Spatial Assessment of forest litterfall in Central Himalayas (India): Comparison of geospatial, remote sensing and data-driven estimates

doi:10.21203/rs.3.rs-3315721/v1

A comparison of multiple approaches for annual litterfall estimation and spatial assessment of forests was carried out for the state of Uttarakhand (Geog. Area = 54533 sq. km, Forest Area- 24652.32 sq. km) in Central Himalayas, India. Non-spatial approach used meta-analysis of published litterfall studies in Uttarakhand (29 studies with 115 measurements over sites/years) classified by forest types and area under forest types estimated by remote sensing by Forest Survey of India. The measured mean litterfall ranged from a high of 7.88 t/ha/yr for the sub-tropical broad- leaved forests to a low of 3.70 t/ha/yr in plantations. Spatial models of litterfall used a data-driven approach with 100 measurements and a random forest (RF) model that used bioclimate, elevation and forest type as covariates at a spatial grid of 1km resolution. This estimate was compared with published global (Li et al., 2019) and European (Neumann et al., 2018) spatial models. The total litterfall with five different forest-type area and estimated mean litterfall varied between 12.34 to 14.69 Mt/yr and with spatial allocation to forest type map estimated 14.02 Mt/yr litterfall. Data-driven spatial model using Random Forest approach estimated 13.305 Mt/yr of total litterfall. Use of spatial litterfall models developed for other study areas resulted in estimates that ranged from 9.11–15.81 Mt/yr. The study provides important insights towards developing a spatial gridded annual litterfall dataset for India and its use for studying the dynamics of forest carbon cycle.

Geospatial assessment of Litterfall flux

Random Forest

Forest Type wise Total Annual Litterfall

Data-driven litterfall estimates

Machine Learning

Forest litterfall is a critical flux that relates net primary production of a forest to carbon and nutrient cycling as well as litter and soil organic carbon pools (Vogt et al., 1986). Recent literature on litterfall has investigated role of plantation age, vegetation characteristics on total and leaf litterfall and resolved the seasonality of litterfall. Quantification of litter fall fluxes and carbon pools are essential to study the carbon cycle within the forests. In depth understating of carbon sequestration potential of forests and their contributions to net carbon emissions is a pre requisite. A spatial analysis could reveal the districts undergoing changes as well as district with higher potential for sequestration and support action for mitigation. These computations facilitate a better understanding of the productivity within the forest ecosystems. Remote sensing techniques can be coupled with biospheric and ecosphere models and can prove to be a promising method to investigate information at regional, national and global scales.

Recent methodological enhancements use improved geostatistical and machine learning methods on a wider set of geospatial covariates (Shen et al., 2019, Geng et al., 2022, Zhao et al., 2022) and use of remote sensing data in estimating litterfall (Wang et al., 2016, Hu et al., 2019, Shen et al., 2019). Recently there is spurt in studies reporting field measurements and their statistical models (Wen and He, 2016; Shen et al., 2017, You et al., 2017, Jia et al., 2018; Liu et al., 2019, Zhang et al., 2018), which however do not address the spatial variation at regional scale. Detailed observation on total and leaf litterfall have been reported on sample plots. Using individual studies, global and regional database of litterfall and associated observations have been compiled such as (a) Global – Bray and Gorham, 1964, Vogt et al., 1986, Zhang et al., 2014, Shen at el., 2019, Li et al., 2019, Holland et al., 2015 (A global data base of litterfall, litter and nutrients contains 1497 data records from 575 locations extracted from 685 publications), etc. and (b) Regional- Chinese forests (Zhao et al., 2022, Jia et al., 2020), European forests (Neumann et al., 2018) etc. Use of global data sets for regional variability and climate-based models & quantification of total litterfall flux have been critical in quantifying litterfall flux for different forest types (Mathews 1997) as well as to develop relation with climatic factors (Meentemeyer et al., 1982).

In the past two decades, there had been an increase in the number of publications on litterfall as well as simultaneous increase in the high-resolution forest datasets with the advancements in remote sensing, Global Positioning System (GPS) and Geographic information system (GIS) technologies in India. Preliminary literature survey for litterfall measurements in India indicated uneven coverage over forest types and regions. A recent global synthesis on litterfall (Holland et al., 2015) also under-reports Indian data. Indian studies as forest-wise summaries are provided by Dadhwal et al., 1997, Chhabra & Dadhwal 2004., and Ahirwal et al., 2021 with environmental correlates. In spite of publications of number of global syntheses, availability of larger set of Indian site-level data (listed in Supplementary Materials as Litterfall Studies in India.pdf) and improvement through multi-parameter use of remote sensing and data-machine learning methods (Fararoda et al., 2021; Mishra et al., 2020; Belgiu and Drăguţ, 2016; Breiman 2001; Srinet et al., 2019; Sreenivas et al., 2014), there have been no further reassessment of litterfall over Indian forests. Hence, there arises a need to revisit and predict the spatial and temporal variability of updated forest litterfall and the distribution maps using the climate data and the advanced geospatial and machine learning techniques. Machine learning has become a powerful tool for studying ecological parameters and modelling litterfall.

