A PADDY GROWTH STAGES CLASSIFICATION FOR PISAR-L2 DATA USING FUZZY MODEL

A PADDY GROWTH STAGES CLASSIFICATION FOR PISAR-L2 DATA USING FUZZY MODEL S. Mulyono, Sumargana, I. Fauziyah, E. Kustiyanto Agency for The Assessment and Application of Technology (BPPT), Indonesia BPPT 2 nd building 19 th floor, Jl. MH Thamrin no: 8 Jakarta 134 Phone: 62 21 316 976 Fax: 62 21 316 972 Email: {sidik.mulyono, sumargana, ilvi.fauziyah, eko.kustiyanto}@bppt.go.id ABSTRACT Recently, remote sensing technology plays an important roles in agriculture application, especially for paddy growth stages classification, which is an critical process in prediction of crop production using a remote sensing technology. In this paper, fuzzy model is used to classify six paddy growth stages from 4 polarizations of PiSAR-L2 airborne data. The developed fuzzy model was build and tested on 277 sets of PiSAR-L2 airborne data, which involved paddy field area around Karawang, Subang, and Indramayu district of West Java. Three kinds of membership functions, e.i. linear, phi, and radial basis function (RBF), are studied and compared to obtain the best fit model for classification evaluated by Cohen s kappa coefficient. The results show that the proposed method with three type of membership function have competitive in accuracy, which kappa coefficient reached more than 9%, and it is suitable for for real applications in PiSAR images for paddy growth stages classification. Keywords: paddy, growth stages, classification, fuzzy model, PiSAR INTRODUCTION Indonesia has more than 8.1 million hectare of paddy field to fulfill national food need at least 33 million ton per year [6]. Direct human observations based statistical calculation and estimation of rice production currently used in Indonesia, often causes irregularities since the results tend to be excessive or over-estimated. This often causes speculation to open the policy of rice imports. The condition is further worsened by the problematic import data. Beside enhancement of paddy field area and improvement of other agricultural technical aspects, more reliable harvest area estimation for paddy fields is a critical issue and really needed to support the National Food Security Program that have been promoted and coordinated by Indonesian government. In addition, an accurate and timely rice conditions monitoring and rice harvest area estimation are certainly needed. This problem on rice productivity could be solved by a remote-sensing technology that has the capability to provide information in a timely response with a high degree of confidence. With the advantages of this technology, the latest information and analysis regarding the current state of agricultural land and crops can be presented periodically, quickly and accurately. Moreover, a yield prediction of crop production can be done quickly and covers a large area without damaging the plants. Classification of growth stages is an important component in prediction of paddy production using remote-sensing technology. In our previous work, we proposed a rice yield prediction method based on airborne hyperspectral images with spectral domain using Genetic Algorithms based New Sequence Principal Component Regression (GA-NSPCR) [1]. In the absence of hyperspectral imagery, we extended our previous method to work with 1, meter spatial resolution of MODIS images to classify nine paddy growth stages classification with One Against All (OAA) strategy and balanced branches strategy (BB-OAA) of Support Vector Machines (SVM) [2]. We further extend our previous work on paddy growth stages classification with basic Extreme Learning Machine (ELM) using hyperspectral data [4], and with a new ensemble incremental approach of ELM [3] using MODIS surface reflectance product computed from six bands of MODIS with 5 meter spatial resolution to build some classifiers using ELM which is directly applied for multiclass classification. We also developed a fuzzy model for classification of paddy growth stages based on synthetic MODIS data from hyperspectral data [5]. To improve the accuracy of fuzzy model for six growth stages classification, this paper proposes the rules described in digital code (refered to item 1 above) instead of defined in common subjects and verbs of fuzzy logic, which selected from available data samples by fuzzification process, in such away

can represent all PiSAR-L2 data. This paper addresses this very important and urgent issue on how to effectively predict paddy production through growth stages classification using fuzzy model based on 4 polarization of PiSAR-L2 airborne data with high accuracy for reliable yield prediction. MATERIALS AND METHODS / EXPERIMENTAL Data samples Several PiSAR-L2 images were acquired by airborne campaign involved paddy field area around Karawang, Subang, and Indramayu district of West Java on 4 August in 212, which carried out under cooperation between Indonesian Agency for Assessment and Application of Technology (BPPT) and Japan Aerospace Exploration Agency (JAXA). The area frame sampling (AFS) survey was also conducted parallel with this campaign, to collect ground reference data related to six paddy growth stages, e.i. vegetative-1, vegetative-2, generative, harvesting, bareland, and ploughing, according to the terminology issued by Ministry of Agriculture. In this survey, 26 sampling segments are determined in advance and have 5x5 meter area, where each segment consists of 25 grids of 1 x 1 meter area. Distribution of this sampling segments is shown in figure 1. Figure 1. Distribution of sampling segments laid on Karawang, Subang, and Indramayu District of West Java Figure 2. Lay-out of grids in each sampling segment First of all, the coordinates of the midpoint of each grid 1 x 1 meters in each segments is determined, where more than five hundred of data samples obtained from the midpoint of the grid that containing paddy growth stages information. In next step, coordinates of all midpoint of grids are overlaid with PiSAR images, to collect the digital number (DN) of images for four polarizations, e.i. H-H (horizontal transmit and horizontal receive), H-V (horizontal transmit and vertical receive), V-H(vertical transmit and horizontal receive), and V-V (vertical transmit and vertical receive) related to each growth stages condition. In order to convert into normalized backscatter value, these digital number then are calculated and normalized by equation (1) and (2) bc i = log (DN i ) (1)

