Handbook of HydroInformatics: Volume II: Advanced Machine Learning Techniques

By Saeid Eslamian

Advanced Machine Learning Techniques includes the theoretical foundations of modern machine learning, as well as advanced methods and frameworks used in modern machine learning. Handbook of HydroInformatics, Volume II: Advanced Machine Learning Techniques presents both the art of designing good learning algorithms, as well as the science of analyzing an algorithm's computational and statistical properties and performance guarantees. The global contributors cover theoretical foundational topics such as computational and statistical convergence rates, minimax estimation, and concentration of measure as well as advanced machine learning methods, such as nonparametric density estimation, nonparametric regression, and Bayesian estimation; additionally, advanced frameworks such as privacy, causality, and stochastic learning algorithms are also included. Lastly, the volume presents Cloud and Cluster Computing, Data Fusion Techniques, Empirical Orthogonal Functions and Teleconnection, Internet of Things, Kernel-Based Modeling, Large Eddy Simulation, Patter Recognition, Uncertainty-Based Resiliency Evaluation, and Volume-Based Inverse Mode.

This is an interdisciplinary book, and the audience includes postgraduates and early-career researchers interested in: Computer Science, Mathematical Science, Applied Science, Earth and Geoscience, Geography, Civil Engineering, Engineering, Water Science, Atmospheric Science, Social Science, Environment Science, Natural Resources, Chemical Engineering.

• Key insights from 24 contributors in the fields of data management research, climate change and resilience, insufficient data problem, etc.
• Offers applied examples and case studies in each chapter, providing the reader with real world scenarios for comparison.
• Defines both the designing of good learning algorithms, as well as the science of analyzing an algorithm's computational and statistical properties and performance guarantees.

• Language: English
• Publisher: Elsevier Science
• Release date: Dec 6, 2022
• ISBN: 9780128219508

Book preview

Preface

Advanced Machine Learning Techniques, Volume 2 of the Handbook of HydroInformatics series, includes 24 chapters containing the theoretical foundations of modern machine learning, as well as advanced methods and frameworks used in modern machine learning. It presents the art of designing good learning algorithms, as well as the science of analyzing an algorithm’s computational and statistical properties and performance guarantees.

The global contributors to this handbook cover theoretical foundational and advanced framework topics including: extremely randomized trees, random forest, group method of data handling, artificial neural networks (ANNs), outlier robust extreme learning machine (ORELM), AI-based model fusion approach, spatiotemporal variation of land use and land cover (LULC) data, trend analysis of extreme events, probable maximum precipitation (PMP) estimation, deep learning, long short-term memory (LSTM), hydrometeorological time series, dimensionality reduction by graphical modeling, eco-hydrological functioning, emotional artificial neural networks (EANNs), reference crop evapotranspiration machine learning approaches, exploring nature-based adaptation solutions, fuzzy-based large-scale teleconnection, hydrological model training and validation, geographic information systems (GIS), remote sensing, priority actions program/regional activity center (PAP/RAC) models, metrics of hydro-climatological performance, univariate and multivariate L-moments, copula functions, voting-based extreme learning machine (V-ELM), reference evapotranspiration methods of estimation, hybrid wavelet-random forest, wavelet-M5 models, resilience and disaster risk management effectiveness index, sequential Monte Carlo (SMC) methods, smart cities and hydroinformatics, support vector regression (SVR) models optimized with gray wolf optimization (GWO), and genetic algorithms (GAs).

This volume is a true interdisciplinary work, and the intended audience includes postgraduates and early-career researchers interested in computer science, mathematical science, applied science, Earth and geoscience, geography, civil engineering, engineering, water science, atmospheric science, social science, environment science, natural resources, and chemical engineering.

The Handbook of HydroInformatics corresponds to courses that could be taught at the following levels: undergraduate, postgraduate, research students, and short course programs. Typical course names of this type include: HydroInformatics, Soft Computing, Learning Machine Algorithms, Statistical Hydrology, Artificial Intelligence, Optimization, Advanced Engineering Statistics, Time Series, Stochastic Processes, Mathematical Modeling, Data Science, Data Mining, etc.

The three-volume Handbook of HydroInformatics is recommended not only for universities and colleges, but also for research centers, governmental departments, policy makers, engineering consultants, federal emergency management agencies, and related bodies.

Key features are as follows:

•Contains key insights from 24 contributors in the fields of data management research, climate change and resilience, insufficient data problems, etc.

•Offers applied examples and case studies in each chapter, providing the reader with real-world scenarios for comparison.

•Defines the designing of good learning algorithms, as well as the science of analyzing an algorithm’s computational and statistical properties and performance guarantees.