Litterfall estimates for Indian forests and their dynamics with respect to changes in forest cover, crown cover density and increase in plantation forests are critical to understand the forest carbon cycle changes. Methodology of meta-analysis of published studies and the spatial assessment using data-driven approach and Remote Sensing (RS)- derived inputs as demonstrated in this study can be adopted for nation-wide multiple time assessment. This study uses published forest type maps, remote sensing (RS), and machine-learning (ML) based techniques to capture the spatial variability in litterfall. Backward linkage of spatially gridded litterfall to forest fraction, and biomass and forward linkage to litter and soil organic pool would be critical in accurate assessment of changing terrestrial carbon cycle and monitoring the progress of India to committed national goals as part of Paris agreement 2015.

For developing a methodology and inter comparison amongst various approaches, study area is confined to the state of Uttarakhand in Central Himalayas as it had the largest concentration of published data on litterfall. Here, the use of RS & ML is explored for 1km resolution spatial mapping of litterfall for a study region of Uttarakhand. Also related to litterfall is observations on litterpool and SOC. Major Objectives of the study were to compare the estimates of annual litterfall for state of Uttarakhand from four methodological approaches (i) Meta-analysis of local litterfall measurements with remote sensing-derived forest type area, (ii) Total Annual Litterfall using four different methods. iii) To model and generate a gridded layer of Litterfall at a spatial resolution of 1km*1km using varying satellite data inputs and to compare them with other existing global models. iv) To apply a new data-driven approach (Random Forest Technique) to estimate the Total Annual Litterfall and generate a spatial map. Comparison of output was performed on both state-level total litterfall and scatter plot of spatial output from geospatial and data-driven approach.

Study area

In India, the largest published studies on litterfall have been conducted over the state of Uttarakhand (Fig. 1). Hence, the study aimed to focus the spatial assessment of litterfall over the Uttarakhand. Forests prove to be an integral part of the socio- economic and cultural life of people in Uttarakhand. The terrain is mostly hilly in this region, except in the regions of Bhabar and Tarai. Thick forest patches of Pine sps., Quercus sps., Shorea robusta, Rhododendron sps., Tectona grandis, Dalbergia sissoo, Cedrus deodara, Adina cordifolia, Acacia catechu and Terminalia bellirica tree species occupy the area at varying latitudes.

Forest type groups identified by Champion and Seth (1968) in Uttarakhand include 44 types, namely, Dry deciduous (DD), Moist Deciduous (MD), Sub tropical Broad leaved (STB), Sub Tropical Conifer (STC), Montane Temperate Broadleaved (MTB), Montane Temperate Conifer (MTC), Sub Alpine Forests (SAF), and Plantations (Pltn). Recent remote sensing-based forest and vegetation mapping studies include Roy et al., (2016) and Reddy et al., (2015). The elevations play an important role in determining the forest type distribution.

This study adopts a meta-analysis approach to derive non-spatial estimate of forest litterfall (LF-NS) and adopts four different approaches for developing spatial forest litterfall assessment. The first approach to spatialization integrates EO-based forest type and vegetation category maps (LF-SFTM). The second approach used published geographic and climate-based models for mapping litterfall (LF-SCLIM). The third approach applied machine-learning to relate published litterfall to forest type, location, climate, physiography and remote sensing derived indices (LF-SML). The fourth approach evaluated EO based indices and geophysical variables to develop new models or apply some published equations (LF-SEO).

Majority of the previous studies on litterfall were concentrated over the region of Uttarakhand. A total of 115 total litterfall measurements reported in literature were used as inputs for modelling. A valuable database of these observations was created along with additional information like altitude, tree density, basal cover, age, total annual litterfall, leaf litterfall, litter, soil organic carbon, leaf biomass and above ground biomass, depending upon the availability of data for each site. Spatial assessment also requires the geolocation of each study site. When not specifically indicated in the publication it was estimated from site description. These geolocations were also confirmed by visualizing them on Google Earth Pro.

