Land use/land cover and land surface temperature analysis in Wayanad district, India, using satellite imagery

ABSTRACT This paper assesses Land Use/Land Cover (LULC) classification and Land Surface Temperature (LST) in Wayanad district during the years 2004 and 2018. The LULC classification of Wayanad district is identified using IRS P6 (Linear Imaging Self Scanner) LISS- III, and LST using thermal band of (Enhanced Thematic Mapper Plus) ETM+ imageries. Maximum likelihood classification (MLC) technique is opted to categorize six land-use features: water body, paddy field, forest, dense, agricultural crops and built-up. From 2004 to 2018, impacts of changes in features are correlated with the raised LST. Overall vegetation cover shows an increasing pattern during the study period. The water bodies in Wayanad district improved from 4.30 to 32.68 sq.km due to construction of two dams: Banasurasagar and Karappuzha. However, agricultural crops and paddy field area have decreased by 4.7% in last 14 years. Decreasing rate of agricultural crops can be directly linked to population growth, thereby developing various built-up zones for basic needs. Forest and dense vegetated cover area are increased nearly 2.3 and 3.0%, respectively, during the study period, while bamboo degradation has also been witnessed from 2008 to 2013. The built-up class shows growth from 1.48 to 5.69% of total land area during 2004 and 2018. LULC have noticeable influences on LST with a negative correlation between vegetation cover and LST with a decrease of 1.75oC. The study findings can help the local authorities to implement urban planning regulations for public awareness and policy makers for a sustainable planning and management in forthcoming years.


General
LULC changes mapping and its impacts at regional scales serve a wide range objective to mitigate and manage various disasters like landslides, global warming, urban flooding, etc., (Reis 2008). The change-detection study negatively affects the climatic patterns, hazard susceptibility, biodiversity loss, global and local socio-economic dynamics (Mas et al. 2004;Dwivedi, Sreenivas, and Ramana 2005;Zhao, Lin, and Warner 2004). Rapid conversion of various land covers to different land use patterns is observed globally (Lambin, Geist, and Lepers 2003). Landcover variations mainly depend on population growth of an area along with human intervention (Achmad et al. 2015), agricultural demands (Cammerer, Thieken, and Verburg 2013;Li, Zhou, and Ouyang 2013;Dale et al. 1997), natural calamities (Dubovyk, Sliuzas, and Flacke 2011), economic and urbanization development (Rimal et al. 2019;Khan et al. 2014), and other factors (Mustafa et al. 2018). With realization that global environment is strongly influenced by land surface, concerns about LULC emerged in research areas. Availability of land-use change statistics aids in the decision-making process for environmental planning and management (Prenzel 2004;Fan, Weng, and Wang 2007). Land cover conversion modifes surfacealbedo, which in turn increases energy exchanges between atmosphere and surface thus introducing an impact on local climate (Sagan, Toon, and Pollack 1979). Metropolises with diverse physical surfaces than the surrounding rural areas show impact on their microclimate (Cai, Du, and Xue 2011). There are numerous developmental activities and policy implementations by government revolving around the changes in urban development resulting into forest loss of nearly 40% over India during 1880and 1980(Flint 1994. Only a few studies are conducted on some Indian metropolitan cities like Delhi (Mallick, Kant, and Bharath 2008), Jaipur (Jalan and Sharma 2014) and Chennai (Amirtham and Devadas 2009), but fewer research works are carried out in Wayanad district of Kerala on this aspect. From 2008 onwards, a large-scale degradation of bamboo plantation has occurred in Wayanad district thus growing concerns in forest department. To keep the vulnerability of landslides (Kuriakose, Sankar, and Muraleedharan 2009) and increased migration of permanent residents to Wayanad district, the present study of land-cover change detection is engrossed at this juncture.
Global coverage of Land Surface Temperature (LST) can be derived from series of satellites like Landsat, Terra and Aqua with higher spatial resolutions (Zhu, Lu, and Jia 2013). The transformation of vegetated areas to build up (Mallick, Kant, and Bharath 2008) and conversion of wetland and marshland into agricultural cultivation land or bare waste land (Pal and Akoma 2009) plays vigorous role in LST increase. The other causative factors responsible for urbanization increase in correlation with climate change (Mustafa et al. 2018). The author cited a vibrant investigation of factors like topography of the area, zoning status, population density, employment opportunities, Euclidean distance from national and state highways, railways, commercial areas, religious monuments, etc., responsible to increase in urban sprawl. A hybrid model is used to predict LULC by combining logistic reasoning and cellular automata growth models for Ahmedabad city and concluded that the population increase contributes maximum to the LULC changes because of extra land-living parcel requirement (Mustafa et al. 2018). Also, an investigation on allocation of different built-up classes for the study area simplifies the study of the aforementioned factors' influence on various climatic factors like temperature, rainfall, air pollution, etc. (Verma et al. 2016;Rollet et al. 1998;Schneider 2012;Mohan et al. 2012;Loibl and Toetzer 2003). Study of the relation between LST and LULC changes helps to solve problems related to climatic change and analyse interactions between human and environment (Jha, Dutt, and Bawa 2000). LST values of all pixels with their thermal reflectance have correlation with neighbourhood pixel DN values (Song et al. 2014). Modelling current surface temperatures with respect to LULC plays a vital role in mitigating rapid climatic change. New policies and guidelines in urban or rural land-use planning can be adopted for reducing heating effect based on the results.
Remote sensing is proved to be a very useful advanced tool for LULC variation detection. Remotely sensed data and Geographical Information Systems (GIS) are considered as one of the effective responsive tools for urban climate studies (Deep and Saklani 2014). The use of satellite imageries creates lenience to fetch ground truth data of protected zones as well as tough terrains like Wayanad, India. The necessity of spatial data to obtain LST and LULC is driving investigators to identify pertinent satellite-based approach, which will overcome the above-said difficulties in surveying remote locations. Remotely sensed data has become one of significant applications for mapping change in LULC with varied applications (Lo and Choi 2004). This study for change detection is mainly aimed to assess the change that occurred in LULC in Wayanad district during 2004 and 2018, and establish its correlation with LST. The LULC map and LST map are developed and observed using Linear Imaging Self Scanner (LISS) III and Enhanced Thematic Mapper Plus (ETM+), respectively. The correlation of individual land-use features with LST change is also analysed in this study.