Saeid Eslamian, Department of Water Engineering, College of Agriculture, Isfahan University of Technology, Isfahan, Iran

Faezeh Eslamian, McGill University, Montreal, QC, Canada

Chapter 1: Analyzing spatiotemporal variation of land use and land cover data

Dinagarapandi Pandia; Saravanan Kothadaramana; Mohan Kuppusamya; Saeid Eslamianb,c    a School of Civil Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu, India

b Department of Water Engineering, College of Agriculture, Isfahan University of Technology, Isfahan, Iran

c Center of Excellence for Risk Management and Natural Hazards, Isfahan University of Technology, Isfahan, Iran

Abstract

Land Use and Land Cover (LULC) contributes to a substantial demand for hydrological studies. The LULC change is to determine the spatial deviation between 1996 and 2016 over the Chittar catchment. The Landsat data is used to develop the LULC classifications, which include residential area, active cropland, inactive farmland, forest, scrubland, without scrubland, and river/tanks. The significant deviation in a residential area, inactive cropland, without scrubland and river/tanks obtained from the spatial distribution of 1996–2016 LULC data. These LULC changes from 1996 to 2016 reveal the increasing trend of the residential area and inactive cropland; vice versa, scrubland and river/tanks are decreasing trends. Then, the image classification technique is applied for determining the vegetation distribution using Leaf area index (LAI) and Normalized Multiband Drought Index (NMDI) methods. Finally, Visual interpretation is compared to the image classification of LULC data and then, LULC classification methods are adopted in water resources applications.

Keywords

Chittar catchment; LULC; Landsat data; Spatial deviation; Image classification

Acknowledgment

First of all, Author need to thank the VIT University, Chennai, for allowing me to do the research work. Author gradually thankful for USGS Earth explorer and European space agency for providing the Landsat series of data and Sentinel 2 data.

1: Introduction

Land Use and Land Cover (LULC) changes are one of the critical parameters affecting watershed hydrology (Al-Fares, 2013; Garg et al., 2017) due to environmental hazards. There is a significant ecological threat to the conversion of cropland to urban areas due to urbanization. However, LULC changes occur with its LULC classes, where the crop plant rotates with its seasonal (Wagner et al., 2013). In semiarid, seasonal variations lead to scarcity of water, impacting seasonal cropland to uncultivated. This transition has an impact on social and economic development. Most studies concentrate on the effects of LULC shifts on water resources, although developing countries prefer to focus on the main research topics. Mainly, LULC Changes contribute to an interpretation of the morphology of the surface water body; thus, it is more successful in finding the availability of water as seasonal or annual (Fathian et al., 2016). LULC changes are analyzed for examining the effect of the hydrological response. LULC is used as a part of developing the physical based hydrological model (Garg et al., 2017).

Electromagnetic radiation (EMR) has different spectral wavelengths with different morphological characteristics. The EMR for optical remote sensing denotes the reflectance of morphological features. Morphology of LULC classes is usually interpreted with the optical remote sensing data in the range of visible near infra-red (VNIR) wavelength, i.e., 0.4–0.7 μm (Rogan and Chen, 2004; Coppin et al., 2004; Al-Fares, 2013; Hussain et al., 2014; Cheng and Han, 2016). Optical remote sensing is more appropriate for visual interpretation. LULC features are segregated with the help of visual interpretation. The multispectral bands can also interpret the LULC features; here, compare the image and visual interpretation. Additionally, image interpretation needs the Short wave infra-red (SWIR) wavelength, i.e., 0.8 to 2.1 μm. Image interpretation is a technical initiate for higher-level data analysis. Both the image and visual interpretation are used in this study and compared. LULC classification is to achieve maximum accuracy from high-resolution satellite imagery. Nowadays, Sentinel 2, Resourcesat-2 and Landsat 8 OLI are freely accessible for high-resolution earth observation (EO) satellites. These EO satellites collect reliable data information with their spectral reflectance. Spectral values define the earth’s features by their nature, followed by the spatially distributed signature. This spectrum varies from the launch vehicles; mainly, Landsat data has the multispectral bands likely red, green, blue, VNIR, SWIR. As per the temporal resolution, Landsat imagery offers the 8-day repeat cycle and starts from the early 1970s. By this advantage of Landsat data, most of the researchers used for board applications, namely earth sciences, water resources, coastal dynamics, environmental impact assessment, land surface process, etc. (Jha et al., 2000; Jayakumar and Arockiasamy, 2003; Dewan and Yamaguchi, 2009; Al-Fares, 2013; Mishra et al., 2019).