Method 1: Litterfall estimates from Literature Survey

Survey of published literature on total litterfall (TLF, Mgha-1a-1) measurements over sites in Central Himalaya (State of Uttarakhand) was carried out. Studies which had at least one year of data were selected. The additional information extracted from publication included geographic location (latitude, longitude), forest/vegetation type. When geographic locations were not explicitly provided, they were estimated from map of study area or name of village/site when map was not included in publication. Due to adoption of different classification schemes by authors, the standardization of forest/vegetation class was carried out to equivalent class adopted in this study (Champion and Seth, 1968, FSI 2021, Sudhakar Reddy et al., 2015). A list of publications with information regarding geolocation and forest class is provided in Supplementary material.

Litterfall- Litterpool Relation

Forest Survey of India as part of State of Forest Report and Carbon stock estimation reported carbon pools by forest type/class for the year 2021. The pools reported are aboveground biomass, litter, coarse woody debris (CWD) and soil organic carbon (SOC). Mean litterflux averaged over forest type was related to FSI litterpool.

Relation to Previous Estimates on Litterfall

Chhabra and Dadhwal, (2004), had estimated the total annual litterfall for the Indian forests. An attempt was made to compare these values with that of the results from the present study.

Litterfall- SOC Spatial Maps

Sreenivas et al. have mapped SOC pool for India at high resolution and reported the SOC at 2 depth class at average resolution/grid size of 5×5 km. The spatially modelled litterflux was also related to SOC of Sreenivas et al., 2016.

Method 2: The Mean Estimates of Litterfall by Vegetation Type (Adopted from FSI)

Forest Survey of India has classified the Uttarakhand vegetation types into 44 classes according to the Champion and Seth Classification. In the present study, these classes were regrouped and modified into eight major vegetation classes. Litterfall Flux was computed for each of the classes using the mean values of litterfall assessed during the present study. An attempt was done to compare the Classified outputs (based on forest type) from various sources by determining the Litterfall for each forest type. Classified outputs from various sources were gathered and were used to estimate Litterfall Flux using the mean values of total annual litterfall from meta-analysis. The spatial layers of forests and forest types were as follows: Copernicus Global Land Cover Layers (2017) (Buchhorn et al., 2020), Classified Vegetation Map of India from the National Carbon Project (Reddy et al., 2015), Land Use Land Cover map from ORNL website (P.S. Roy et al., 2016). Computed mean litterfall for each forest type was assigned as attribute to corresponding forest type polygons. Resulting litterfall GIS layer was converted to 1km raster grid for further analysis as described in Section 3.3

Method 3: Spatial assessment of litterfall (LF-SFTM)

National level grids of 5km* 5km, adopted from Reddy et al., (2015) were used to generate fully aligned spatial grids of 1km*1km and further analysis was restricted to 54533 grid cells of Uttarakhand by assigning an attribute of area under forest in each grid.The number of grids which contained forests was 43881 and the total forest area was 2.529 Mha. In each 1km grid the area under 8 regrouped forest type was available as an attribute. Spatial models of litterfall developed for Europe by Neumann et al., (2018) with temperature, precipitation and LAI was applied to Uttarakhand grids for inter-comparison (Table 1). Data sources and their characteristics are given in Table 2.

Table 1

Litterfall spatial assessment models evaluated for Uttarakhand
Phenomenon	Equations Used	Units	Source
Foliar LF (Dry Foliar Litterfall Rate) =	Exp(2.241 + 0.650log(T + 10) + 0.211 log (P))	g/sq. m/yr	Liu et al., 2004
Foliar Carbon=	LAI/8.2*0.260	kg/ sq. m	Pietsch et al., 2005
Annual Total Litterfall=	125.562 + 19.888LAI + 12.160T + 0.115P	g/sq. m/yr	Neumann et al., 2018
Foliar LF Carbon=‍	NPP*0.34	kg/ sq. m/yr	Malhi et al., 2011

Table 2

List of Input layers used and their major characteristics
Data Products	Year	Source	Resolution	Reference
Leaf Area Index (LAI)	2017	MODIS	500 m	ee.ImageCollection ("MODIS/061/MOD15A2H")
‍Net Primary Productivity (NPP)	2017	MODIS	500 m	ee.ImageCollection ("MODIS/061/MOD17A3HGF")
Annual Precipitation	Climate	WorldClim	1 km	Fick and Hijmans, 2017
Annual Mean Temperature	Climate	WorldClim	1 km	Fick and Hijmans, 2017

Method 4: Modelling Litterfall through Machine Learning (LF- SML)

The bioclimatic variables represent the average values for the 1970–2000 period and are derived from monthly temperature and rainfall values (Fick and Hijmans, 2017). The data was downloaded from WorldClim website (www.worldclim.org/data/worldclim21) at 30 second resolution. The topographic variables used were elevation, slope and aspect (all derived from SRTM data (Farr et al., 2007). All the variables in the study were processed at 1 km spatial resolution. For model training and testing, annual litter fall values reported at ground locations over the entire study region (retrieved from literature) were used. Since data was processed at 1 km resolution, when more than one field point was lying inside a single 1km by 1km grid, mean of the annual litter fall values for these points was taken, finally resulting in 100 ground truth locations.

