Mining Sequence Pattern on Hotspot Data to Identify Fire Spot in Peatland

Transcription

1 International Journal of Computing & Information Sciences Vol. 12, No. 1, December Automated Context-Driven Composition of Pervasive Services to Alleviate Non-Functional Concerns Imas Sukaesih Sitanggang and Erna Fatayati Pages DOI: Mining Sequence Pattern on Hotspot Data to Identify Fire Spot in Peatland Imas Sukaesih Sitanggang 1 Erna Fatayati 2 Computer Science Department, Bogor Agricultural University Darmaga Campus, Jl. Meranti Wing 20, Level V, Bogor 16680, INDONESIA 1 imas.sitanggang@ipb.ac.id, 2 ernafata@gmail.com Abstract: Fires on peatland occurs each year in Sumatra Island, Indonesia. A hotspot is an indicator of fire occurrences in dry land or in peatlands. Peatland fires can be identified by extracting sequential patterns in a hotspot dataset. Hotspots occurred two to five days consecutively have high potency to become fire spots indicating real fires (fire spots). This work aims to generate sequential patterns on hotspot datasets in Sumatra Island Indonesia in 2014 and 2015 using the SPADE algorithm. In addition association rule mining was conducted to obtain association between the locations of hotspot sequences and weather conditions. The results show that sequence patterns of hotspot in Sumatra in 2014 were occurred in the villages which have the weather conditions: average humidity between 70.1% and 78.2%, average temperature between C and C, and precipitation of 0 and 0.9 mm. In addition, sequence patterns of hotspot in Sumatra in 2015 were occurred in the villages which have the weather conditions: average humidity between 68.6 and 77.7%, average temperature between and C, and precipitation of 0 mm. The sequence patterns of hotspot are strong indicator for fire spots. Identifying the sequence patterns of hotspots associated with the weather conditions of the locations where the sequences occurred will be beneficial for related parties in peatland fires prevention. Keywords: Association Rule Mining, Data Mining, Hotspot, Peatland fires, Sequential Pattern Mining. Received: May 25, 2016 Revised: August 14, 2016 Accepted: September 12, Introduction Peatland has a very important role in maintaining the balance of the ecosystem, especially in water regulation, biodiversity conservation, climate change mitigation and support human welfare. Peatland area covers more than 400 million hectares in 180 countries and represents a one-third of global wetlands [7]. Despite tropical peatlands cover only 10-12% of total peatland in the world, but tropical peatlands have a significant role as natural resource that is valuable and important in global environment [9]. Forest fires and peatland is one of the environmental problems in Indonesia that have an impact on the global environment and the ASEAN region, particularly with regard to "transboundary haze pollution" and the emission of greenhouse gases such as carbon dioxide that is an important element in global warming. Emissions of carbon produced from peat fires are the biggest contributors of carbon emissions nationwide followed by changes in land and forest sector, industry, transport and households. Controlling forest and peatland fires requires an early warning system and an early detection system. Hotspot is still an indicator of forest and land fires in Indonesia. This is due to the accuracy of hotspots that is acceptable and accessibility of hotspot data. Not all the hotspots become strong indicator of forest and land fires. The incidence of hotspots consecutively at least two days in a location can be a strong indicator of land and forest fires, particularly in peatland. Spatio-temporal data mining approach, sequential pattern mining particularly, can be used to obtain sequence patterns of hotspots in question. Furthermore, sequence patterns of hotspots are analyzed based on the characteristics of the area where the sequences found. The characteristics include the type and depth of peat soil as well as weather conditions in the region. Sequential pattern mining is extracting sequential patterns that its support value exceeds the pre-defined minimum support. Based on the support values, patterns that are less attractive can be ignored so that mining process becomes more efficient [12]. Some sequential pattern mining algorithms include

2 144 International Journal of Computing & Information Sciences Vol. 12, No. 1, December 2016 Generalized Sequential Pattern (GSP), SPADE, Prefixspan, and Clospan. GSP is one of the algorithms for solving sequential pattern based on the Apriori principle in pruning candidate sequences. SPADE (Sequential PAttern Discovery using Equivalent Class) was developed by [11]. The SPADE algorithm applies a vertical format sequential pattern mining method in which a sequence database is mapped to a large set of item or event. Prefixspan uses the projected data base by creating a prefix of the sequence in extracting sequential patterns [8]. CloSpan (Mining Closed Sequential Patterns) reduces the number of redundant patterns by applying Backward Subpattern and Backward Superpattern pruning to prune redundant search space. Research related to the application of sequential pattern mining techniques on hotspot datasets have