Site description
Study area (as shown in Figure 1) is selected as Wayanad district (11°44ʹ N-11°97ʹ N and 75°77ʹ E-76°43ʹ E), located in North-Eastern region of Kerala, India. The altitude of study area varies from 700 to 2100 m above mean sea level on the crest of the Western Ghats and covers an area of 2130 sq.km. As the Western Ghats is a part of UNESCO's world heritage site and world's biodiversity hotspot, a diverse range of flora and fauna can be observed in this green paradise. Nearly 40% of the total land area of district is under forest-protected zone (Sand 2016). The backbone of district's economy is mainly based on agriculture. Wayanad is famous for its rice and pepper production and the other agricultural crops produced in the district are coffee, tea, coconut, plantain and cardamom.
The increased population and landslides in this high range causes a large LULC variation in the stud area. A large area of the district is under forest-protected zone. About 25.6% forest cover loss is observed between 1973 and 1995 in the southern parts of the Western Ghats (Jha, Dutt, and Bawa 2000). Wayanad district is the only district of Kerala sharing boarder together with Karnataka (north and north-east direction) and Tamil Nadu (south-east direction) states of India. The annual mean rainfall of the district is 2322 mm and average temperature for last 5 years ranges from 18 to 29°C. The population of the Wayanad district is 8,17,420 with assorted locales of the two adjoining states (Sand 2016). The granite quarries in various parts of the districts convert most of the ecologically sensitive areas into landslide-prone areas. Usual landslide activities (Kuriakose, Sankar, and Muraleedharan 2009) often result in land change in the region. Rapid urban developmental activities, vast degradation of bamboo plantations and annual variation of weather conditions in the district are motivations to select the study area.