Result 1: Meta-analysis of site-level litterfall measurements in Uttarakhand

Literature search identified 29 publications which had reported total litterfall for 128 sites in Uttarakhand while TLF were reported from only 115 sites. Most of the measurements are of recent period as 25 of the 29 publications are of 2000–2020. Based on location and information provided the study site forests were classified adopting a truncated Champion and Seth (1968) classification into 6 major forest types, namely, Montane Temperate Forest, Montane Conifer forests, Subtropical Broad leaved/ Mixed forests, Subtropical Conifer forests, Tropical moist Deciduous Forests and Plantations. The range and mean of TLF is given in Table 3. The total litterfall over the region varied from 3.697 t/ha/yr in plantation to 6.248 t/ha/yr in the montane temperate forests and the highest being observed in the subtropical broadleaved forests (7.879 t/ha/yr).

Table 3

Summary on total leaf litterfall in Uttarakhand forests
Vegetation Types	Total Litterfall (t/ha/yr)
	n	Mean	Std. Err	Range
MTF Montane Temperate Forest	20	6.248	1.270	1.71–9.4
MCF Montane Conifer	12	4.650	0.808	0.26–8.95
STMF Subtropical Broad leaved/ Mixed	28	7.879	1.118	0.97–29.06
STCF Subtropical Conifer	20	6.221	0.440	3.51–9.67
TMDF Tropical moist Deciduous Forests	15	5.325	0.558	1.25–8.94
TDDF Tropical Dry Deciduous Forest*	55	5.196	0.508	0.10-18.42
Plantations	21	3.697	0.421	0.40–7.24
* TDDF based on pan-India studies, n = No. of Observations

Litterfall to Litter Ratio

A Vegetation Type-wise assessment of Litterfall (t/ha) over Uttarakhand from the present study was compared with the Litter pool (t/ha) assessment of FSI 2021. The Ratio of litterfall to litter was computed for each vegetation type and are as summarized in Table 4. The highest litterfall to litter ratio was observed in the Montane conifer vegetation followed by the Tropical Moist Deciduous forests. The lowest was observed among the Montane Temperate forests. On an average, the litter to litterfall ratio over the region was assessed to be 0.272.

Table 4

Litterfall to litter ratio
Vegetation Types	Litterfall (t/ha) (Present Study)	Litter (t/ha) (FSI, 2021)	Ratio (Litter/Litterfall)
Montane Temperate Forest	6.25	1.1	0.176
Montane Conifer	4.65	2.01	0.432
Subtropical Broad leaved/ Mixed	7.88	1.57	0.199
Subtropical Conifer	6.22	1.57	0.252
Tropical moist Deciduous Forests	5.33	2.13	0.400
Tropical Dry Deciduous Forest	5.2	1.04	0.200
Plantations	3.7	0.9	0.243
Average			0.272

A comparison with the Previous estimates

The Total Annual Litterfall as reported from the present study was 14.015 Mt/yr. When the total litterfall was recomputed using the mean litterfall values of Chhabra and Dadhwal, (2004), a total annual litterfall of 17.729 Mt/yr was yielded. Hence an approximate reduction of nearly 20% is observed through the years (Table 5).

Table 5

A comparative study with the previous assessment
Vegetation Types	Litterfall (Present Study)	Litterfall (2004)	Forest Area	Total Annual Litterfall
	(t/ha/yr)	(t/ha/yr)	(ha)	(Mt/yr)
Montane Temperate Forest	6.55	5.66	547750	3.100
Montane Conifer	4.65	6.7	37943.4	0.254
Subtropical Broad leaved/ Mixed	7.88	8.93	54745.2	0.489
Subtropical Conifer	6.22	6.7	628224	4.209
Tropical moist Deciduous Forests	5.33	8.93	979721	8.749
Tropical Dry Deciduous Forest	5.2	5.97	106769	0.637
Plantations*	3.7	3.7	78650.3	0.291
Total	17.729
*From present study (due to absence of observations in the previous study)

Correlation of litterfall with Soil Organic Carbon (SOC)

The soil organic carbon data, which was generated under the ISRO-GBP project through spatial modeling at 5 km x 5km resolution was downloaded from the website of Bhuvan, NRSC (https://bhuvan-app3.nrsc.gov.in/data/download/index.php?c=p&s=NI&g=all) and was further analysed (NRSC, 2016). The NICES (National Information system for Climate and Environment Studies) 5km equal area grids provided the mean values of SOC. A grid wise correlation between the total annual litterfall (n = 122) and SOC yielded a coefficient of determination of 0.35 (Fig. 2).

