Projected land cover changes in 2031 in Awangpone watershed, South Sulawesi, Indonesia

One of the factors of watershed damage is that water catchment areas are decreasing due to land use changes caused by various human activities such as logging, forest encroachment, and development carried out in watersheds. Continuous changes without consideration in land use management will adversely affect all elements in a watershed, especially land closures that have a major impact on water availability. This study aims to project land cover changes by 2031 in the Awangpone river basin, Bone Regency, South Sulawesi. This study used a Cellular Automata-Artificial Neural Network (CA-ANN) simulation to make projections of land cover changes in 2031. The results of this study show a change in land cover from 2020 to 2031. Land closures that experienced an increase in area were in the residential class of 214.15 ha or 1.69% and the fishpond class of 58.38 ha or 0.46%. Meanwhile, dryland agricultural classes mixed with shrubs, shrubs, mangrove forests and rice fields have experienced a reduction in area. The changes that occurred both the reduction and the increase in the area of the land closure class in the Awangpone watershed were influenced by land conversion, where the addition of settlements caused a class reduction in other land closures.


Introduction
The problem faced by most major cities in Indonesia is urbanization which results in an increased need for land/ housing.Physical development that is not based on environmental sustainability will have negative consequences in the future, which can be in the form of floods or reduced groundwater.Moreover, the decrease in the vegetation can make negative influence to the stability slope [1].Water infiltration is inseparable from green open space in a city.Ideally every city has green open space covering an area of 30% of the city area which can be lakes (reservoirs, lakes, or situ), parks, and urban forests [2].
The Awangpone Watershed with an area of 12,840 ha is one of the watersheds whose river flow is used by the Bone community, especially the capital of Bone Regency.The Awangpone watershed condition plays a very important role in supporting the availability of water that is interrelated with each other both upstream and downstream, especially the 5 sub-districts located in the Awangpone watershed area.The upstream Awangpone watershed is one of the most important areas in the overall watershed system.This is because the land use changes that occur in the Upper Awangpone watershed will have 1230 (2023) 012046 IOP Publishing doi:10.1088/1755-1315/1230/1/012046 2 further implications for the area below it (downstream).So that any changes that occur / are made in the watershed must be carefully calculated.
The shortage of clean water in Bone County over the past 5 years has become a threat to the clean water crisis and the food crisis.High population, land use change, and reduced forest areas and water catchment areas greatly affect the availability of diminishing water.One of the factors of watershed damage is that water catchment areas are decreasing due to land use changes caused by human activities such as logging, forest encroachment, and development carried out in watersheds.This study aims to project land cover changes by 2031 in the Awangpone river basin, Bone Regency, South Sulawesi.

Material and methods
The tools used in this research are: GPS receivers, cameras, writing stationery, and laptops equipped with Geographic Information System (GIS) software.The materials needed are in the form of map data used in this study, namely the watershed boundary map, the 2009 land use / closure map, the 2015 and 2020 land use/ cover maps, river network maps, 30 m DEM daisies, road network maps and community activity center maps.

Study area
This research was conducted in the Awangpone Watershed which covers several areas in Bone Regency.This research was conducted from December 2020 to March 2023.

Preparation data
The data collected in this study are primary data and secondary data.Primary data include digital elevation model (DEM) data obtained from topo to raster analysis, research site boundary data and land cover data obtained directly through satellite image interpretation activities.Checking land cover data in the field is carried out by adjusting the classification of land cover results of image interpretation by purposive sampling based on accessibility.The accuracy test of the image interpretation results is carried out using the confusion matrix method which uses the calculation of overall accuracy (overall accuracy) and kappa accuracy (kappa index value).

Factors driving land cover change.
The driving factor is an independent variable in building a transition probability model, in previous studies the driving factors used are existing factors that will cause changes in land cover, in this case the distance factor of existing land to the surrounding land use [3].Distance an influential factor in spatial dynamics, including changes in land cover.

Projected stage of land cover change. Projection of land cover changes is carried out using an Artificial
Neural Network (ANN) system and Cellular Automata (CA) simulation on QGIS software with Mollusce Plugins.ANN can be represented by linear algebraic notation [4]: Where,  : Output  : Activity Functions , : Learning Parameters ,  : Bias or noise Artificial neural networks study patterns from given inputs (2009 and 2020 land closures as well as independent variables) in the form of a transition matrix of changes from the previous year to the actual year and then produce solutions (possible) changes that will occur.Formal form of Cellular Automatan A, consisting of four components formulated as follows [4]: The results of the 2031 land closure projections that have been tested for accuracy are then overlayed with the 2020 land closure map to see changes in the area of land cover.Data validation was carried out using a previously projected 2020 land closure projection map with land closure input data for 2009 and 2015.[5] The kappa value of 0.81 -1.00 indicates an excellent accuracy value, the kappa value of 0.61 -0.80 is good, the kappa value of 0.41 -0.60 is medium, 0.21 -0.40 is less than medium, and the kappa value of <0.21 is said to be bad.