Satellite data used
In this study, remotely sensed satellite imageries are used for the LULC classification and the LST estimation of and Near Infrared (Band 4) are used for LULC classification and Thermal band (Band 6) of ETM+ is used for LST estimation. LISS-III satellite imageries used for LULC classification are purchased from National Remote Sensing Centre (NRSC), Hyderabad, and ETM+ imageries used for LST analysis are freely downloaded from the United States Geological Survey (USGS) website. The detailed description of datasets used in this study is presented in Table 1.
The origin of the study and selection of period of comparison is chosen to address two major aspects of land-use changes and its impacts on rising temperature in Wayanad district. Firstly, a pre-developed cloud free coherent satellite image is obtained to produce LULC map of specific study area that can be used as a benchmark digital data in future for LULC variation analysis. Next, cloud-free satellites images are essential for LULC mapping and LST analysis after necessary radiometric and geometric corrections. The presence of atmospheric clouds acts as barrier for satellite's image capturing payload. After considering all conditions, the satellite imageries from LISS-III sensor acquired on   are used for LST estimation. A preparatory version of ArcGIS 10.1 and Envisat is used for pre-processing the satellite images and Figure 2 shows a detailed methodology used in this study.

Image pre-processing
Satellite imageries are necessary to pre-process for more accurate and precise information. Initially, the satellite images obtained are geometrically corrected for LISS-III images using UTM coordinate system and radiometrically corrected for ETM+ images as a part of standard pre-processing procedure. The geometric correction can be achieved by choosing accurate Ground Control Points (GCP) on satellite imagery and on suitable geometric model (Thakkar et al. 2017). Georeferenced Landsat 7 ETM + images acquired from USGS Earth Explorer website after 31 May 2003 have traces of zig-zag lines over satellite ground path. This error is occurred due to failure occurred to malfunction of Scan Line Corrector (SLC): ETM+ sensor. Therefore, radiometric correction is performed on ETM+ (SLC failed) images to make it more useful. Radiometric correction is performed using Landsat toolbox developed for ArcGIS 10.1 (Foody 2002) to correct distortions occurred in ETM+ satellite imageries.
To classify the satellite images, supervised and unsupervised learning techniques have adopted grounded based on intricacy involved in the image classification . Review study on the image classification techniques presents all types of classification techniques and their inferences once equated to the validation dataset (Li et al. 2014). There are seven types (Maximum Likelihood Classifier, MLC) (Settle and Briggs 1987;Shalaby and Tateishi 2007), Naive Bayes Classifier (Minimum Distance-to-Means Classifier) (Atkinson and Lewis 2000), Mahalanobis Distance Classifier (Deer and Eklund 2003), Neural Networks (Kavzoglu and Mather 2003) Support Vector Machines (SVM) (Huang, Davis, and Townshend 2002;Pal and Mather 2005;Marconcini, Camps-Valls, andBruzzone 2009), Decision Trees (McIver andFriedl 2002;Jiang et al. 2012), Boosted Trees (Friedl and Brodley 1997), Random Forest (Gislason, Benediktsson, and Sveinsson 2006)) of supervised and two types (k-means (Rollet et al. 1998;Blanzieri and Melgani 2008) and ISO-data (Dhodhi et al. 1999) of unsupervised techniques available in Envisat and are most extensively used techniques. Unsupervised classification techniques work on pixel-based analysis in which pixels are grouped into families as defined by the user in the number of classes (minimum and maximum classes in case of ISO-data technique). K-means algorithm works on the principle of finding the scalar Euclidean distance between pixels and merging them into classes either as maximum iterative classes with minimal error or as defined by the operator (whichever is smaller). In supervised learning-based algorithms, the training datasets are selected in the order of 10 n (where n is the number of bands selected for image  classification). According to the review of literature on above-mentioned supervised technique, the MLC is chosen as the most effective classification technique (after image enhancement using principal component analysis) (Li et al. 2014).

Training dataset
In general, supervised and unsupervised classification techniques are adopted to classify various features in the area and determine their land characteristics. In the present study, we are applying the MLC (supervised classification) technique for better LULC mapping. Extraction of training data is one of the critical and time-consuming activities in the supervised classification process. Training sample point selection for any supervised classification technique is helpful in estimating mean vector and co-variance matrix aiding to calculate discriminant function for individual feature. The quantity and quality of input training dataset is the principal factor that primarily increases LULC accuracy level, i.e. higher the number and quality of training samples, higher will be the accuracy of LULC classification. So, training pixels of six classes are designated from the regions after inspecting reference map and post-field surveys. The characteristics of false colour image (obtained using various band combinations) for LULC classification is given in Table 2. The total number of training pixels must be higher enough for better and accurate LULC classification. Number of input training pixels for each classification must be kept minimum 30 times the total number of bands used for the study (Mather and Koch 2010). The MLC technique is adopted to classify six land-use classes by taking a minimum of 100 sample training points for each feature class (Schneider 2012).

