2.1

Research Location

The study area is located around Menoreh Hills, which is part of Kulon Progo Regency. This location was selected as the study area due to its topographic characteristics, ranging from gentle slopes to hilly and even mountainous. It was an ideal condition to be used as a basis for such research, in which the geometric correction process considers the differences in topography. In addition, in the study area, there were variations in land cover types, particularly in the form of high to low density vegetation, built up areas, and water bodies.

Figure 1.

Research location

2.2

Data processing stage

2.2.2

Geometric correction of Landsat 8 L1G images

Ground control points (GCP) were taken using reference images, that is Sentinel-2A images. Sentinel-2A imagery has a 10 m of spatial resolution on visible-NIR spectrum. The use of the VNIR spectrum due to the need for imagery analysis is limited to the visual interpretation of land-covers only. The results of the visual interpretation are used as a reference map for accuracy testing in the multispectral classification of Landsat 8 Imagery, so the pixel accuracy test for polygons will be better if the land cover polygon more detailed. Polygon-based accuracy testing for validation of extensive land cover maps is often seen as a cost-effective means of associating some level of confidence with land cover products[7]. On the other hand, the spatial resolution of sentinel images is three times more detailed than Landsat 8 imagery, so that at least one pixel value Sentinel-2A is one-third pixel of Landsat 8. If there is an error in accuracy, the accuracy is less than 0.5 pixels Landsat 8 image, then the value is still allowed to test image accuracy with a resolution of 30 m (Landsat Image)[8]. The pattern of distribution of ground control points basically produces the same imagery geometrically. However, the distribution needs to be considered because it is related to the final results which consider geometric accuracy and semantic accuracy of the map. In this study three models of GCP’S distribution patterns were used, i.e.: (a) pattern of control points on the edges of the image; (b) poving control point pattern (c) the pattern of control points is spread evenly. At least the magnitude of the RMSE value that can be received is 0.5 pixels. The RMSE and STD formulas are as follows[9]:

x′,y′ :the image coordinate resulting from geometric correction

(x, y)_orisin : represents GCP coordinates in the reference field

n : the number of GCP’S

2.2.5

Sampling

This study uses stratified random sampling because it considers land cover classes in the image and in the field, and the affordability of objects in the field. In the accuracy test, the position of the object in the image with the position of the object in the field must have the same coordinates. Authors as mentioned in[10] suggested an equation that can be used as a reference in determining the minimum sample area, that is:

note:

A = minimum sample size in the field

P = image pixel size

L = approximate location accuracy, in pixel size (0-1)

2.2.6

Accuracy Test

A common method used to assess errors is the Error Matrix or often known as the Confusion Matrix[11]. Overall accuracy can be generated from the comparison of the correct number of pixels to the total number of pixels in the image. If the results of classification of remote sensing data show an accuracy of 85%, the data information is acceptable. Assessment of accuracy tests also uses the kappa coefficient[10]. Formulate the Kappa coefficient as follows[13]:

note:

K : Kappa’s Coefficient

N : the number of observation total

r : the number of row in error matrix

x_ii : the observations number of row i column i

x_i+ : observations total on row i

x_+i : observations total on column i

3.1

Visual Interpretation of Land-Cover

The land-cover classification is based on the level 2 multipurpose classification scheme of Danoedoro’s[6] which has been modified on a scale of 1: 10.000 to 1: 30.000. The land-cover class produced from the results of visual interpretation amounted to 14 classes, that is: Deep water (C11), turbid water (C14), woody broadleaves 1 (C21), woody broadeaves 2 (C21), non-woody broadeaves 1 (C22_1), non-woody broadeaves 2 (C22_2), non-woody broadeaves 3 (C22_3), needle leaves (C23), dry soil (C31), moist soil (C32), moist and wet surface (C33), Asphalt and cemented surface (C41), hardened clay surface (C42), asbestos surface and zinc (C44). The results of the land cover classification using visual interpretation methods produced 1394 land cover polygons divided into 14 classes (figure 3). Sampling for land cover accuracy test is based on the Slovin formula[14], which is formulated as follows:.

Figure 3.

Land-cover map of visual interpetation result of Sentinel-2A Imagery

note:

n : number of sampel

N : number of population

e : desired level of significance

Population number or land cover polygon is 1394, so based on the Slovin formula obtained 263 minimum samples with a significance level of 0.55 (94.5%). The results of the field data accuracy test produce land-cover maps with an overall accuracy of 90% with a kappa index value of 0.85. Table 2 shows the producer’s accuracy and user’s accuracy of land cover results from visual interpretation.

Table 2

producer’s accuracy and user’s accuracy

3.2.

Geometric Correction

Geometric correction using polynomials requires some object to be used as tie points. The objects in the form of static appearance, such as roads, rivers, reservoirs, bridges, and permanent buildings. Geometric correction using affine method with 6 ground control points that were distributed on the edge of the image produced an RMSE of 0, 355539. Geometric correction with the second order polynomial method using 10 ground control point that were distributed around the image produces RMSE of 0.178053; whereas geometric correction with 3rd order polynomial method of field control points used amounted to 17 binding points which were evenly distributed in the image resulting in an RMSE of 0.100343. The RMSE value from geometric correction using the 3rd order polynomial was the best value as compared to the affine and 2nd order polynomial methods. The residual value of the image geometric correction is presented in Table 3.

Table 3.

Residual value of geometric correction

This study found that the factors influencing the RMSE values were the distribution of ground control points, number of ground control points, and order of polynomials. As viewed from those three factors, the level or order of polynomials was the most influential one. The more number of ground control points that used, followed by an increase of the order of polynomials, the better the RMSE value.