been conducted by [1] and [6]. Reference [1] applied the Clospan algorithm [5] which is its implementation is available in the Sequential Pattern Mining Framework (SPMF) [10]. In this work, sequential patters were extracted from the dataset with several minimum support values in the range 1% to 20%. The dataset consists of hotspots and precipitation in the period 2001 to 2010 in Riau Province Indonesia. The study showed that hotspots were occurred sequentially in temperature of C. Another study by [6] applied the Prefixspan algorithm [4] to determine sequential patterns on hotspot datasets. Sequential patterns of hotspots with a time span of 2 to 3 days are considered as strong indicator for forest and land fires in Riau Province Indonesia. The study states that about 16.95% of 18,856 hotspots in 2014 were occurred sequentially [6]. The objective of this study is to discover sequential patterns of hotspots in Sumatra Island Indonesia in 2014 and Furthermore, associations between the sequential patterns of hotspots and weather condition in Sumatra are identified using the Apriori algorithm. The weather data include humidity (%), average temperature ( C), and rainfall (mm). In section (2) we discuss the study area and data used in this work. Section (3) discusses the steps performed to obtain the objective of this work. In addition, in Section (3) we provide an overview of sequential pattern mining and association rule mining. Result and discussions are briefly discussed in Section (4). We end with summary and future work in Section (5). 2. Study Area and Data The data used in this study are hotspots in Sumatra Island Indonesia in the period of January 2014 to December The hotspots data are selected only those have confidence equal to or greater than 70%. The data were collected from National Aeronautics and Space Administration (NASA) Fire Information for Resource Management (FIRMS) ( The attributes of hotspots that are used to generate sequence pattern of hotspots include longitude, latitude, and acq_date. Longitude and latitude represent the location of a hotspot. acq_date denotes the acquisition date of a hotspot. Number of hotspots is 14,237 in the dataset Sumatra 2014 and 16,720 in the dataset Sumatra In addition to hotspot data, peatland layer and district border layer of Sumatra Island were used to select the study area and to analysis the sequence patters of hotspot. Peatland layer consists of three main features namely peat type, peat depth, and land cover. This layer was collected from Wetland International. Weather data were collected online from Bureau of Meteorology and Climatology Indonesia that are available on in Microsoft Excel format (.xls). The weather dataset is in the period including average temperature ( C), average humidity (%), and precipitation (mm). 3. Methodology There are three main steps in this work: data preprocessing, sequential pattern generation from hotspot datasets, association rule mining on the dataset of hotspot sequences and weather condition in Sumatra Indonesia 2014 and Data pre-processing consists of several stages, namely, data selection, data cleaning, data transformation, and sequential dataset generation. Data selection was conducted to select the data and attributes that are used in research. In addition, we eliminated hotspots which are outside the peatlands in Sumatra. Data transformation was performed in preparation of sequential dataset. Data cleaning handled missing values on weather data that are represented as 9999 and The 9999 indicates the absence of data, while 8888 indicates the data are not measurable. Missing values were replaced by the average value of weather data on a day before and after the date of missing values. In the step of sequential data generation, a task relevant dataset was prepared as the input of the SPADE algorithm. The dataset includes SID, Tid, size and items. SID represents the attribute of longitude and latitude, Tid denotes date code, size represents number of hotspot occurrences in the same location and date, and items denotes a list of date code that appear as many value of size. These preprocessing steps were done utilizing the software R, Quantum GIS PostgreSQL and PostGIS. 3.1 Sequential Pattern Mining Sequential pattern is a sorted list of an item, data, or event. Set of the items contained in a sequence is called sequence elements. Suppose D is a database of customer transactions. I = I 1, I 2,..., I m is a set of m items. T is a transaction that contains {customer_id, transaction_time, item_purchased}. s i is an itemset that contains itemset I. S is a sequence containing a sorted list of itemset <s 1, s 2,..., s n> [12]. In sequential pattern