LULC accuracy assessment
For the accuracy assessment, the classified results are compared with available reference datasets, which is assumed to be correct for defining a classification. Several methods are adopted to analyse user and overall accuracy of remote-sensed data (Aronica and Lanza 2005). LULC change accuracy is influenced by factors like sensor component associated issues and data pre-processing methods used with standard conditions at the image acquisition time (Morisette and Khorram 2000). In this study, we are using error matrix or confusion matrix technique (Foody 2002) for post-classification comparison method. Agreement and disagreement of pixels are generally compiled in the error matrix method. The rows of the error matrix table represent the reference data while the columns represent classified data (Morisette and Khorram 2000). For 2004 and 2018, the numbers of sampling points for postclassification validation of the six classes are 319 and 297, respectively, and using Equation 1, the percentage overall accuracy of each feature is calculated.
Overall Accuracy ¼ Total Number of Correct pixels Total Number of Pixels Ã 100 (1)

Land surface temperature analysis
All the satellite imageries comprise of pixel arrays that contain intensity value and location address which are digitalized and stored in the form of Digital Number (DN). The DNs are processed to convert into temperature scale using Equation 2. Image pixels are converted from DNs are rescaled to spectral radiance by using radiometric rescaling coefficient and are further converted to brightness temperature using thermal constants. When digital data are used for analysis, conversion of raw DNs to equivalent radiance or reflectance value for the comparison with other datasets is necessary (McFeeters 1996). The formulas used for conversion process are given below.

Conversion to radiance
The Scan Line Corrected (SLC) low-gain thermal band of ETM+ image is primarily converted from DN back to radiance scale (USGS LANDSAT 7 (L7) DATA USERS HANDBOOK Version 1.0 2018) using Equation (2).

Land Use/Land Cover (LULC) classification
Usually, LULC classes are chosen and fixed as per the requirements of specific study and application. In this study, we have classified the area into six LULC classes. Anderson's standard classification system (Anderson et al. 2017 From 2004 to 2018, the paddy field area is reduced from 352.32 to 219.01 sq.km, i.e. a loss of nearly 6.26% paddy cultivation area is observed. Currently, paddy cultivation is not at all profitable due to higher labour cost and low market rates. So, farmers are pushed to convert paddy field into other cash crops and food crops like banana, tapioca, ginger and yam cultivating areas.    During the site visit, we also observed ground water depletion in open wells near paddy field which are now converted into other agricultural crops land. Fallow paddy field can have a negative effect on ground water level (Anderson et al. 2017). Paddy fields near to roadsides are nowadays converted into built-up areas for economic benefits.
Mostly, agricultural crops are transforming into builtups due to increase in population. According to census data, 817,420 people reside in the study area and have 190,894 households (Nagar and Kawdiar 2018). Agricultural crop land areas are also decreased nearly 4.7% within 14 years, i.e. from 865.22 to 764.40 sq.km. Some agricultural crop lands are mainly cleared for cash crops like yam and ginger. Few agricultural crop land areas are misclassified under forest class due to the aged trees in agricultural cultivating lands, as a reason of which pixels show similar values of forest and dense class.
Better growth of vegetation is observed for forest and dense vegetated LULC classes between 2004 and 2018. There is an overall increase from 620.63 to 672.01 sq.km in forest class. Dense vegetated class shows an increase from 12 to 15% (255.56-320.26 sq.km). The strict rules and regulations are followed by government for recent years against human activities in protected areas. At present, both forest and dense LULC classes are not disturbed by human to certain extend; hence, the vegetation in those classes shows slight increase. The increase in forest area helps to restrict the urban development and provides additional benefits to restore the ecology of the area.
A dramatic increase of built-up class is experienced in the study area between 2004 and 2018. The built-up area is increased from 31.55 to 121.21 sq.km during the study period, i.e. an increase of 5.29% in 2018. Population increase is one of the driving factors for drastic growth in built-up area. The rocky areas are also classified under builtup class due to similar pixel values with normal built-up area. The operation of stone mining activities at various location of study area also increased the built-up class area. Approximately, 55 stone mining activities are working till 2016 (Nagar and Kawdiar 2018). The rural built-ups are growing well rather than urban built-ups in the study area. On analysing resultant LULC maps of both years, most of the roadways are mapped better in LULC map of 2018 due to clearance of vegetation near the roadways and wider road than 2004. The variation of LULC classes during 2004 and 2018 is illustrated in the Figure 5.

