Land Suitability Map of GIS Raster and ANN

63 The overall accuracy of validation dataset is 96, and this indicates that the performance system was good. This accuracy can be accepted to the prediction of suitability class processing. As discussed in previously, lower accuracy was achieved due to less representation between the input and output pattern. Also, it may occur because the number of training dataset was less suitable. However, the overall accuracy of 95 was achieved, and this indicates that the system has met the requirement to be continued for prediction. Table 4.2. Accuracy of ANN for validation dataset processing No. Output Label Number of Training Dataset Number of Valid Prediction Accuracy 1. S1 25 25 100 2. S2 83 77 97 3. S3 23 19 83 4. N 2 2 100 Total 233 216 96

4.2. Land Suitability Map of GIS Raster and ANN

Based on the overall accuracy in validation datasets as shown in Table 4.2, then prediction for map classification were conducted. Mapping the suitability classification resulting from ANN was carried out in ArcGIS 9.0 by applying the database join and mapping the legend editor. The map result are shown in Figure 4.2, where each suitability was placed on the land of the coastal area, where 21,202.20 ha 11.24 represent of S1, 87,190.83 ha 46.21 of S2, 6,958.26 ha 3.69 of S3 and 61,251.12 ha 32.47 of N. Most of study areas were covered by its suitability, 6.20 of ‘No Data’ and only 0.19 represents unclassified area. ‘No Data’ class represents the area that was not included in data processing by using raster format. Unclassified areas were produced by ANN, which uncategorized to S1, S2, S3, and N. This may occur due to really unclassified, where predicted output values of ANN in form of [0000] . Otherwise, it may occur due to the ANN output were 64 [1100] or [0110] or [1010]. This mean that those features of output was considered to ‘Unclassified’ due to these output were belonged two or more different suitability. The number of each record was unclassified for the prediction data set are shown in Table 4.3. Table 4.3. The “Unclassified” numbers of the prediction data set. No. S1 S2 S3 N Total Note 1. 8 Unclassified 2. 1 1 5 Belong to S1 and S2 3. 1 1 1 Belong to S1 and S3 4. 1 1 1 Belong to S2 and S3 Total 15 On the map, “Unclassified areas” were distributed on certain location. It may occur due to lack representation of training dataset, so that this affects on the weight value and final map result. Or this can be caused by lack of number of training dataset. The composition value of suitability S1 S2 S3 N was de-normalized from number between 0 and 1. Those values represent the integer format that needed to make the classification. Actually, these values can be modified to make the fix suitability classes. However, this modification will need additional effort in MS Visual Basic 6.0 program. After conducting check and recheck on all above by modifying iteration and logistic constant, then the final overall accuracy was improve 98. It was achieved by setting the momentum constant β of 0.3, logistic constant of 0.2 and iteration number of 500, on which the iteration was stabled at 377 epochs. This shows that the system performance was good, and was indicated by the number of unclassified areas, which was only 0.19 . Geographically, the S1 are located on the back of buffer zone and placed on the ponds zone of spatial plan map. S2 were distributed outside the ponds zone and placed on the buffer zone and excludes of S1 zones. S3 areas are located on a limited 65 area; they are distributed along the main land coast. N is distributed on main land and parts of it were placed on “No Data” of study area. The detailed information of suitability is shown in Figure 4.2. Based on the literature, S1 is the best choice due to no limitation to be developed for shrimp ponds. The S1 distribution is considered appropriate for spatial plan arrangement.

4.3. Land Suitability Analysis from Vector – Map Overlay