Estimates of total forest litterfall for Uttarakhand obtained from widely different approaches and data inputs converged to median estimate of 14.09 MTyr-1 with range of 9.112 Mtyr-1-15.814MTyr-1. Since estimates from non-spatial approach do not capture spatial variability of litterfall, results from data-driven and remote sensing derived geophysical models were compared. Spatial assessment of forest litterfall with remote sensing derived maps and biophysical inputs, spatial climate data sets.

Result 2: Total Litterfall for Uttarakhand (Adopted from FSI)

State-level total forest litterfall was estimated by using five sources of area under various forest types. These include tabular results of Forest Survey of India 2021 assessment and three spatial vegetation/forest type mapping studies (Roy et al., 2016, Reddy et al., 2015) and one global land cover assessment (Copernicus Global Land Cover Layers, 2017). These studies had large variation in number of vegetation/forest types covered, spatial resolution of input RS data (30m – 100m), covered different period. It is not surprising that the total forest area although different was within 10 percent (2.26 Mha – 2.49 Mha). The estimated total annual litterfall was 12.34–14.69 Mta-1. Uncertainty of this litterafall ranged from 0.90–4.87 percent. This approach of forest type integration with mean litterfall cannot capture the spatial variability in litterfall within any class. The two additional techniques, namely data-driven approach and spatial model-driven approach were explored to characterize spatial variability. The number of forest vegetation classes pertaining to each classification along with the area estimates for each class is summarized in Table 6.

Table 6

Estimated Total litterfall using multiple sources of forest-type spatial layers
Year of RS Data	RS Data (Spatial Resolution)	No. Of Forest Type Categories	Forest Area (State Total) ‘000 ha	Total Litterfall (Mt/yr)
2005	ETM+, LISS III (100m)	4	2489.1	14.66 ± 4.68
2013	AWiFs (56m)	18	2465.2	13.64 ± 1.41
2017	CGLS – LC 100 (.ca 1000m)	4	2263.9	13.12 ± 4.87
2020	Non-spatial	6	2464.2	14.69 ± 3.28
2021	LISS-II/ III Visual Interpretation	28	2487.8	12.34 ± 0.90

Result 3: Modelling Total Annual Litterfall using spatial grids of 1km*1km resolution

A summation of the product of each vegetation type to its average litterfall provided the Total Annual Litterfall for each grid. Hence a grid-level spatial layer of forest area (Fig. 3) as well as Litterfall was generated (Fig. 4). Type wise estimates of total annual litterfall is depicted in Table 7. Here the total annual litterfall over the region was estimated to be 14.015 million tonnes per year. A district- wise analysis of the state of Uttarakhand (Fig. 5) revealed that the highest production of littrefall is recorded from the district of Garhwal (2.14 million tonnes) followed by Uttarkashi, Nainital, Chamoli and Tehri Garhwal respectively. Litterfall in the districts of Pithoragarh, Almora and Dehradun were comparable. The least litterfall was observed in the district of Haridwar.

Table 7

Forest Type- wise estimates of Total Annual Litterfall
Forest Types	Total Annual Litterfall (Million tonnes)
Dry deciduous (DD)	0.554
Moist Deciduous (MD)	5.208
Sub-tropical Broad leaved (STB)	0.431
Sub-Tropical Conifer (STC)	3.551
Montane Temperate Broadleaved (MTB)	3.520
Montane Temperate Conifer (MTC)	0.244
Sub Alpine Forests (SAF)	0.270
Plantations (Pltn)	0.237
Uttarakhand Total	14.015

Comparative study with other models

Studies on meta-analysis of litterfall at global scale have explored the control of spatial variation in litterfall due to climatic and geographic factors. These include temperature and rainfall. In this study recent models of Neumann et al., 2018, Liu et al., 2004, Pietsch et al., 2005; White et al., 2000, which use parameters like temperature, precipitation, Leaf Area Index and Specific leaf area, were applied to spatial layers for Uttarakhand derived from climate and phenological data. In order to capture spatial variability, a 1km resolution, highest grid at which standard climatic datasets are available was used. Results from use of these models are provided in Figs. 6–9. In order to estimate the total litterfall at state/district level, model output of litterfall per hectare was weighted by forest fraction in each 1km grid. These results are depicted in Table 5.