The review of Landsat is used in several studies in the LULC class transition. For a long term of 32 years, a transition of LULC classes defined an increase in a built-up area and caused the declines in the forest and agricultural areas at Hazira, Gujarat, due to industrial development and population pressure (Chauhan and Nayak, 2005). The visual interpretation of Landsat 5 and Sentinel 2 data are used to analyze the 30 years of LULC change for the Rani Khola watershed of the Sikkim Himalaya (Mishra et al., 2019). Similarly, forests in the Western Ghats, India, were noted the degraded between 1973 and 1995 (Jha et al., 2000). Subsequently, Eastern Ghats of South India were molested and converted the cropland to land with or without shrub from 1990 to 1999 (Jayakumar and Arockiasamy, 2003). The transition of LULC classes is to check with the ground truth data using kappa statistics (Dewan and Yamaguchi, 2009; Al-Fares, 2013). Kappa statistics are used to measure accuracy (Cohen, 1960; Congalton, 1991); LULC transformation used kappa statistics to analyze the variance of a matrix for the long term. Besides, Chittar watershed is an agricultural prone area due to the shift of rural to urban areas, resulting in a decreasing rainfall pattern (Dinagara, 2018). Sentinel 2 is one of the optical remote sensing used to release the finer resolution applied for LULC classification compared to other accessible data. Further, Leaf Area Index (LAI) also classified the LULC and that interpreted from the spatiotemporal satellite images like Landsat and MODIS data (Middinti et al., 2017; Blinn et al., 2019). Thus, LAI measures the indirect relationship of soil moisture and its main account to reflect the water susceptibility. LAI model was developed with various spectral vegetation index namely Normalized difference vegetation index (NDVI), Enhanced vegetation index (EVI), soil-adjusted vegetation index (SAVI), etc. (Middinti et al., 2017). Indexes are used the blue, red, NIR, SWIR multispectral bands and that collection of bands is available in Landsat data. LAI model was used explicitly for forest area where canopy cover is much more than other sites. Vegetation cover was usually used for reference evapotranspiration with their density classes (Cleugh et al., 2007) and reliable for the rainfall interception (Dietz et al., 2006). Normalized Multiband Drought Index (NMDI) is applicable to estimate the vegetation and soil moisture from the SWIR range from 1.6 to 2.35 (Wang and Qu, 2007). In this case study, the main factor is to analyze the LULC pattern for the Chittar watershed, Tamil Nadu, India. The visual interpretations are applied and calculate the changes in terms of kappa statistics. Data uncertainty is to understand for evaluation purposes. From the multispectral data, image interpretation is to analyze the vegetation classes with the help of LAI model. Finally, the LULC change examines the impact of the hydrological response.

The study area, i.e., Chittar watershed, lies in the part of the Tharamirabarani river basin, southern Tamil Nadu state, India (Fig. 1). The watershed lies in the global coordinates of 77° 9′E and 9° 12′ N to 77° 48′E to 8° 48′ N. Mostly, seasonal crops like paddy, maize, black gram, banjara, sorghum, etc. are cultivated along the river bed. So this watershed comprises of agricultural land. It covers about 78% of Chittar watershed. Chittar watershed has some of the towns like Sengottai, Tenkasi, Kadayanallur, Alankulam, etc. The study area comprises of 8 LULC classes within the 1660 sq km.

Fig. 1

Fig. 1 Geographical location of study area.

2: Data preparation

LULC is classified as per Level 1 to 3 of the NRSC Land Use and Land Cover Monitoring Division (2011–12). In Level 1, built-up land, i.e., residential area, agricultural land, wasteland, water bodies, and forest are used as LULC classes (Roy and Inamdar, 2019). In level 2, agricultural land is divided into crop, i.e., active cropland, plantation and fallow, i.e., inactive cropland; Wasteland is divided into land with shrub, i.e., scrubland and land without shrubland on the Chittar watershed. The crop is cultivated with the paddy; then banana plantation is observed in this Chittar watershed (Fig. 2). Talukdar and Pal (2019) were used the Normalized Differences Water Index (NDWI) and Modified Normalized Differences Water Index (MNDWI) images derived from Landsat data were studied to sustenance of ecological flow in riparian wetlands. NDVI, SAVI, and EVI are used to develop the LAI model for Indian tropical forests (Middinti et al., 2017). NDMI gives an admirable range for LAI values. Thus, LAI also releases the porous zone for hydrological components. Hydrological components are to analyze the impact of hydro-disaster prone areas. The limitation of the methodology is not able to understand the reliable validation with ground truth and comprise single LULC data. In this section, visual and image interpretation are applied into LULC classification.

Fig. 2

Fig. 2 (A) Cropland like paddy, maize, black gram, banana, etc., inactive cropland (fallow land); without Scrubland and with scrubland (B) Chittar river bridge and canal irrigation near Alaigaipuram.

The Landsat data are not feasible in this study, with 10% cloud-free for my study area. This uncertainty cannot be removed easily; some image techniques are needed to apply for cloud removal. Image techniques are not availed free; that kind of cloud removal is tedious to work. So, merely 10% of cloud-free Landsat data are downloaded for the month of the summer season. Uncertainty of Landsat is usually noticed during the postmonsoon season. Averagely, 50% of Cloud cover is more in postmonsoon Landsat data. Thus, inactive and active cropland is mentioned because lacking postmonsoon Landsat is not visually interpreted, long-term fallow land named inactive cropland. Cropland and scrubland have similarities to retain smooth texture to rough texture with the green patches named as scrubland; rectangular pattern with green spots called as cropland or long term brownish named as inactive cropland. The white spots with a square design called as a residential area. The coarse rectangular pattern named plantation; grayish patches or whitish irregular patterns named without scrubland. The hills regions at the upright of Chittar watershed called a forest, part of Western Ghats. Diamond or triangular shape pattern with smooth greenish texture or whitish brown color is named tanks; long linear pathway with the bluish color called a river.