Results and discussion
Land cover resulted from the interpretation of imagery in landsat 7 imagery in 2009, landsat 8 imagery in 2015 and 2020.From the results of this imagery interpretation, it shows that there are six classes of land closure located in the Awangpone watershed area with different areas.The class and extent of land cover in the Awangpone watershed can be seen in Table 1 and land cover in 2020 can be seen in Figure 1.
Table 1 From the results of the image interpretation, it shows that the area of land closure classes in 2009 and 2020 as in Table 1 has changed both additions and subtractions.The largest addition of area in the residential class experienced an increase of 4.37% of the total watershed area or 554.49ha, this addition occurred due to the conversion of the rice field class, mixed dryland agriculture with shrubs, shrubs, fishponds and secondary mangrove forests, The increase in area also occurred in the fishpond class which increased by 0.38% of the total watershed area or 48.56 ha.This addition occurs due to the dominant conversion of the secondary mangrove forest class.This is in line with research conducted by [6] that coastal community activities that convert mangrove forests are used for residential land, agriculture or plantations, anchorage, salt making, and mining.
The largest reduction in land cover area occurred in the bush mixed dryland agricultural class by 3.41% of the total watershed area or 432.50 ha reduction in the bush mixed dryland agricultural class converted to residential, fishpond and rice field classes.In the shrub class experienced a reduction of 0.91% of the total watershed area or 115.77ha, the shrub class was converted to a dry land agricultural class mixed with shrubs, settlements and rice fields.A map of the changes in land cover from 2009 to 2020 can be seen in Figure 2.

Projected stage of land cover change
The data used in this projection is in the form of data on dependent variables and independent variables.Dependent variables include land closure in 2009 (initial) and land closure in 2015 (final).Independent variables (driving factors) are the distance to the road, the distance to the settlement, the distance to the public facility point and the distance to the river.A map of road distances, residential distances, distances of the central points of community activities can be seen in Figure 3.

Evaluating pearson's correlation
The relationship between independent variables to land cover changes is calculated in the range of values 0-1, where 0 indicates no linkage, while value 1 indicates a very close relationship between independent variables used to land cover changes.Independent variables that have a high correlation value are assumed to have the same characteristics, so they can give rise to multicollinearity that can interfere with the regression process [7].There is a factor that has a high correlation value or is above 0.7 then one of the factors can be eliminated to prevent multicollinearity [3].The results of the calculation of Evaluating correlation can be seen in table 2. The results of evaluating correlation in table 2 can be seen that all independent variables used have a relationship with varying correlation values, but in independent variables the distance to settlements has the highest correlation value, which is close to 0.7 this happens because the spread of road networks, settlements and points of public facilities have similarities, especially in the middle and downstream areas of the watershed.To prevent the occurrence of multicollinearity in the next stage, only three independent variables are used, namely the distance to the road, the distance to the point of public facilities and the distance to the river.

Area changes
The 2009-2015 land closure change is a dependent variable that will be the rule for changing a land closure in the projected year.In the tools area changes, a table of transition matrix changes is generated (Transition Matrix).The change transition matrix is a relationship between pixels and land cover classes made into a rule that can change the value of a cell or pixel in the next period of time [8].The Transition Matrix table can be seen in Table 3.The Transition Matrix shows simple rules for land cover changes in the next period that will be rules for Cellular Automata simulations.This rule is obtained from the calculation of the area of land closure in 2009 which is the basis for reference then compared to the area in 2015 to see the opportunity for change.From Table 3, it can be seen that the greatest chance of land cover change occurs in the secondary mangrove forest class that will be converted into fishponds, namely 0.310553, the change in the shrub class to bush mixed dryland agriculture is 0.288706.The greatest chance of a land closure class that would be the same land closure occurs in the Residential class of 1.000000.

