Chlorophyll Changes Using Neural Network and Vegetation Indices in Tropical deciduous forest

Background: Ongoing climate and Earth’s atmosphere changes create profound effect on distribution and composition of forest, as well as on the fauna that depends on forest. The Sentinel-2A satellite data eases the mapping of Leaf Chlorophyll Content (LCC) at higher spatial and temporal resolution. In the present study, the temporal dimension of LCC was evaluated as an indicator of plant stress. LCC was retrieved using the inversion of the radiative transfer model based on an articial neural network. The data used for Spatio-temporal modelling of LCC was Landsat data. Result: From the Sentinel imagery derived vegetation indices, it was found that the narrowband indices having high correlation with LCC were pigment specic simple ratio and normalized difference index (45) (R 2 > 0.7; p < 0.001) centred at 665 nm, 705 nm, and 740 nm. Landsat 8 infrared percentage vegetation index had a strong relationship with LCC (R 2 =0.8). The Spatio-temporal (1997 to 2017) plant stress were detected using changes in LCC through an equation of correlation. The negative changes and deterioration of LCC were seen in the forest during the year 1997 to 20I7(rate = -1.2 µgcm -2 year -1 ) showing higher rate of forest health decline.


Introduction
Forests worldwide are becoming susceptible to an increasing size of abiotic and biotic disturbances [1][2][3][4]. The forest makes up high productive plants and resembles a signi cant carbon sink on the planet, which entails understanding of their spatiotemporal changes utmost important [5][6]. Forests are prone to degradation largely because of anthropogenic disturbances as well as climate changes [7][8]. With coupling with GIS, remote sensing technology proved to be a powerful tool for forest analysis [9].
The simulating forest photosynthetic activity is crucial to de ne tropical environmental controls and ecosystem metabolism [10][11]6]. Chlorophyll is a natural green particle in plant cells and plays a crucial part in the photosynthesis process. There are two classes of chlorophyll in plants, chlorophyll a and b which function as a photoreceptor in photosynthesis. The de ciency of these two chlorophyll contents may result in an unhealthy forest. Low chlorophyll content can limit primary production and hence, photosynthetic potential [12]. Chlorophyll is an indicator of physiological status, senescence, and stress [13][14]. The forest health condition changes need the understanding of chlorophyll content changes. Yang et al. 2017 [15] worked on the temperate forest by showing the seasonal variation of chlorophyll on ground based Normalized Difference Vegetation Index (NDVI) imagery. Pastor-Guzman et al.,2015 [16] performed on spatiotemporal mapping of chlorophyll in the Mangrove forest by using Landsat derived NDVI-Green. Zarco-Tejada et al.,2018 [17] analysed the temporal changes of red edge spectra, which is sensitive to chlorophyll for estimating canopy defoliation and pigment degradation. Croft et al., 2014 [18] analysed growing season temporal variation of leaf chlorophyll content in a temperate deciduous forest.
Satellite remote sensing has di culties in the estimation of chlorophyll beneath the canopy. Top of canopy re ectance is empirically related to chlorophyll concentration in the forest [19,15]. This principle applied to create radiative transfer models based on the arti cial neural network for estimating chlorophyll concentration. The original inspiration of this technique comes from the working of the central nervous system and with coordination of the neurons (and their dendrites, axons, and synapses), which make up signi cant information processing systems. In a neural network model, nodes ("neurons," units" or "processing elements") are coming together to make a network of nodes, therefore called "neural network. For operational applications, the inversion techniques are applied, which is based on a pretrained re ectance database. The model so develop is a mixture of PROSPECTS [20] and SAIL [21] called PROSAIL to see biophysical properties of the forest.
Remote sensing-based appraisal of biophysical characteristics of the forest is often getting through developing empirical relationships between eld-measured vegetation properties and vegetation indices [18]. While using the PROSAIL model and satellite-determined vegetation indices to evaluate the chlorophyll content of the leaf in different vegetation types, only few of studies focused on sub-tropical forest [22][23]. Given the signi cance of foliar pigments as surrogates of forest physiological status, phenology, health condition, and the gross primary productivity, it is crucial to evaluate the exactness of vegetation indices to foresee LCC.
Forest in the study area is affected by fragmentation [24] and connectivity weakening [25]. The forest of Jharkhand comprised 29% of its total geographical area, and it is prone to degradation due to many natural and anthropogenic factors including forest re hazard and drought, which lead to loss of green cover [26]. This study examines the usefulness of sentinel 2A imagery for retrieving leaf chlorophyll content in the forest. Our primary aim is to prove that the multispectral sensor Landsat 8 able to create maps of spatiotemporal variation of leaf chlorophyll focus in the sub-tropical forest. Change detection of chlorophyll content using combination of Sentinel 2A and Landsat is new in the research.