3: Visual interpretations

Using the VNIR region of Landsat imagery, LULC is digitized in the geographical information system (GIS) environment employing physical interpretation (Mengistu and Salami, 2007; Areendran et al., 2013). Landsat TM, Landsat ETM+ (Enhanced thematic mapper) and Landsat 8 OLI (Operational land imager) imagery has 30 m of spatial resolution and VNIR region consists of blue (0.45–0.52 μm), green (0.52–0.60 μm), red (0.63–0.69 μm), Near-infrared (0.77–0.90 μm), and Short-wave Infrared (1.55–2.35 μm) multispectral released by Earth Observation Satellites. The object detection techniques such as image preprocessing, image segmentation, image filtering, image classifiers, etc. were surveyed (Cheng and Han, 2016). Object detection can be the knowledge-based image that classifies the land features using visual interpretation (Mishra et al., 2019). Visual interpretation is most widely preferred because it does not need the data calibration process in advanced GIS tools. The limitation of visual interpretation is not frequently accessible 10% cloud-free data and more time combustion for my study area.

Landsat TM and 8 OLI imagery are used to prepare for 1996, 2001, 2006, 2011, and 2016 (Table 1); visual interpretation is adopted to prepare by the image tone, texture, pattern, association, etc. In this study, National Remote Sensing Centre (NRSC) LULC classification is done using Landsat imagery. The features are active and inactive cropland, Plantation, Forest, Scrubland and without scrubland, Residential area, and Rivers/tanks prepared, as shown in Fig. 3. Forest can also be subdivided into Deciduous and Evergreen. Forest area is reserved by the government norms in the Chittar watershed. So, change in the forest area will be insensitive to this visual interpretation. LULC features, whereas residential area, rivers/tanks, cropland, scrubland, often change their areal distribution.

Table 1

Fig. 3

Fig. 3 LULC map of eight subclasses for the period of (A) 1996, (B) 2001, (C) 2006, (D) 2011, and (E) 2016.

4: LULC distribution

Forest is upright of Chittar watershed, where it receives the higher rainfall distribution flows toward the east to join the Thamirabarani river. Land features of agricultural land are more dominant, i.e., active cropland in this watershed from 1996 to 2016 (Fig. 4). A mid-portion of the watershed, conversion of land features like scrubland, without scrublands commonly appear from 1996 to 2016. Plantation occurred more in 2001 compared to other periods. From 1996, Inactive cropland converted from the plantation and active cropland. Then, Inactive cropland is slowly raising its distribution along the river bedside after 2011. Finally, 2016 of the river bedside appear the active cropland.

Fig. 4

Fig. 4 Change analysis of LULC subclasses at 10 years (A) 1996–2006, (B) 2001–11, (C) 2006–16, 15 years, (D) 1996–2011, (E) 2001–16 and 20 years, (F) 1996–2016. The X -axis is noted as 1—forest, 2—without scrubland, 3—scrubland, 4—active cropland, 5—plantation, 6—inactive cropland, 7—residential and 8—rivers/tanks.

Similarly, Residential area increase with the southwest of this watershed; conversion of rural to urban receives from the 2011 and 2016. Tanks are dry in the summer season and wet in postmonsoon, but the number of tanks reduced as in areal distribution. River line width is also decreased by significant encroachment through irrigation activities; the residential construction invades some tanks.

Sentinel 2 data with a resolution of 10 m started as free access from 2013 (Mishra et al., 2019). In this paper, 2021 data are used to classify the LULC for Chittar watershed. Finally, the uncertainty of Landsat data understands that the coarse resolution compares with the Sentinel 2 data. Thus, it also defines the limitation of visual interpretation over spatial resolution. However, kappa statistics are used to compare the old to present data; few ground truth data are needed for visual interpretation (Chen et al., 2018).

Sentinel data of 10 m resolution are preferable for reliable LULC generation (Mishra et al., 2019). Because finer resolution gives them more predictable to interpret land features (Table 2). Thus, LULC subclasses are classified into eight parts, as shown in Fig. 5. Increase trend of 2019 is in without scrubland, plantation, residential and river/tanks from 2016. The decreasing trend of 2021 is in the forest and active cropland established in 2016. The change of plantation, river/tanks and active cropland denotes the finer resolution of Sentinel data that helpful for the visual interpretation of LULC generation. Area contributions of LULC subclasses are prepared and interpreted with the increasing trend in residential and inactive cropland. River/tanks are decrease trend up to 2016. LULC subclasses likely without scrubland, scrubland, plantation and active cropland are the changing trend. The recent LULC maps of Chittar watershed revealed the changes in hydrological parameters. Main hydrological parameters impact the precipitation, evapotranspiration, surface runoff etc. So, the importance of LULC change has a more significant impact on hydrological modeling.