Total Annual Litterfall when computed as a function of Leaf Area Index (LAI), Annual Mean Temperature and Precipitation (Neumann et al., 2018), resulted in a grid -wise spatial layer (Fig. 6). Total estimated litterfall over the entire study area was 14.09 million tonnes, which was exactly comparable with our modelled values of Total annual litterfall (14.315 million tonnes per year). The foliar Carbon was estimated as a function of LAI and Specific Leaf Area (Pietsch et al., 2005; White et al., 2000). The litterfall flux was calculated from the resultant as a product of foliar carbon and a fraction of 0.45. The total annual litterfall thus assessed was 15.814 million tonnes over the entire region. The litterfall flux is spatially depicted in Fig. 7. Foliar litterfall when computed as a function of temperature and precipitation (Liu et al., 2004) gave a modelled output with total annual litterfall of 9.112 million tonnes (Fig. 8).When the annual litterfall was estimated using the net primary productivity (NPP) as a proxy measure of biomass abundance (Malhi et al., 2011) and a fraction of 0.34 (litterfall fraction as suggested by Malhi), the resultant litterfall accounted to 11.179 million tonnes (Fig. 9). This proves the positive correlation of litterfall on NPP, irrespective of the tree species. The rest of the models generated outputs incomparable to the existing studies. One of the major reasons for this could be attributed to the terrain and meteorological factors prevailing there.

Result 4: Random Forest

Recent developments in data science and machine-learning algorithms have opened avenue for data-driven modelling for variety of applications. provided another approach to develop models. The data driven approach such as Random Forest (RF) can handle both continuous and discrete variables, identify best predictor amongst large number of possible predictor and also partition the data space to identify different best predictor subsets. RF is being adopted for a large number of applications and its use for modelling litterfall has been demonstrated in recent studies. Adoption of RF also permits exploring a larger set of predictors. However, for use of such an approach, adequate sampling of the target variable with sufficient observations is critical. Amongst various states, the forest in Uttarakhand have been most extensively studied and a survey on published litterfall indicated that highest number of studies were reported from Uttarakhand.

The RF model was developed using the package random Forest 4.7–1.1 in Rstudio environment running R version 4.1.2. The 100 data points were split in 80 − 20 ratio for model training-testing. Two parameters need to be optimized for constructing an effective RF model: ntree the number of trees to be grown by the model, and mtry the number of predictors to be randomly selected at each splitting step while growing the trees. The ntree parameter can be optimized by checking when the OOB (out of bag) error stabilises with respect to different number of trees, while for optimizing mtry parameter, tuneRF function can be used. RF also has the ability to provide variable importance of different predictor variables, based on the effect of permuting the variable on models mean square error and node purity. Furthermore, instead of using all the variables for model building, a fewer number of predictors which perform better (based on model RMSE) can be identified. For this, recursive feature elimination (RFE) algorithm with repeated cross-validation method was utilised. The variables thus identified by the RFE algorithm were finally employed to generate the RF model.

The RF model for Annual Litterfall was generated using 22 predictors (19 Bioclimatic and 3 topographic variables), with ntree = 500 and mtry = 7 after optimizing these parameters. Bioclimatic variables from BIO12-BIO19 represent Precipitation (12 for Annual, 13 = Wettest month, 14 = Driest month, 15 = Co-efficient of Variation, 16 = Wettest quarter, 17 = Driest quarter, 18- Warmest quarter, 19- Coldest quarter). Bioclimatic variables from BIO1-BIO11 indicates temperature (BIO1 for Annual mean, 2 = Mean Diurnal Range, 3 = Isothermality, 4 = Temperature Seasonality, 5 = Max Temperature of Warmest Month, 6 = Min Temperature of Coldest Month, 7 = Temperature Annual Range, 8 = Mean Temperature of Wettest Quarter, 9 = Mean Temperature of Driest Quarter, 10 = Mean Temperature of Warmest Quarter, 11 = Mean Temperature of Coldest Quarter). According to the variable importance ranking obtained by the RF model, the top-ranking variables were bio13, bio16, bio3, bio15, and bio12 based on %IncMSE as well as %IncNodePurity (Fig. 10). For selection of subset of variables which provide better model RMSE, RFE algorithm was run. The external resampling method used in RFE algorithm was repeated with 5 repeats and 10-folds. The optimum number of variables selected by the RFE algorithm was found to be 5 (Fig. 11). These top 5 variables- bio13, bio15, bio16, bio3, and bio12 were eventually used to generate the final RF model for annual litter fall prediction. Using testing data, it was found that the final model has RMSE value of 1.98 Mg/ha. The output annual litter fall map of Uttarakhand at 1km resolution generated using the above model is shown in Fig. 12. This spatial layer when overlaid with the 1km*1km spatial grid of Uttarakhand gave us the grid wise estimates of Total annual litterfall. The total over the state accounted to 13.305 million tonnes of litterfall flux for a year. The results were in comparison with our modelled output of litterfall per annum. A validation of the data driven output with our modelled one (Fig. 13) yielded a coefficient of determination (R2) of 0.92.