Data collection
The eld study allowed us to nd which forest cover types had what amount of leaf chlorophyll content.
The numerous portions of the study area transformed into open and scrub forest from dense forest (Fig. 2). The present study used Sentinel 2A data for retrieving LCC and Landsat data for Spatio temporal mapping of LCC (http://earthexplorer.usgs.gov) ( Table 1). Where, pρ is the DOS corrected image and Cos ∅ z is the cosine form of solar zenith angle and E 0 is the exoatmospheric solar spectral irradiance.
The Radiance (L λ ) was calculated through the below-given equation (Eq. 2) Where DN = Digital Number and the rel ectance (P λ ) was computed by given below equation (Landsat Science User Data Handbook) (Eq. 3). P λ = π * L λ * d 2 / ESUN λ *Cosθ (3) Where P and L are re ectance respectively and at-satellite spectral radiance, ESUN is mean solar exoatmospheric irradiances for a band, d is the Earth-Sun distance, and cos ø is the solar incidence angle.

For Sentinel 2 A data
The sentinel 2A undergoes various stages of atmospheric correction using SNAP software. The Sent2core plugin was used for transformation of Sentinel level 1 image to level 2 resembles conversion of digital number of images to surface re ectance. It helps to correct images from the presence of aerosol, ozone level distortion, type of locality, mid-latitude abnormality and topographical variations. The IdePix Sentinel MSI plugin was used for masking of land/water/ cloud in the imagery.

Computing chlorophyll content
The method of inversion of PROSAIL using SNAP software(step.esa.int/main/toolboxes/snap) is given in Fig. 3 and its illustration below.
Where X* is the normalized input, X the original value, Xmin and Xmax respectively the minimum and maximum values.
Every arti cial neural networking has the hidden layer and the output layer. The layer outlined by its number of neurons, which has biases, weight, and transfer function which is developed earlier and embedded in PROSAIL model.
The neurons have transfer function called tangent sigmoid transfer function (Eq. 5); While an output layer with the transfer function is linear (y = x). It is simply involving the application of reverse function used to enter normalization given in Eq. 6. Y = 0.5 (Y*+1) * (Ymax-Ymin) + Ymin (6) where Y* is normalized output value (LCC) issuance from neural network, and Ymin and Ymax were calculated over the neural network training dataset.

Page 6/25
The accuracy assessment of LCC is given in Fig. 4. It shown good Root-mean-square Deviation (RMSE) of 56.29 µg cm − 2 . In case of LCC, signi cant correlation coe cient of 0.87 seen using arti cial neural networking [28].

Vegetation indices
A vast number of vegetation indices have been used and evaluated for measuring plant physiological conditions using simulated and empirical data [29]. 9 vegetation indices were calculated using the Sentinel 2A and Landsat 8 OLI data sets, from which 7 were broadband indices while 2 vegetation indices were narrow-band indices ( Table 2). The chlorophyll concentration is well predicted in the wavelengths between 400 nm to 725 nm with multiple scattering in the Near Infrared (NIR) region. Therefore, vegetation indices covered these ranges are useful in modelling LCC [39]. connected to estimate biophysical variables [40]. In the present study, indices were taken because most of the biophysical measurements, for example chlorophyll estimation using remote sensing depend upon the spectral proprieties of vegetation by a semi-empirical linkage with vegetation index [41,42]. In other means the spectral vegetation indices are popularized for retrieving chlorophyll content from re ectance factors.

Modelling change detection
The forest health and stress were examined by LCC value, its decrease in value shows forest health decline and increment in plant stress. The Spatio-temporal changes of chlorophyll were computed using a grid statistics algorithm in SAGA GIS. Then, the scatter plot matrix was used to nd the relationship of LCC time to LCC time+1 using Geoda software (https://spatial.uchicago.edu). Time was the year taken for temporal variation of Chlorophyll.

Linear regression of LCC and sentinel derived vegetation indices
Leaf re ectance was used to derive the vegetation indices listed in Table 2. The spatial map of Sentinel derived vegetation indices is given in Fig. 5. At leaf-level, the relationships between the vegetation indices and LCC were computed. The narrowband vegetation indices with the most explanatory power (R 2 > 0.7) were Normalized difference vegetation index (45), and pigment speci c simple ratio (Fig. 6). When coe cients of determination computed, vegetation indices pursued a similar pattern showing an expansion in the level of measured variation in LCC.

Chlorophyll Concentration and Landsat 8 vegetation indices
One Landsat 8 image acquired in November 2017 were used to derive broadband vegetation indices (Table2) for comparison with LCC derived from sentinel 2A. The spatial map of Landsat derived vegetation indices is given in Fig. 7. The average LCC plotted against its corresponding pixel on the Landsat 8 vegetation indices and the coe cient of determination was computed. The correlation analyses showed that Landsat 8 infrared percentage index is the broadband vegetation index was most sensitive to LCC (R 2 = 0. 8) (Fig. 8). The linear model was used that produced this large correlation, described by Eq. 7, to construct an LCC map.