bc i bc i = max (bc) (2) Normalized backscatter profiles of four polarizations for each paddy growth stages can be seen in figure 3. The configuration of data samples is shown in table 1, where each growth stages is represented by six digit binary code.. a. Vegetative-1 stage b. Vegetative-2 stage c. Generative stage d. Harvesting stage e. Bareland stage f. Ploughing stage Figure 3. Backscatter profiles for paddy growth stages

Tabel 1. Number of samples No Growth stages Number of samples Binary code 1 Vegetative 1 53 1 2 Vegetative 2 29 1 3 Generative 61 1 4 Harvesting 54 1 5 Bareland 35 1 6 Ploughing 43 1 Classification Using Fuzzy Model Fuzzy logic is an extension of fuzzy set theory initiated by prof. Lotfi Zadeh of the University of California USA, in 1965. Fuzzy logic is a methodology for operational states of a system of law with the expression of language, rather than mathematical equations. Basically fuzzy logic controller is a control method that simulates human thinking by involving uncertainties in system-physical systems. Fuzzy logic controllers make decisions as set by the input signal and the output that occurred earlier in the form of linguistic variables using the general rules. The structure of fuzzy model for paddy growth stages classification using in this paper consists of several parts, e.i. : 1. Crisp input: The input to the model is crisp, and each input consists of PiSAR data set for each polarization, converted and normalized by equation (1) and (2). Each input is devided into 7 ranges, e.i. Very-very-Low (VVL), Very-low (VL), Low (L), Medium (M), High (H), Very-high (VH), and Very-very-high (VVH), which can be written in digital code as, 1, 2, 3, 4, 5, and 6 respectively. 2. Fuzzification: The fuzzification comprises the process of transforming crisp values into membership degree for linguistic terms of fuzzy sets. The membership function is used to associate a grade to each linguistic term. To transform each crisp input into membership degree in fuzzy, we use three type of membership function, e.i. linear, phi, and gaussian function (RBF), as shown in figure 5. 3. IF-THEN rules: Fuzzy sets and fuzzy sets operations are the subjects and verbs of fuzzy logic. If-Then rule statements are used to formulate the conditional statements that comprise fuzzy logic. In this paper, we use compound rules, which are constructed with much more number of rules for each growth stage, that selected from available data samples, in such away that could represent all PiSAR data. The novelty of this paper is that all the fuzzy rules are described in digital code (refered to item 1 above) instead of defined in the subjects and verbs of fuzzy logic, as shown in table 2. 4. Defuzzification: Defuzzification is the process of producing a quantifiable result in fuzzy logic, given fuzzy sets and corresponding membership degrees. These will have a number of rules that transform a number of variables into a fuzzy result, that is, the result is described in terms of membership in fuzzy sets. For inference, we adopted Tsukamoto method in this paper. 5. Binary coding: Crisp output is converted into binary code with condition as follow: If max(crisp output) then binary code = 1 else (3) Where the index of max(crisp output) is denoted the growth stage Table 2. The examples of IF-THEN rules in digital code Rule no IF THEN rules Digital code 1 If HH is H, and HV is L, and VH is L, and VV is VVH, then Y=VVH 4 2 2 6 6 2 If HH is H, and HV is L, and VH is M, and VV is VVH, then Y=VVH 4 2 3 6 6 3 If HH is L, and HV is VL, and VH is VL, and VV is VVH, then Y=VVH 2 1 1 6 6 4 If HH is VVH, and HV is H, and VH is H, and VV is VVH, then Y=VVL 6 4 4 6 5 If HH is VVH, and HV is M, and VH is M, and VV is M, then Y=VVL 6 3 3 3

IF-THEN Rules for Binary coding HH Vegetative-1 class Crisp output 1 /1 Crisp input HV VH VV Fuzzification Vegetative-2 class Generative class Harvesting class Bareland class Ploughing class Defuzzification Crisp output 2 Crisp output 3 Crisp output 4 Crisp output 5 Crisp output 6 /1 /1 /1 /1 /1 Growth Stage Figure 4. Fuzzy model for growth stages classification a. Linear function b. π(phi) function c. RBF function Figure 5. Membership functions for H-H polarization used in fuzzy model