Transition potential modeling
The probability of change with the Artificial Neural Network (Multilayer Perceptron) method is obtained by conducting training on dependent variables and independent variables.The learning process in the ANN method is carried out Learning and Recall on a phenomenon repeatedly until it gets a pattern with high accuracy [9].The training process is carried out repeatedly by changing the variations of hidden layer parameters, iterations, and learning rates, to get the smallest error value [10].At this stage, the smallest Min Validation Overall error value was found to be 0.02394 from all training experiments conducted.The results of the training model can be seen in Table 4.The parameter values in Table 4 are the values used so that the smallest error value is found.In this training, the variation of parameters used is the neighborhood parameter or neighborliness with a value of three pixels (neighborliness Moore and von Neuman).Neighborliness here is defined as cells adjacent to the cell concerned [10].The maximum iterations parameter or repetition process with a value of 100 and at a repeat value of 100 the smallest error results are obtained.Increasing the number of iterations to the optimum iteration limit increases the error value.Large iteration values do not always provide good accuracy [11].
A learning rate parameter with a value of 0.100 learning rate that is too low will cause the algorithm to converge more and more.The momentum parameter determines the magnitude of the weight change 1230 (2023) 012046 IOP Publishing doi:10.1088/1755-1315/1230/1/0120468 of a training, Momentum with a value of 0.055 is considered to have the best performance [11].The optimum hidden layer value of 10 results in a good or low error value.From the results of several network simulations carried out, it shows that variations in different parameters are very influential, especially learning rate, hidden layer, and iteration.

Cellular Automata simulations
The cellular automata simulation stage produces a projected map of land cover in 2020 which refers to transition potential modeling results from the training process.The class and extent of land cover in 2020 simulation results can be seen in Table 5.The area of each land cover class in 2020 as shown in Table 5 is the result of cellular automata simulations.In Table 5 it can be seen that within a period of five years there was a change in area when compared to the land cover class in 2015 there were four closure classes that experienced a reduction namely the dry land agricultural class mixed with shrubs, secondary mangrove forests, shrubs, and rice fields.In addition to reductions, there was also an increase in area in the class of settlements and fishponds.

Validations
The validation results of the projected land cover change resulted in a Kappa value of 0.93.[12] Kappa value is said to be very good if the Kappa value > 0.80.The results of calculating the Kappa value using validation tools can be seen in Figure 4. Based on the validation results carried out with a kappa value of 0.93, it shows that the cellular automata simulation has very good accuracy.So, it is concluded that these cellular automata simulation can be used to project changes in land cover in 2031.The projection of land closure changes in 2031 is carried out by running a validated Cellular Automata simulation.In the projected land cover change in 2031, there are six classes of land cover including, secondary mangrove forests, settlements, mixed dryland agriculture, rice fields, shrubs and fishponds.The area of land cover in 2020 and 2031 can be seen in Table 6.Based on the results of projected land cover changes in 2031, there are four classes of land that have been reduced, namely secondary mangrove forests, rice fields, shrubs and dryland farms mixed with fishpond bushes.In addition, there are two classes that have experienced an increase in area, namely the residential and fishpond classes.The percentage of land cover changes in 2020 and 2031 can be seen in Figure 5.The highest area increase occurred in the residential class by 1.69% and fishponds by 0.46%.In addition to the addition, there was also a reduction where the highest reduction occurred in the dryland agricultural class of mixed shrubs which was reduced by 0.9% while the lowest reduction occurred in the shrub class by 0.25%.
The addition of residential classes is due to the dominant conversion given by the rice field class, the rest of the agricultural class of land mixed with shrubs and shrubs, the addition of the dominant settlement class occurs in the central area of the Awangpone watershed which leads to the downstream area.The fishpond class experienced an increase in area of 0.46% due to the dominant conversion given In addition to the addition, there was also a reduction where the agricultural class of dry land mixed with shrubs and rice fields became the class that experienced the largest reduction, namely 0.9% for dryland farming mixed with shrubs and 0.79% for rice fields.This happens because it is predominantly converted into a residential class.
Changes that occurred both the reduction and the increase in the area of the land closure class in the Awangpone watershed were influenced by land conversion, where the addition of settlements led to a reduction in other classes.The 2031 land cover map can be seen in Figure 6.

Figure 3 .
Figure 3. (a) Map of distance to road, (b) distance to settlement and (c) distance to river.

Figure 4 .
Figure 4.The result of the calculation of the kappa value.

Figure 5 .
Figure 5. Percentage change in land closure class in 2020 and 2031.
forest class.The change in the class of mangrove forests to become the dominant fishpond occurred in the downstream area of the watershed and a considerable change occurred in the Barebbo district.

Table 2 .
Evaluating correlation on independent variables.

Table 4 .
The results of the training model of the Artificial Neural Network method.

Table 5 .
Land cover in 2020 simulated by cellular automata.

Table 6 .
Land cover class area in 2020 and 2031.