No abstract

Empirical Best Linear Unbiased Predictor (EBLUP) has been widely used to predict parameters in area with small or even zero sample size. The problem is when this model should be used to predict the parameters of non-sampled area. Ordinary EBLUP predicted the parameters using synthetic model which ignore the area random effects because lack of non-sampled area information. Thus, those prediction will be distorted based on a single line of the synthetic model. One of idea that developed in this paper is to modify the prediction model by adding cluster information by assuming that there are similiarities among particular areas. These information will be added into the model to modify the intercept of prediction model. Another approach is by adding random effects of auxiliary variable into the previous model in order to modify both intercept and slope of the prediction model. In this paper, simulation process is carried out to study the performance of the proposed models compared with ordinary EBLUP. All models evaluated based on the value of Relative Bias (RB) and Relative Root Mean Squares Error (RRMSE). The results show that the addition of cluster information can improve the ability of the model to predict on non-sampled areas.

Data on rice production is crucial for planning and monitoring national food security in a developing country such as Indonesia, and the classification of the growth phases of rice plants is important for supporting this data. In contrast to conventional field surveys, remote sensing technology such as Landsat-8 satellite imagery offers more scalable, inexpensive and real-time solutions. However, utilising Landsat-8 for classification of rice-plant phase required spectral pattern information from one season, because these spectral patterns show the existence of temporal autocorrelation among features. The aim of this study is to propose a supervised random forest method for developing a classification model of rice-plant phase which can handle the temporal autocorrelation existing among features. A random forest is a machine learning method that is insensitive to multicollinearity, and so by using a random forest we can make features engineering to select the best multitemporal features for the classification model. The experimental results deliver accuracy of 0.236 if we use one temporal feature of vegetation index; if we use more temporal features, the accuracy increases to 0.7091. In this study, we show that the existence of temporal autocorrelation must be captured in the model to improve classification accuracy.

Logistic regression analysis is one of classification methods which is both most popular and common used. This classifier works well when the class distribution in response variable is balanced. In many real cases, the imbalanced class dataset frequently was found. This problem can affect of being difficult at obtaining a good predictive model for minority class dataset. The prediction accuracy generated will be good for majority class but not for minority class. SMOTEBagging is a combination of SMOTE and Bagging algorithm which is used to solve this problem. The purpose of this study is to create a powerful model at classifying the imbalanced data and to improve the classification performance of weak classifier. This study used credit scoring data which is imbalanced data consisting of 17 explanatory variables involved. The result from this study showed that the sensitivity and AUC value from SMOTEBagging Logistic Regression 6858 Fithria Siti Hanifah et al.(SBLR) model is greater than the sensitivity and AUC value of logistic regression model. Moreover, SMOTEBagging algorithm can increase the accuracy of minority class.

Food crops monitoring in developing countries such as Indonesia plays an essential role to support national goals in food security and self-sufficiency. One of the fundamental challenges is plant phase classification task which could help to estimate yield before harvest. In contrast to the conventional field survey method which required a large amount of human and capital resources, we explore a more scalable, inexpensive and real-time method using publicly available remote sensing data, i.e. Landsat-8 satellite. Landsat-8 provides rich spatiotemporal features which could support the detection of numerous vegetation and crop-related indices. However, to accurately classify the plant phase, the existing features require additional spectral pattern from different seasons. We found out the existence of temporal autocorrelation among features of food crops plant phase. The aim of this study is to propose a supervised random forest for features engineering to select the best multitemporal features for the classification of rice plant phase. In this study, we focus on the rice plant phase classification in Banyuwangi Regency, Indonesia as a case study. The ground truth data are the monthly kerangka sampel area (KSA) of average rice plant phase at the regency level which officially released by BPS-Statistics Indonesia. The experimental result shows the accuracy of 0.573 with one temporal feature. Furthermore, incorporating four consecutive temporal features gives higher accuracy gain to 0.727 which shows the temporal autocorrelation. Based on the extensive evaluations, our findings and contributions in this study include: (1) insight to capture the temporal autocorrelation to increase the model accuracy (2) a machine learning classification model which is not sensitive to multicollinearity. Our proposed method provides the potential benefit for the government and statistical agencies towards a more scalable agricultural survey.