RESULTS AND DISCUSSION Fuzzy model for paddy growth stages classification is developed by constructing the program codes in interactive data language (IDL). The accuracy of classification result for each membership function of fuzzy model is evaluated by Cohen s kappa coefficient calculated from confusion matrix. A confusion matrix (Kohavi and Provost, 1998) contains information about actual and predicted classifications done by a classification system. Performance of such systems is commonly evaluated using the data in the matrix. Kappa values are interpreted as follows: < less than chance agreement,.1.2 slight agreement,.21.4 fair agreement,.41.6 moderate agreement,.61.8 substantial agreement, and.81.99 almost perfect agreement. Table 3,4, 5 show that all type of membership function are strong competitive in term of accuracy, and have kappa value with almost perfect agreement (more than 9%). This phenomena also proved that fuzzy model classifier is not affected by the structure of data set used in learning and is robust to noises in the data. Table 3. Confusion matrix for linear MF (Kappa=.921138) Ground reference data (Actual) Veg-1 Veg-2 Gen Hrv Brl Plg Prediction Plg Brl Hrv Gen Veg-2 Veg-1 51 1 1 19 2 7 1 1 59 1 1 54 34 2 1 43 Table 4. Confusion matrix for π MF (Kappa=.92989) Ground reference data (Actual) Veg-1 Veg-2 Gen Hrv Brl Plg Prediction Plg Brl Hrv Gen Veg-2 Veg-1 52 1 22 1 7 1 58 2 1 54 1 33 2 1 43 Figure 6 and 7 show the distribution map of growth stages using fuzzy model with RBF membership function, where the black color shows the objects that are not defined as paddy growth stages.

Table 5. Confusion matrix for RBF MF (Kappa=.925524) Veg-2 Gen Hrv Brl Plg 52 1 2 2 7 1 58 2 Hrv 1 54 Brl 34 2 1 43 Gen Veg-2 Veg-1 Veg-1 Plg Prediction Ground reference data (Actual) a. Original PiSAR Image b. Growth stage map Figure 6. Growth stage distribution map using 5x5 pixels of PiSAR image a. Original PiSAR Image b. Growth stage map Figure 7. Growth stage distribution map using 1x1 pixels of PiSAR image

CONCLUSION The proposed fuzzy classifier is robust to noises in the data and are able to capture gradual changes of PiSAR data by fitting its membership functions. The three types of membership functions used in this paper guarantee that the fuzzy classifier will not overfit the data. All fuzzy classifier is is suitable and strong recommended for real applications in PiSAR images for paddy growth stages classification. ACKNOWLEDGMENT We would like to thank to Indonesian BPPT and Japan Aerospace Exploration Agency (JAXA) for their support by providing the PiSAR airborne data dan ground reference data of paddy. REFERENCES [1] S. Mulyono, M. Ivan Fanany, T. Basaruddin, Genetic Algorithm Based New Sequence Principal Component Regression (NS-PCR) For Feature Selection And Yield Prediction Using Hyperspectral Remote Sensing Data, International Geosciences and Remote Sensing Symposium, 212 [2] S. Mulyono, M. Ivan Fanany, T. Basaruddin, A Paddy Growth Stages Classification Using MODIS Remote Sensing Images with Balanced Branches Support Vector Machines, International Conference on Advanced Computer Science and Information Systems, 212 [3] S. Mulyono, T. A. Pianto, M. Ivan Fanany, T. Basaruddin, An Ensemble Incremental Approach of Extreme Learning Machine (ELM) For Paddy Growth Stages Classification Using MODIS Remote Sensing Images, International Conference on Advanced Computer Science and Information Systems, 213 [4] N. Suhandono, S. Mulyono, F. Maspiyanti, M. Ivan Fanany, An Extreme Leaning Machine Model For Growth Stages Classification of Rice Plants From Hyperspectral Images Subdistrict Indramayu, The Second Indonesia -Japanese Conference on Knowledge Creation & Intelligent Computing, 213. [5] M. Widjaja, S. Mulyono, A. Darmawan, Fuzzy Classifier of Paddy Growth Stages Based on Synthetic MODIS Data, International Conference on Advance Computer Science and Information System, 212 [6] Roadmap of National Rice Production Enhancement (P2BN) Towards 1 Million Tons Surplus Rice in 214, annual report, Ministry of Agriculture. [7] Roman Liberacki, Effect of Adopted Rules of Inference and Methods of Defuzification on The Final Result of The Evaluation of Reliability Made Using The Fuzzi Logic Method, Journal of Polish CIMAC, Gdansk University of Technology [8] Stephen M Haley and J Scott Osberg, Kappa Coefficient Calculation Using Multiple Ratings Per Subject: A Special Communication, Journal of the American Physical Therapy/Volume 69, Number 11/November 1989.