3 Mining Sequence Pattern on Hotspot Data to Identify Fire Spot in Peatland 145 mining, we discover subsequences that are frequently found in the sequence dataset. Suppose given sequence α = <a 1, a 2,..., a n> and β = <b 1, b 2,..., b m>, α is a subsequence of β, denoted α β if there is an integer 1 j 1 <j 2 <... <j n m such that a 1 b j1, a 2 b j2,..., a n b jn. β is a supersequence of α [2]. Sequential pattern mining extracts sequential patterns that its support value exceeds pre-defined minimum support. Support of the sequence s is defined as the fraction of transactions that support these sequences. Based on the minimum support value, the patterns that are less attractive can be ignored so that the mining process becomes more efficient [12]. This study applied SPADE (Sequential Pattern Discovery using Equivalence classes) [11] to find sequence patterns on hotspot datasets. Figure 1 shows the outline of SPADE algorithm to determine frequent 1-sequences, frequent 2-sequences, up to frequent k- sequences [11]. Figure 1. SPADE algorithm [11]. Extracting frequent 1-sequences is the first step in the SPADE algorithm after determining a minimum support. Frequent 1-sequences are all sequences that have length of 1 event and its support value greater than or equal to the minimum support. For the next step, all items that are frequent 1-sequences will be a parent class in the creating frequent k-sequences. In the process of frequent k-sequences generation, a k- sequence is obtained by combining a k-1 frequent sequence with another k-1 frequent sequence. Figure 2 shows the process to generate all possible sequences ([11]. Figure 2. Enumerate-Frequent-Seq function [11] 3.2 Association Rule Mining Association rule mining is a data mining technique that aims to find an interesting relationship between the items hidden in a large dataset. The interesting measures that are commonly used in association rule mining are as follows [3]: 1. is measured as the percentage of transactions in D that contains A B, denoted as. 2. is measured as the percentage of transactions in D containing A that also contain B, denoted as. 3. The lift (A,B) is calculated using the following formula: (1) Based on the value of lift(a,b) in Equation (1), the relation of occurrence of A and B is described as follows. If lift(a,b) is greater than 1, then A and B are positively correlated meaning that the occurrence of A implies the occurrence of B. If lift(a,b) is less than 1, then A and B are negatively correlated. A and B are independent if lift(a,b) is equal to 1. It means that there is no correlation between A and B. This work applied the Apriori algorithm to extract association between sequence of hotspots and weather condition in Sumatra Indonesia 2014 and The Apriori algorithm discovers frequent itemsets and association rules in a transactional dataset that have support and confidence greater than the user-specified minimum support (minsup) and minimum confidence (minconf) respectively. The Apriori algorithm is as follows [2]: 1) L 1 = {large 1-itemsets}; 2) for (k = 2; L k 1 ; k++) do begin 3) C k = apriori-gen(l k 1); // New candidates 4) forall transactions t D do begin 5) C t = subset(c k, t); // Candidates contained in t 6) forall candidates c C t do 7) c.count ++; 8) end 9) L k = {c C k c.count minsup} 10) end 11) Answer = kl k; Figure 3. Apriori algorithm [2] L k is a set of large k-itemsets in Figure 3. This set contains k-itemsets that its support values meet minimum support. C k is a set of candidate k-itemsets. Itemsets in this set are potentially to be large itemsets. In the Apriori algorithm, the apriori-gen function has the argument L k 1 i.e. the set of all large (k 1)-itemsets. The output of this function is a superset of the set of all large k-itemsets [2]. 4. Results and Discussion Sequential patterns of hotspots were obtained at the minimum support of 0.01 by utilizing the SPADE algorithm that is available in R package arulessequence. Number of 2-frequent sequences and 3-frequent sequences generated is 28 and 91 from the dataset of hotspot Sumatra 2014 and the dataset of

4 146 International Journal of Computing & Information Sciences Vol. 12, No. 1, December 2016 hotspot Sumatra 2015 respectively. Frequent sequences generated on the dataset of hotspot Sumatra 2014 are as follows (Figure 4). hotspots were occurred on October 24, 2015 (date code 297) and these occurrences were followed by other hotspots on October 26, 2015 (date code 299). This sequence patterns are found in 22 villages in Sumatera Islands. Our analysis focuses on the 2-frequent sequences and the 3-frequent sequences in which hotspots in the sequences are occurred consecutively in the duration of 2-5 days. Hotspots that appear in these sequences can be a strong indicator of peatland fires and we denote these patterns as fire spots. Figure 4. 2-frequent sequences and 3-frequent sequences of hotspot Sumatra 2014 The most frequent sequences in the hotspot dataset in Sumatera 2014 is <{70},{72}> with the support of (Figure 4). This sequence states that hotspots were occurred on March 11, 2014 (date code 70) and these occurrences were followed by other hotspots on March 13, 2014 (date code 72). This sequence patterns are found in 63 villages in Sumatera Islands including Sepahat in sub district of Bukit Batu, Bengkalis district, Riau Province. Ground verification was conducted on this site on June 12, Figure 5 shows burn area in 2014 in sub district of Bukit Batu, district of Bengkalis, Riau Province. Figure 5. Burn area in 2014 in sub district of Bukit Batu, district of Bengkalis, Riau Province (Photo taken on 13 June 13, 2015) Frequent sequences generated on the dataset of hotspot Sumatra 2015 are as follows (Figure 6). Figure 6. 