Table 2

Fig. 5

Fig. 5 2021 LULC from the optical Sentinel-2 data.

5: LULC change detection

As per 10, 15 and 20 years of LULC change between the periods of 1996 to 2016 changed for the eight subclasses of LULC (Fig. 4). In the 10 years, forest and active cropland are most similar to 1996–2016. Scrubland, without scrubland, inactive cropland and plantation are converted into active cropland. Some parts of active cropland converted into scrubland, without scrubland, inactive cropland and plantation. Next, the high impact of land conversion, scrubland, is converted into without scrubland. These 10 years, conversions of active cropland lead to sensitivity for agricultural practices and cause economic reliability for their states. Some parts of scrubland literally converted due to seasonal monsoon. More pumping of groundwater resources is used for cultivated crop plants, giving the hydrological impact. This pumping causes the depletion of groundwater resources. In the 15 years, the first sets of 15 years are more similar to 1996–2011, except residential area majorly converts from other lands as well as without scrubland into active cropland changes are equally distributed. Plantation converts into other lands. At second sets of 15 years, forest, active cropland, plantation and river/tanks are quite similar from 2001 to 2016. The residential area and inactive croplands mostly converted from other lands. These 15 years of LULC changes, conversion of a residential area, without scrubland and inactive croplands, leads to sensitivity in agricultural practices and rural to urban conversion. Rural to urban conversion occurred due to industrialization, population growth and construction technology development in their states. Simultaneously, agricultural practices need technology improvement for their irrigation activities. Irrigation activities prefer the canal or source of precipitation. Precipitation requirement is liable for 4 months (north-east monsoon) of rainy seasons. Thus, the Chittar watershed requires more water resources through groundwater or surface water. These resources attain economic stability for their states. In the 20 years, the second sets of 15 years are similar to 1996–2016 and show the same results. These show the same habitats that appear from 2011 to 2016.

In the lack of ground truth data, kappa statistics analyze the change from the old to the present data. Reference data are considered old data and classified data used from the present. For example, the LULC change analysis for 1996 and 2006 used 1996 as reference data and 2006 as classified data. As per Cohen (1960) stated the overall accuracy, producer’s accuracy and user’s accuracy also implemented for the LULC change analysis. The three kinds of intervals are taken using kappa statistics, whereas three sets in 10 years intervals, two sets of 15 years intervals and lastly 20 years intervals. The kappa statistics give the mark reference data deviate to the classified data. In this kappa statistics, the producer’s accuracy and user accuracy define the spatial change detection methods understand between the reference and classified data (Dewan and Yamaguchi, 2009; Al-Fares, 2013). A low kappa value means the high deviation, i.e., more LULC change; high value, i.e., minimal LULC change.

Kappa statistics is the spatial matching with the different sets of LULC data. Thus, LULC data are prepared for 1996 to 2016. In this case, LULC changes of 8 subclasses are to understand the spatial deviation with a minimum of 10 years and a maximum of 20 years LULC. Spatial deviation of LULC data using Kappa statistics is used to positive answers by terms of user accuracy, producer accuracy and overall accuracy (Table 3); similar kind of negative answers by terms of Omission error, commission error and overall error (Table 4). Positive answer for 10 year intervals gives the 0.65, 0.59 and 0.63; negative answers of 0.35, 0.41 and 0.37. So, the change occurred with the second set of 10 years of LULC data. Positive answer for 15-year intervals gives 0.65 and 0.60; negative answers of 0.35 and 0.40. So, the change occurred with the second set of 15 years of LULC data. Positive answer for 20-year intervals gives the 0.61; negative answers of 0.39. After 2001, LULC changes are developing in agricultural practices, population growth, industrialization, monsoon failure, etc., giving the ≥0.40 of negative answers. The landscape is more dynamic in nature; analyses the changes of LULC subclasses, as mentioned above. Changes in active cropland, without scrubland and residential areas, are continuously noticed; thus, there is a 90% chance of spatial deviation from the adjacent data over Chittar watershed.

Table 3

Note: UA, user accuracy; PA, producer accuracy; OA, overall accuracy.

Table 4

Note: OE, omission error; CE, commission error; OVE, overall error.

