DEVELOPMENT OF MODIFIED CART ALGORITHM FOR CLASSIFYING PADDY GROWTH STAGES WITH GRID SEARCH HYPERPARAMETER TUNING METHOD IN KARAWANG REGENCY, WEST JAVA

• Main Author: Fawziyya Masnur, Nadira
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75525
• Institution: Institut Teknologi Bandung

Indonesia is an agricultural country with one of the main agricultural commodities being rice. One of the areas that is the center of rice cultivation in Indonesia, especially in Karawang Regency, West Java. There are several methods of measuring rice production data, one of which is the ASF statistical method. Apart from using these statistical methods, several studies are showing that machine learning models can be developed to predict the rice growing phase precisely and accurately. In this final project, the development of the CART algorithm which is being researched at BRIN is carried out as a method of measuring rice production data. The development of a modified CART algorithm is aimed at improving the performance of the CART model in classifying rice growing phases in the study area of Karawang district, West Java in 2020–2021. The development was carried out using KSA sample data extrapolated with attribute data from the Sentinel-1A SAR satellite imagery. The CART model is modified using the grid search hyperparameter tuning method and ensemble stacking with a random forest model, support vector machine, and gradient boosting as the final estimator. The ensemble stacking method uses a passthrough parameter with a value of True which indicates that the training attribute data is also used to train the final estimator. In addition, several experiments were carried out to modify the data in the form of adding NDVI and NDBI attributes obtained from Landsat-8 OLI imagery and oversampling the rice growing phase class samples (1, 2, 3, and 4). Model development is carried out by fitting the 2020 ASF sample data in the study area. Furthermore, model testing is carried out by comparing the model predictions to the KSA sample data in 2021. The evaluation metrics used are overall accuracy, f-1 score in the rice growing phase class, and kappa score. The result of the best modified CART algorithm obtained from this series of final assignments is a stacking CART model with gradient boosting that uses sentinel-1A SAR image data that has been oversampled on its training data. The CART base classifier model used in stacking CART with gradient boosting is a tuned CART hyperparameter model with optimal hyperparameter results being maxnodes of 1000 and minsamplesleaf of 2. The best modified CART algorithm model appears to increase the OA of the baseline CART model with the lowest difference of 0.70% in training data September 2020 and the highest difference was 6.50% in the training data January 2020. In testing the model using NDVI and NDBI attribute data, it was found that in months with image data that had a low level of cloud cover it succeeded in increasing the OA model, but for months with a high level of cloud cover is seen to reduce the OA of the model. In addition, the application of oversampling also shows an increase in model performance in estimating the area of the rice growing phase class area when compared to the estimated area based on the calculation of the ASF statistical method. Therefore, it is found that stacking CART with gradient boosting has better performance than other models.