst climate change and environmental degradation challenges.

By using the results of the LULC analysis as input into the Mock Model discharge simulation, water resource managers can design more effective mitigation strategies, such as creating water catchment areas or forest rehabilitation in critical watersheds This approach supports flood risk management, improves water use efficiency, and designs data-driven policies supporting sustainability of watershed ecosystems. If look at the flood prediction modeling results detected using SRTM data processing results, it can be seen that flood conditions in 2022 will increase by an area of This proves that between 2016 and 2024, LULC will experience a decrease in vegetation of 4,267.7 hectares and an increase in built up area and bareland under development.

The limitation in this research is that the RF method can be used in conjunction with the supervised method. This is because much correct train data is needed for higher accuracy. However, the unsupervised RF in this research has a good application, but it needs to be collaborated with supervised methods to train visible objects better. This research was helped by the choice of Sentinel-2A imagery with a resolution of 10 m, making the image clearer.