2-frequent sequences and 3-frequent sequences of hotspot Sumatra 2015 The most frequent sequences in the hotspot dataset in Sumatera 2015 is <{297}, {299}> with the support of (Figure 6). This sequence states that Weather condition of the hotspot sequences are also identified in this study using the association rule mining approach. Number of association rules generated on the hotspot dataset Sumatra 2014 is 115 rules. These rules contain weather conditions that are coincided with the location of the hotspot sequences. The weather conditions are as follows rainfall in the interval mm, the average temperature between and C, and the average humidity between 70.1 and 8.2%. The strongest rule is {sid = SID1768 (Sepahat village in Riau Province)} => {rainfall = (0-0.9 mm)} that has support of , confidence of 0.9, and lift of Number of association rules generated on the hotspot dataset Sumatra 2015 is 146 rules. These rules contain weather conditions that are coincided with the location of the hotspot sequences. The weather conditions are as follows no rain (rainfall of 0 mm), the average temperature between and C, and the average humidity between 68.6 and 77.7%. The strongest rule is {sid=sid4928 (Ujung Tanjung village in South Sumatra province} => {no rain} that has support of , confidence of 1, and lift of Association rules verification will be conducted in the future works by comparing the patterns with the secondary data related to real peatland fires in the study area. Such data are collected by governmental institutions under the Ministry of Environment and Forestry, Republic of Indonesia. 5. Summary and Future Work This study has successfully discovered the sequence patterns of hotspot in Sumatra Indonesia 2014 and 2015 by applying the sequential pattern mining algorithm SPADE. The analysis focuses on the 2-frequent sequences and 3-frequent sequences in which the incidence of hotspots consecutively at the duration of 2-3 days can be a strong indicator of peatland fires. The sequence patterns were combined with weather data to identified association between hotspot sequences and weather conditions in the location where the sequences take place. The results show that the hotspot sequences in Sumatra in 2014 are

5 Mining Sequence Pattern on Hotspot Data to Identify Fire Spot in Peatland 147 frequently occurred in the following weather conditions: average humidity between 70.1% and 78.2%, average temperature between C and C, and precipitation of 0 and 0.9 mm. In addition, the hotspot sequences in Sumatra in 2015 are frequently occurred in the following weather conditions: average humidity between 68.6% and 77.7%, average temperature between C and C, and precipitation of 0 mm. In the future work, we will conduct hotspot sequences verification through field work. In addition, verification will also be performed by utilizing satellite image for burn area in Sumatera island Indonesia. References [1] T Agustina, IS Sitanggang. Sequential Patterns for Hotspot Occurrences Based Weather Data using Clospan Algorithm. The 3rd International Conference on Adaptive and Intelligent Agroindustry (ICAIA); 2015 August 3-4; Bogor, Indonesia (ID). pp doi: /ICAIA [2] R. Agrawal and R. Srikant, Fast algorithms for mining association rules, Proceedings of 20 th Int. Conf. Very Large Data Bases, VLDB, Eds. Morgan Kaufmann, pp , [3] J Han, M Kamber, J Pei. Data mining: Concepts and Techniques. San Francisco,USA: Morgan Kauffman Publisher, [4] J Han, J Pei, X Yan. Sequential pattern mining by pattern-growth: Principles and extensions. StudFuzz pp [5] J Han, X Yan, R Ashfar. Clospan: Mining closed sequential pattern in large dataset [6] NZ Nurulhaq, IS Sitanggang IS. Sequential pattern mining on hotspot data in Riau Province using the PrefixSpan algorithm. The 3rd International Conference on Adaptive and Intelligent Agroindustry (ICAIA); 2015 August 3-4; Bogor, Indonesia (ID). pp doi: /ICAIA [7] F Parish, A Sirin, D Charman, H Joosten, T Minayeva, M Silvius and L Stringer. (Eds). Assessment of Peatlands, Biodiversity and Climate Change: Main Report. Global Enviroment Centre, Kuala Lumpur and Wetlands International, Wageningen [8] J Pei, J Han, B Mortazawi-Asl, J Wang, H Pinto, Q Chen, U Dayal, M Hsu. Mining sequential patterns by pattern-growth: the prefixspan approach. IEEE Transactions on Knowledge and Data Engineering. 16(11): , [9] L Syaufina Kebakaran Hutan dan Lahan di Indonesia: Perilaku Api, Penyebab, dan Dampak Kebakaran. Bayumedia, Malang. (In Bahasa). [10] PF Viger. Sequential pattern mining framework (SPMF) versi 0.94 [Online] Available: com/spmf [11] MJ Zaki. SPADE: An Efficient Algorithm for Mining Frequent Sequences. Machine. Netherlands: Kluwer Academic Publisher. 42, 31-60, [12] Q Zhao, S.S Bhowmick. Sequential pattern mining: a survey. Technical Report, CAIS, Nanyang Technological University, Singapore, No Imas Sukaesih Sitanggang, received the PhD Degree in Computer Science from Faculty of Computer Science and Information Technology, Universiti Putra Malaysia in She is a lecturer in Computer Science Department, Bogor Agricultural University, Indonesia. Her main research interests include spatial data mining and spatial data processing. Erna Fatayati is a student at Computer Science Department, Faculty of Natural Science and Mathematics, Bogor Agricultural University, Indonesia.