Preliminary comparison with independent estimates of litter pool and soil organic carbon indicates that such a spatial estimate of litterfall would improve their estimates. As litter pools are open to human and fire driven losses and SOC to vegetation cover and land-use changes, such spatial assessments would improve our understanding of changes in terrestrial carbon cycle. The study also highlights the important role of RS inputs.

Spain (1984) defines the carbon flux through litterfall as one of the major ways of the return of the nutrients from the aerial parts of the plants and the dead organic matter into the surface of the soil. It in turn proves to be a major source of energy for the forest floor saprobiota. Mathews (1997) reported the global annual litter production of 97 Pg and 100 Pg C yr − 1, using a combination of model based and data- based techniques based on relationship existing between the respiration of soil and root. He also observed a lower litterfall production over the tropical dry deciduous forests as well as moist and evergreen vegetation types. The major reason accounted for this was the lower net primary productivity in the arid and semi-arid regions. Another major contributing factor was the anthropogenic disturbances, due to which an accountable number of litterfall observations were collected from the open forest category. He estimated the total litterfall C flux for the period 1980-82 to be 210 ± 20 Tg C yr − 1, among which, 150 ± 13 Tg C yr − 1 comes from the leaf litterfall.

Though numerous measurements of litterfall are available at point level, they are not open for spatial estimation of forest processes such as primary productivity and Leaf Area Index (LAI). Early estimates of Litterfall over nations and biomes are based only on-site analysis. Prominent among them are those provided by ORNL DAAC, where Holland et al., (2014) measured above ground litter carbon, litterfall and littermass along with phosphorous, nitrogen and concentration of nutrients in litter pools and litterfall from 685 sources of literature and compiled them to a comprehensive database starting from 1827 to 1997. It also included their description, elevation, longitude and latitude, climatic information based on precipitation and temperature and certain characteristics of vegetation from site. Out of their total observations, only 87 observations were recorded from India. Hence, it is well evident that though global models were developed, they were not developed further and in all the studies, Indian studies are underrepresented.

Litterfall studies in India are carrried out on a regional scale, which are inadequate to estimate the total litterpool. In Uttarakhand forests, Thadani et al., (2015) reported that the average sequestered carbon is 3.12 ± 0.4 t ha-1 year-1, the maximum being from Eucalyptus spp. and the minimum from the forests of Pinus wallichiana (5.99 ± 1.9 t ha-1 year-1 and 1.25 ± 0.3 t ha-1 year-1 respectively). It is also interesting to note that, annually, nearly 5.9 million metric tons of carbon gets added to forest by tree sequestration. Hence, after the Cancun (2010) climate framework, REDD+ (Reduction Emissions from Deforestation and Degradation of forests and enhancement of forest carbon stocks) was included in 2012 post studies. They also observed that the various invasive species like Parthenium spp., Lantana camara and Eupatorium spp. have invaded the mountains of Himalayas of Uttarakhand along the rivers and roads.

Amongst the Indian forests, the total litterpool estimated from the average estimates of carbon densities of litter pool by Ajtay et al (1979) ranged between 390–410 Tg C between the years 1985–1986. Whereas, when estimated between the years 1985- 86, it was approximately 410–440 Tg C. Ray et al. (2013) observed that the litterfall fluxes over mangroves of Sundarbans dominated by Avicennia to be highest during the post-monsoon season and the annual litterfall accounting to be 2.07 Tg C/yr. The litterpools when estimated for subtropical and tropical forests by Palm et al., 1986 was reported to be 17 Tg C and 160 Tg C for the period 1980. Kaul et al. (2009) reported a decrease in the average carbon net flux from 5.65 Tg C yr-1 to 1.09 Tg C yr-1, while making a comparison between the years 1982–1992 and 1992–2002 due to land use changes. As a result of deforestation, the net C emissions assessed were 9.9 TgC and 3.2 TgC respectively. Over these years, a cumulative net C flux of 45.9 Tg C was observed as a result of land use changes. Yadav et al. (2022) studied a few selected tropical forests of Haryana and reported their potential for carbon sequestration to be 3.55–4.35 Mg C ha-1 year-1.