6: Image interpretation

The LULC data are prepared from the satellite data using image classification techniques such as supervised, unsupervised, fuzzy logic, subpixel classification, support vector machine, end member collection, etc. But, the image classification technique is allowed with trained algorithms, i.e., ground truth data gives better efficiency of LULC data through image interpretation. Here, image interpretation is the process applied for the postmonsoon period of Landsat data. The data cover 2001, 2011, and 2020 for the Chittar watershed. After converting the reflectance from Digital Number (DN) values, multispectral bands are ready to process the bands to estimate the NDVI, EVI, SAVI and NMDI (Table 5). So, vegetation indexes classify the LULC data from the satellite data.

Table 5

7: LAI model

LAI model is used multispectral bands to classify the vegetation cover, i.e., LULC classification (Middinti et al., 2017; Blinn et al., 2019). The reflectance of multispectral bands indicates the range of vegetation occurs over the region. Vegetation impels the maximum soil moisture with the fertility of nutrients content. However, vegetation emits the energy by the process of evapotranspiration, as well as rainfall interception might be possible in canopy cover. Broadleaf tree and dwarf tree includes in the vegetation cover. NDVI widely used to estimate the vegetation cover using this Landsat series of data. NDVI enhances the effect of both vegetation and soil properties. But, SAVI gives the very sensitivity in the soil backgrounds as compare to NDVI. Then, EVI can expose high biomass. Thereby, EVI is used to maximum recover the vegetation signal from the Red and NIR bands. In this case study, the LAI model is to develop for vegetation cover using spectral values. NDVI, EVI and SAVI are discussed briefly in Table 5. From the Landsat data, Blue, Red, NIR and SWIR bands are used as DN values. DN values convert into spectral reflectance. Red and NIR bands are used to estimate the NDVI and SAVI. But, EVI is used the blue, red and NIR bands. As eq. 22.1 said, the LAI model adds the NDVI, SAVI and EVI. The maximum value is taken as highly probable for vegetation cover. Vegetation cover, i.e., LULC classes is classified with five classes; water bodies, scrubland, forest, low vegetation and medium vegetation from the LAI model (left side Fig. 6).

Fig. 6

Fig. 6 Compare the LAI (left) and NMDI (right) for years 2001, 2011 and 2020.

By adding NDVI, SAVI and EVI, LAI figured out in the Chittar watershed (Fig. 6); water bodies of 2001 LAI are not visible but slowly increase after 2011. The thick forest in LAI is clearly observed in the upper watershed. Scrubland raises 2011 LAI. Low vegetation lies maximum in 2001 and slowly reduced 2011 and 2020 of LAI. Here, the postmonsoon period of Landsat data are collected and classified with their spectral reflectance. The study area fills with the water in the postmonsoon. So, the LAI model 2020 observed the water bodies.

si1_e    (1)

Additionally, NDMI helps us to compares the vegetation similarity, which implies the hydrological drought. NMDI is used the NIR and SWIR to estimate the soil moisture (right side Fig. 6). Thus, drought analysis is carried for this Chittar watershed; NMDI also included to analyze the water stress. Hence, LAI and NMDI are comparatively equal; this case study compares the temporal scale of 2001, 2011 and 2020. NMDI value is much similar to LAI value (Fig. 6). The distribution of NMDI gets down and maximum value occupies the distribution after 2011. Due to the moderate range of NMDI, the watershed is not much affected by the server drought-prone zone.

The LAI model and NMDI are classified as five land cover classes for 2001, 2011 and 2020. The LAI model is implemented to analyze the vegetation cover, which shows the land cover of the Chittar watershed. Respectively, NMDI is used to analyze the drought-prone zone. LAI of land cover class changes is calculated as a percentage shown in Table 6. One percent of Water in 2011 has compared with others. Low vegetation is comparatively reciprocal to the scrubland and thick vegetation. This model implies the more evapotranspiration reduce the soil moisture content in the Chittar watershed. Then, NMDI 2001 was equally distributed for five classes, higher than 15%. In 2011 and 2020, NMDI was distributed in the range of Moderate and High. NMDI was similar to their LAI. So, vegetation cover reduces and increases in the drought-prone zone.

Table 6

8: Compare the visual interpretation vs image interpretation

As per the study, console both visual and image interpretation. Thereby, LULC change detection is deeply mentioned by the visual interpretation in the GIS environment. Thus, it manually interprets the 8 LULC classes in the Chittar watershed. After the LULC changes happen for the eight classes are discussed. In the 8 LULC classes, the spatial distribution may occur manual human error by the visual interpretation. The limitation of visual interpretation satisfies with digital image processing. Image interpretation purely depends upon the spatial and spectral resolution. Data interpretation is discussed with the basic concepts applied in this case study (Table 7).