Variations in chlorophyll Spatio-temporal
Landsat 8 OLI satellite image acquired in November 2017 because vegetation vigour was high in the month. Using Eq. (1), LCC was mapped for every pixel based on Landsat data. A similar method was applied to Landsat images obtained at various dates to map LCC (Fig. 9). Maps can show the spatialtemporal distribution of LCC. Larger LCC values were seen at the core of the forest, the higher amount of chlorophyll presented in the very dense part of the woods.
The area where a high amount of chlorophyll was found has protected by the forest (layer of trees acts as a boundary itself, as they are free from outer disturbance. The low amount of chlorophyll was present in edges of the forest because they were in proximity to external interference from non -forest activities. Regarding the temporal variability, the maps depicted a decreasing gradient of LCC from November 1997 to November 2017 (Fig. 10). As far as anyone is concerned, this was the rst occasion when that Landsat 8 used to outline LCC of sub-tropical timberland at the temporal scale.

Modelling Change detection
The LCC value was in decreasing trend from 1997 to 2017(rate = -1.2µgcm − 2 year − 1 ) (Fig. 10). By the change analysis, overall, 33% of plants stress or forest health decline were predicted in the forest mainly because of pollution from crasher mine and other industrialization activities. The decreasing trend of variance from 1997 to 2017 showed large area of vegetation have low value of LCC in 2017 compared to 1992.
Scatter plot matrix of same variable from different base year is presented in Fig. 11, the base year was shown in x axis. From the regression analysis of two period taken from base year and independent variable (1997 to 2017), all shows signi cance changes (Fig. 11). The momentous changes occur between 1997 to 2017 having R 2 less than 0.6. Major changes occurred between the base year 2000 and independent variable 2013(R 2 = 0.60). Other change between base years and independent variable from 1997 to 2017 can be seen in the Fig. 11.

Discussion
The reason for infrared percentage index shows a strong relationship with LCC because the green edge acted like a red edge in the respect that both sides shift towards longer wavelengths at high LCC [43]. Therefore, in the study area, the low, dense forest having scrubs and grassland unable to show a good estimate of LCC. Thus, when LCC starts increasing, the green band becomes more sensitive, and the spectra of the red band become the lowest; that was the reason Landsat 8 infrared percentage index was best related to LCC. Although the LCC maps stood for a realistic spatiotemporal variation in LCC, many uncertainties were going with to them as it is constructed over assumptions and limitations. For estimating LCC of the forest the inversion of PROSAIL model was used. PROSAIL has been reversed over wide range of foliage canopies [43][44][45][46][47][48][49][50] with various levels of success. LAI is the prime part of measuring biophysical parameters in the forest [51][52]. Additionally, variation in leaf moisture, changes in soil moisture, changes in soil re ectance and differences in plant structure affect the linkage between Landsat 8 vegetation indices and LCC [29], particularly at LAI less than 3 [53]. Although this suspicion requires further examination, there are motivations to accept that LAI does not change much temporally. It is likewise critical to take note that the low LCC values was affected by uncertainty. Other uncertainties in the accuracy of estimation in the years 2013 and 2017 arise because of sensor characteristics changes. The results used a new age of satellites rises above the abilities of the Landsat in terms of spatial resolution, swath width, number of spectral bands, and revisit time [54].
Forest become less productive with such decline of LCC (33%) from declining of gross primary productivity [55] and affects carbon uptakes with overall biogeochemical cycling. In the study area, as the canopy density [24] were declining gives rise to open forest. Open forest has combination of new and old tress, having varying height. In the 2017 and 2013 the LCC value was minimum. These minimum values showed increment of diseases in plants and corresponding stress would rise mortality of trees.

Conclusion
The result showed the relationship between vegetation indices and the biophysical parameters, showing which multispectral indices best explained the variation on LCC. The ability of broadband and narrowband vegetation indices was evaluated to predict forest LCC at different scales. Indices with spectral bands around the red edge (705-753 nm), pigment speci c simple ratio and normalized difference index (45) indices were the most sensitive LCC (R2 > 0.7). A key nding was that the best performing Landsat 8 vegetation index was infrared percentage index, which explained 80% of the variation in LCC. The infrared percentage index better performed than NDVI for the retrieval LCC. Landsat satellite data retrieved vegetation indices was more sensitive to LCC than Sentinel derived vegetation indices.
The linear model portraying the connection between LCC used to delineate the Spatio-temporal variability of LCC in the forest scene. The examination exhibited that the medium-resolution and multispectral Landsat 8 used to measure LCC in tropical woodlands where ground systems and other potential devices were not applied, and the utilization of mapping procedures dependent on satellite information is essential. The LCC showed declining from 1997 to 2017 resembles forest health decline. The 33% of decreased of LCC in 2017 compared to 1997 revealed 33% increase in plant stress.
A practice use of this outcome was that future endeavours to determine LCC in a forest using multispectral remote sensing ought to consider the use of Landsat 8 infrared percentage index. The study likewise certi ed the importance of the red edge bands to anticipate LCC at the leaf level. Prescribed that future research should concentrate on testing existing and recently created algorithms to determine LCC in forest using the new age of satellites that beat the abilities of current sensors. The study received approval from the institutional review board of Central university of Jharkhand.

Consent for publication
The Author guarantees that the Contribution to the Work has not been previously published elsewhere.
Availability of supporting data Not applicable

Competing interests
No competing interest  The 'Norm' symbols indicating to the normalization process. Symbols 'S' and 'L' conform sigmoid (tansig) and linear transfer functions of the neurons respectively.