Vegetation indices
The NDVI enhances all vegetation and tends to have positive value. Soil may exhibit nearby zero value, and all water body features have negative values (McFeeters 1996). Only red band and NIR band of LISS-III images are used for NDVI analysis. In 2004, NDVI values ranges from −0.365079 to +0.763158; whereas for NDVI values during 2018, it ranges from −0.322581 to 0.875862, which means the vegetation cover has increased in 2018 when compared to 2004 in Wayanad district (Figures 6 and 7).
The NDWI is identified to monitor the presence of water bodies. NDWI is used to recheck the increase of water body area during 2004 and 2018. Like NDVI, water index scale also ranges from −1 to +1. The NDWI scale shows −0.4740 to +0.6266 and −0.7288 to +0.5312 during 2004 and 2018, respectively (Figures 8 and 9). The

Land surface temperature
The ETM+ sensor carried on Landsat-7 satellite data is used to study for the LST of Wayanad district between 2004 and 2018. LST is defined as the skin-temperature of earth surface, which is detected by satellite imageries (Zhu, Lu, and Jia 2013). To estimate LST, the most widely used remote-sensing technique based on thermal band is used (Jimenez-Munoz 2003). Since the LISS-III sensor is not having any thermal band, LST analysis using LISS-III satellite imagery is not possible. Hence, Thermal band (Band 6) of ETM+ sensor having wavelength 10.40-12.50 µm acquired both low and high gain data at 60 m resolution. These images are resampled into 30 m resolution using spline interpolation and then used in this study. The low-gain data is used when surface brightness is higher, while high-gain data is used if the surface brightness is lower. Image saturation is comparatively better for low-gain data. Moderate Resolution Imaging Spectro-radiometer (MODIS) sensor carried on Aqua and Terra satellites also have thermal bands (Bands 31 and 32), but resolution of both MODIS and LISS-III sensors is beyond comparison. Hence, we choose lowgain thermal band of ETM+ sensor carried on Landsat-7 with spatial resolution 30 m for the LST estimation.