Table 7

9: Conclusions

Land use and Land cover are usually preferred for urban sprawl and its effects on population growth, industrialization, etc. Mostly, the Land cover is converted into Land use through anthropogenic activities. LULC subclass distribution reveals the change in the lower part of the river bed and the mid-portion of the watershed due to canal irrigation and monsoon drift. It causes an impact on the hydrological process. Here, LULC changes analysis studies preferring to impact of hydrological model development. Results of LULC data clarify the uncertainty of optical satellite data, i.e., Landsat. Spatial resolution is significant uncertainty of Landsat, so sentinel data are preferable for visual interpretation. Then, Kappa statistics is applicable to understand the LULC changes for 1996–2016. It gives 0.60 of maximum correlation and 0.40 error estimation for these LULC changes. Results of LULC changes from 1996 to 2016 happen the increasing trend of the residential area and vice versa decrease trend of river/tanks. Thus, identify the reduction of water resources over this Chittar watershed. Water resources that alias the sensitive hydrological parameters like less precipitation and more evapotranspiration. Sentinel data prepare the 2021 LULC data with the finer resolution. Then, 2021 LULC data are applied as hydrological responses for the physical based hydrological model.

LULC change using visual interpretation compares the image interpretation like LAI and NMDI using image processing of accessible Landsat data. LAI is used to classify the five LULC classes. LULC change analyzed with the 3 years of Landsat data. Finally, data conclude the increase of vegetation cover. The postmonsoon period (Mar to May) leads to high vegetation. But, soil moisture was reduced due to evapotranspiration except for 4 months (September to December) of precipitation. Seasonal crops are more active in the Chittar watershed. NMDI relates to the moderate drought-prone area, so it does not severely affect by the hydrological parameters. Thus, the watershed management plan should be negligible for improving water resources, and the benefits of water resources cause the significant development for agriculture activities, hydraulic power, etc.

References

Al-Fares W. Historical Land Use/Land Cover Classification Using Remote Sensing: A Case Study of the Euphrates River Basin in Syria. Heidelberg, Germany: Springer; 2013. Springer Briefs in Geography..

Areendran G., Rao P., Raj K., Mazumdar S., Puri K. Land use/land cover change dynamics analysis in mining areas of Singrauli district in Madhya Pradesh, India. Trop. Ecol. 2013;542:239–250.

Blinn C.E., House M.N., Wynne R.H., Thomas V.A., Fox T.R., Sumnall M. Landsat 8 based leaf area index estimation in loblolly pine plantations. Forests. 2019;103:222.

Chauhan H.B., Nayak S. Land use/land cover changes near Hazira region, Gujarat using remote sensing satellite data. J. Indian Soc. Remote Sens. 2005;333:413–420.

Chen Y., Sun K., Li D., Bai T., Li W. Improved relative radiometric normalization method of remote sensing images for change detection. J. Appl. Remote. Sens. 2018;124:045018.

Cheng G., Han J. A survey on object detection in optical remote sensing images. ISPRS J. Photogramm. Remote Sens. 2016;117:11–28.

Cleugh H.A., Leuning R., Mu Q., Running S.W. Regional evaporation estimates from flux tower and MODIS satellite data. Remote Sens. Environ. 2007;1063:285–304.

Cohen J. A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 1960;201:37–46.

Congalton R.G. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 1991;37:35–46.

Coppin P., Jonckheere I., Nackaerts K., Muys B., Lambin E. Digital change detection methods in ecosystem monitoring: a review. Int. J. Remote Sens. 2004;259:1565–1596.

Dewan A.M., Yamaguchi Y. Using remote sensing and GIS to detect and monitor land use and land cover change in Dhaka metropolitan of Bangladesh during 1960–2005. Environ. Monit. Assess. 2009;150(1–4):237.

Dietz J., Hölscher D., Leuschner C. Rainfall partitioning in relation to forest structure in differently managed montane forest stands in Central Sulawesi, Indonesia. For. Ecol. Manag. 2006;237(1–3):170–178.

Dinagara P.P. Monthly rainfall and temperature trend of Tirunelveli. Water Energy Int. 2018;61:65–71.

Fathian F., Dehghan Z., Eslamian S. Evaluating the impact of changes in land cover and climate variability on streamflow trends (case study: eastern subbasins of Lake Urmia, Iran). Int. J. Hydrol. Sci. Technol. 2016;6(1):1–26.

Garg V., Aggarwal S.P., Gupta P.K., Nikam B.R., Thakur P.K., Srivastav S.K., Kumar A.S. Assessment of land use land cover change impact on hydrological regime of a basin. Environ. Earth Sci. 2017;7618:635.

Huete A.R. A soil-adjusted vegetation index SAVI. Remote Sens. Environ. 1988;253:295–309.

Huete A.R., Liu H.Q., Batchily K.V., Van Leeuwen W.J.D.A. A comparison of vegetation indices over a global set of TM images for EOS-MODIS. Remote Sens. Environ. 1997;593:440–451.

Hussain M., Chen D., Cheng A., Wei H., Stanley D. Change detection from remotely sensed images: from pixel-based to object-based approaches. ISPRS J. Photogramm. Remote Sens. 2014;80:91–106.