3.3.

Resampling

The nearest neighbor, cubic convolution, and bilinear interpolation algorithms in geometric correction produced images with different pixel values. Generally speaking, visual differences can be seen in the results of resampling with the nearest neighbor algorithm, since the algorithm produces jagged images with broken appearance of pixels. On the other hand, the cubic convolution and bilinear interpolation algorithms produce finer and smoother image appearance. These differences can be seen in linear objects such as rivers, which can be seen in Figure 4.17.

Figure 4.

Comparison of the results of resampling the image resulting from geometric correction

The results of nearest neighbor resampling have fewer mixed pixels than those produced by other two algorithms. In image processing, the nearest neighbor algorithm also requires a fairly short time compared to bilinear interpolation and cubic convolution. The influencing factor is the number of pixels used as a reference, as already explained that bilinear interpolation uses the four closest neighbors, while the cubic convolution uses the 16 closest neighbors. But both algorithms produce finer output pixels. So that the two algorithm is good if used for visual analysis.

3.4

Multispectral Classification Maximum likelihood

The use of the maximum likelihood algorithm is based on an algorithm that is considered statistically well established[1]. In addition, the difference in pixel values in each geometric and resampling correction result was also used as the basis for selecting the maximum likelihood algorithm, because this algorithm considers the pixels’ probability values to be assigned to particular class. In conducting a multispectral classification it is necessary to consider several things such as image composite selection, and training areas or ROIs. Image composite that used was 654 (SWIR, NIR and Red), which was based on the clarity of the appearance of the image, which can distinguish various types of vegetation. The selection of image composites was also based on the statistical parameters used, i.e. Optimum Index Factor (OIF) developed by Chavez et al.[15]

Table 4

OIF calculation results for 6 channels

Retrieval of ROI serves as a reference in creating land cover classes in the study area. The land cover class used is the same as the visual interpretation, which refers to the multipurpose classification of[6]. Differences in visual interpretation using Sentinel-2A imagery and multispectral classification based on 8 Landsat Imagery was that the later did not include the class of asphalt cover and cemented surface (C41). The choice of training area was based on the spectral characteristics of the object, i.e. hue and color. Furthermore, to assess the training area taking, statistical calculations were performed using the separability index. Evaluation of the level of separability of the sample were counducted using Transformed Divergence and distance of Jeffries-Matusita (JM). JM distances, which ranges between 0 and 2, provide a general measure of the separation between the two classes according to their possibilities[16].

The multispectral classification process with the maximum likelihood algorithm was implemented using ENVI 5.3 software. The results of the land cover classification with the maximum likelihood algorithm produced nine land cover maps. Each map was tested for accuracy on maps of visual interpretation and appearance of sentinel-2A images. Accuracy tests was carried out using a square-shaped polygon with an area taking into account the shift of pixel image to a map or reference image. Polygons to test the accuracy of each method are shown in Figure 5.

Figure 5.

comparison of the area of the test polygon in the wide leaf cover class; not woody 1

The results of the multispectral classification were assessed using the configuration matrix and Kappa index tables presented in Tables 5 and Table 6. From these results it can be seen that the transformation using 3^rd order polynomials produced the best accuracy value among other geometric correction methods. The 3^rd order polynomial with the nearest neighbor resampling method produced the highest accuracy of 75.43% with a Kappa coefficient of 0.70298. Meanwhile, the transformation of 3^rd polynomials with bilinear interpolation and cubic convolution resampling methods yielded the same values, i.e. 72.2359%.

Table 5.

Overall accuracy of the results of the multispectral classification

Table 6.

The kappa index coefficient results from the multispectral classification

From the calculation of overrall accuracies, the transformation of the 2^nd order polynomial with the nearest neighbor algorithm produces a land cover map with the lowest accuracy value of 68.0590%. The affine transformation with bigger RMSE value, as compared to the 2^nd order polynomial transformation, produced a fairly good accuracy on the three resampling algorithms. Therefore, this study found that the higher order polynomial transformation was not a guarantee to produce output images with better spectral quality (for multispectral classification) in comparison with the lower order polynomials.

The accuracy of the land-cover classification results did not meet the requirements mentioned by Aronoff[12], which says that the accuracy at least 85%. However, this study was limited to the assessment of the best algorithms to produce land cover maps. Some other factors that lead to the low accuracy were also identified, and three of them were the presence of cloud cover, differences in image recording, and rotation of rice planting. Thin clouds have almost the same digital value as asbestos and zinc surfaces object. In addition, samples taken in the field which are then plotted on the image are often under the cloud in Landsat 8. So that it affects the accuracy of the object when conducting the accuracy-test results of multispectral classification. Rice planting rotation directly affects moist soil and moist and wet surfaces. These two objects are almost the same in the imagery, the difference is that moist soil has a brighter hue and is associated with broad woody leaves 3. Moist open land in the field can be found in post-harvest paddy fields and before planting rice seeds. Dark hue on damp objects and wet surfaces indicate that there is water on the object, besides that the damp open ground objects are side by side with broad, woody leaves 1 so that sometimes on land A rice is already mature while on land B rice is still newly planted and inundated by water. It is known that in the Dry Season (MK) of 2018, the first planting period is carried out in the third week of March to the first week of April. Whereas Sentinel-2A imagery being recorded on May 4, 2018, so that the average paddy is 1 month old. So that the condition of the object in the interpretation of the image is different from the conditions in the field. especially Landsat 8 imagery was recorded on May 29, 2016, the differences of recording condition with field conditions differed 3 years so that the land cover condition changed a lot.