LST vs. NDVI
Vegetative lands absorb an adequate amount of heat through transpiration and release low radiations leading to decrease in surface radiant temperatures. The modifications in topography of the area and its surface type exhibited noticeable difference in the LST. Discernibly high temperatures are observed during summer season due to changes in land-use characteristics of the area from 2004 to 2018. The mean surface temperature of the study area classified under paddy field dropped from 27.43 to 25.52°C during 2004-2018 periods. Paddy field classification shows the highest surface temperature among all other vegetation classifications. Since several paddy fields are converted into banana, ginger, tapioca and yam cultivations, certain region shows a rise in surface temperature of nearly 2.5°C. During paddy cultivation, water can stay on the land. Thus, conversion of paddy field to crop cultivation land reduces the moisture content of soil, which in turn increases the LST of those       However, a variation of 3.0°C can be observed in Mananthavady (11.8014°N, 76.0044°E) region (surface temperature is dropped from 23.9°C to 20.8°C). We have detected the similar variation of dense classification vs other classes in Mananthavady region. But a unique LST trend is noticed in and around Meenangadi (11.6596°N, 76.1726°E) panchayath region which shows almost an equal surface temperature value of 23.8°C in both 2004 and 2018. Meenangadi panchayath in Wayanad district is currently on a mission towards world's first carbon neutral panchayath by 2020 (Nagar and Kawdiar 2018).
Overall, 0.5°C rise in built-up surface temperature is experienced during the study period in Wayanad. Urban settlements in the study area like Kalpetta and Mananthavady show a decrease of 0.66°C in LST, while all other urban built-ups and rural built-ups show an increasing trend of approximately 1°C whereas temperature of rural town areas has a rise of 1.19°C. Hence, urban development in the study area and LULC classification in correspondence with LST variation carry up radiation from surface types like gravels, metallic and concrete. In Wayanad, rural towns are more developing rather than urban townships. The rural settlements show an average surface temperature increase of 0.29°C in the year 2018, i.e. an average of 24.97 and 25.26°C during 2004 and 2018, respectively. An increase of nearly 3.2°C in LST is noted at stone mining quarries in various locations of the district. Conversion of agricultural crops to rocky areas for the mining activities lead to increase of surface temperature from 24.7 to nearly 27.9°C. Similar to the study area, the other parts of Kerala viz., Kuttanand (Singh 2018), Alappuzha (Prasad and Ramesh 2019), Thiruvananthapuram (Arulbalaji and Maya 2019), Bharathapuzha basin (Raj, Azeez, and Use 2010), etc., have shown a prominent surge towards urbanization from the period of 2004-2014. Further, the increase in vegetative land and decrease in LST is observed due to varied agricultural practices. However, the studies on sprawling in urban core cities indicate the increasing LST values of surrounding dense built-up region (Shastri and Ghosh 2019). LULC classification and changes in temperatures due to increasing urbanization are also studied in other parts of India (Chennai (Amirtham andDevadas 2009),(David Sundersingh 1990;Devadas and Rose 2009), Delhi (Mallick, Kant, and Bharath 2008;Mohan et al. 2012;Kikon et al. 2016), Pune (Deosthali 2000), Hyderabad (Sundara Kumar, Udaya Bhaskar, and Padma Kumari 2017; Badarinath et al. 2005), Mumbai (Lei et al. 2008;Grover and Singh 2015), Bangalore (Ramachandra, Bharath, and Gupta 2018), Ahmedabad (Vyas, Shastri, and Joshi 2014), Lucknow (Singh, Kikon, and Verma 2017), Jaipur (Jalan and Sharma 2014;Chandra, Sharma, and Dubey 2018), Surat (Sharma, Ghosh, and Joshi 2013), Kochi (Thomas et al. 2014) and Nagpur (Agarwal, Sharma, and Taxak 2014)), producing a similar negative correlation in case of vegetative features vs. LST and positive in case of nonevaporating surfaces vs. temperature.

Conclusion
The temporal changes in LULC and LST in Wayanad, a district of Kerala from southern India during 2004 and 2018, are analysed in the present study. After the analysis, it is evident that overall vegetation cover increased to a certain extent in the study area from 2004 to 2018. Major LULC changes are observed in the water body and built-up classes. The increase in water body classification during 2018 is mainly due to the commissioning of two dams in the Wayanad during the mid of year 2004, i.e. after the capture of 2004 image used for this study. These changes can be directly or indirectly linked to human intervention and population growth. The infrastructural or other developmental activities due to human intervention in Wayanad district seriously depleted agricultural cultivatable land area. Along with the human intervention into agricultural crops, the paddy field near developing zones are also getting transformed into built-ups. This conversion process grabs the moisture content in topmost soil surface layer. It is clear from the LULC map of 2004 and 2018 that the built-up areas' developmental activities are occurring in rural areas rather than urban areas during 2018.
The LST analysis results from the overall study area show a decreasing pattern between the years 2004 and 2018. LST in paddy classified area shows the highest surface temperature among the vegetation classes. Conversion of paddy field into other agricultural cultivating land and barren lands increases surface temperature scale nearly 2.5°C. The increase in LST after conversion of paddy field is due to loss of moisture from surface soil. Dense shows lowest LST scale among the classes which ranges from 20.5 to 19.1°C. However, there is an increase in surface temperature of 1°C as observed over the builtup areas. Rural built-ups show an increased temperature range than urban built-ups. Overall, the vegetation has increased in certain areas. This is accompanied by a decrease in surface temperature of nearly 2°C in the study area. This is contradictory to what is expected, since there is also an increase in built-up areas associated with increased temperature of approximately 0.5°C. On taking Wayanad, there is an increase in vegetation associated with decrease in surface temperature.