Jayakumar S., Arockiasamy D.I. Land use/land cover mapping and change detection in part of Eastern Ghats of Tamil Nadu using remote sensing and GIS. J. Indian Soc. Remote Sens. 2003;314:251.

Jha C.S., Dutt C.B.S., Bawa K.S. Deforestation and land use changes in Western Ghats, India. Curr. Sci. 2000;79:231–238.

Mengistu D.A., Salami A.T. Application of remote sensing and GIS inland use/land cover mapping and change detection in a part of south western Nigeria. Afr. J. Environ. Sci. Technol. 2007;15:99–109.

Middinti S., Thumaty K.C., Gopalakrishnan R., Jha C.S., Thatiparthi B.R. Estimating the leaf area index in Indian tropical forests using Landsat-8 OLI data. Int. J. Remote Sens. 2017;3823:6769–6789.

Mishra P.K., Rai A., Rai S.C. Land use and land cover change detection using geospatial techniques in the Sikkim Himalaya, India. The Egyptian journal of remote sensing and space. Science. 2019;doi:10.1016/j.ejrs.2019.02.001.

Rogan J., Chen D. Remote sensing technology for mapping and monitoring land-cover and land-use change. Prog. Plan. 2004;614:301–325.

Rouse J.W., Haas R.H., Schell J.A., Deering D.W. Monitoring vegetation systems in the Great Plains with ERTS. NASA Special Publication; 309. 1974;vol. 351.

Roy A., Inamdar A.B. Multi-temporal land use land cover LULC change analysis of a dry semi-arid river basin in western India following a robust multi-sensor satellite image calibration strategy. Heliyon. 2019;54:e01478.

Talukdar S., Pal S. Effects of damming on the hydrological regime of Punarbhaba river basin wetlands. Ecol. Eng. 2019;135:61–74.

Wagner P.D., Kumar S., Schneider K. An assessment of land use change impacts on the water resources of the Mula and Mutha Rivers catchment upstream of Pune, India. Hydrol. Earth Syst. Sci. 2013;176:2233–2246.

Wang L., Qu J.J. NMDI: a normalized multi-band drought index for monitoring soil and vegetation moisture with satellite remote sensing. Geophys. Res. Lett. 2007;34(20):L20405.

Chapter 2: Artificial Intelligence-based model fusion approach in hydroclimatic studies

Vahid Nourania; Elnaz Sharghia; Nazanin Behfara; Fahreddin Sadikoglub; Saeid Eslamianc,d    a Center of Excellence in Hydroinformatics, Faculty of Civil Engineering, University of Tabriz, Tabriz, Iran

b Department of Electrical & Electronic Engineering, Faculty of Engineering, Near East University, Nicosia, Turkey

c Department of Water Engineering, College of Agriculture, Isfahan University of Technology, Isfahan, Iran

d Center of Excellence for Risk Management and Natural Hazards, Isfahan University of Technology, Isfahan, Iran

Abstract

Considerable literature has accumulated over the years regarding the combination of forecasts. The primary conclusion of this line of research is that forecast accuracy can be substantially improved by combining multiple individual forecasts. Furthermore, simple combination methods often work reasonably well relative to more complex combinations. This chapter aims to present model fusion concepts and some of its main approaches with the most focus on artificial intelligence methods. The chapter begins with background information focused on the following themes: Basic definitions and examples from the scientific literature are given to illustrate the process of building model fusion. Then the various methods and technics of model fusion are discussed. Furthermore, a few studies in this field with a brief explanation of their methodology and results are highlighted. Finally, suggestions for future works and challenges ahead are presented.

Keywords

Model fusion; Artificial intelligence; Machine learning; Model ensemble; Hybrid model; Hydroinformatics; Hydroclimate

1: Introduction

Since the hydrologic processes are highly nonlinear, dynamic, spatially distributed, and fragmented processes, they could not easily be described by simple models, and they involve several interlinked elements. Conceptual models require knowledge about the problem domain, whereas artificial intelligence (AI) models do not consider domain knowledge.

It is well known that a combination of many different predictors can improve predictions. Model fusion in machine learning has attracted great attention of the scientific community over the last years. Model fusion has been theoretically and empirically shown to provide significantly more accurate results than single learners in computational intelligence and soft computing, especially while dealing with high dimensional, complex regression and classification problems, nonstationary environments (Brazdil et al., 2009; Kazienko et al., 2013). Another main reason for the popularity of model fusion is the high complementary of its components. The integration of the basic technologies into fusion machine learning solutions facilitates more intelligent search and reasoning methods that match various domain knowledge with empirical data to solve advanced and complex problems (Sun and Wermter, 2000; Cios and Kurgan, 2002). Model fusion approaches can be done using various types of individual models but due to the nature of the black box models, it has more application in the field of black box models. So the most focus in this chapter is on AI methods; however, a few other methods are presented as examples. Also, model fusion can be used for time series prediction or forecasting, but the most focus is on prediction in this