While a large number of studies investigating litterfall in Indian forests and plantations at site level are regularly published, the early major attempt to synthesize these observations for national scale forest litterfall and carbon flux assessment was by Dadhwal et al., (1997). The initial assessment of Indian terrestrial carbon cycle by Dadhwal and Nayak (1993) adopted global forest-wise litterfall from Ajtay et al., (1969). The forest-wise litterfall averages were updated by Chhabra and Dadhwal (2004) and most of the publications since then have not been assimilated and the global litterfall data bases (Holland et al., 2014) fail to capture the large number of Indian studies and data sets. Litterfall flux at national scale for India was estimated as 810–990 Tg C yr-1 (360–420 Tg C yr − 1 is contributed by the forests) by Dadhwal and Nayak (1993) based on global averages (Ajtay et al., 1979) and subsequently using Indian published data for Indian forests by Dadhwal et al., (1997) as 180 ± 24 and 150 ± 13 TgC yr − 1. This was subsequently updated by Chhabra and Dadhwal (2004). It included a total of 44 published literatures, which included 75 measurements of leaf litterfall (6.9 to 3.4 t/ha/yr) and 64 observations of total litterfall (8.5 to 4.3 t/ha/yr). The mean litterfall was used along with remote sensing based forest type inventory by Forest Survey of India (1987 assessment, based on 1980-82 satellite data) to estimate aboveground total and leaf litterfall carbon flux of 188 ± 24 and 152 ± 13 Mt C/yr in forest in India. Recently, Ahirwal et al., (2021) have summarized litterfall data over the Himalayas and looked at interrelations from this meta-analysis.

Litterfall when eventually merged with biomass of microbes will result the formation of soil organic matter (Neumann et al., 2018; Yang et al., 2020). Hence, litterfall plays a key role in determining the content of organic carbon in soil, the nutrient cycling (Chen and Chen, 2018), and regulate the respiration in soil, which in turn will indirectly affect the fertility of soil and thereby the response of global ecosystems to the varying climatic conditions (Gairola et al., 2009; Winkler et al., 2010). Therefore, litterfall plays a key role in calculating the equilibrium of carbon within the forests (Matala et al., 2008). In the present scenario of increasing global emissions, determining carbon budgets and studies are being carried out to target reducing the emission of carbon to the environment (Domke et al., 2016). Litterfall studies have marked importance and implications.

Modern data- driven approaches could provide results with higher accuracies over a larger extend. Generation of spatial maps of litter fall is a value addition to the present study. Still, the limitations of the study need to be emphasized. Classification system chosen for the study could highly influence the results. Areal extend compatibility is highly seen in the scheme of Champion and Seth Classification and it is adopted for the present study. Another uncertainty that remains is about the litter fall datasets that are derived over decadal time scale. Data driven litter fall are based on annual timescale, which in turn would cause an uncertainty owing to the fact that the litter fall is variable inter annually, depending on the seasons. The estimated quantities are prevalent under natural conditions, and hence no human disturbances were considered. There arises a need of including a gridded data of management and human disturbance over the study area. This is the way forward for studying the factors affecting litter fall.

Remote sensing-based forest type mapping of litterfall along with available field observations could be achieved. Summaries based on Leaf Area Index, Net primary productivity and climatic variables like temperature and precipitation aids the spatial depiction of litterfall better. Litterfall estimated using Random forest approach and other models indicates that the spatial variability has been captured. Understory representation of litterfall to litterpool of soil with changing Land Use Land Cover (LULC) needs to be developed.

Taking into account the complexity of forests in India, available data on litterfall estimates are limited. Hence increased number of litterfall estimates along with their associated abiotic and biotic factors helps improving the accuracy of modelling. Natural and anthropogenic disturbances like grazing, fuelwood extraction, burned area, deforestation, transitions in land cover, wind etc. also play a crucial role in the production of forest litterfall. Hence, these components need to be further looked into as a scope of the future study.

Acknowledgements

Authors acknowledge NASA LP DAAC (ee.ImageCollection ("MODIS/061/MOD15A2H"), WorldClim (www.worldclim.org/dataworldclim21), ORNL DAAC (http://daac.ornl.gov/cgi bin/dsviewer.pl?ds_id=1244) and NRSC (https://bhuvan-app3.nrsc.gov.in/data/download/ index.php?c=p&s=NI&g=all) for making MODIS, Global Climatic Data, Litterfall data, Soil Carbon Data, respectively, available through their websites. Authors are grateful to NIAS for making research facilities available. This research was carried out with funds provided by Indira Gandhi Memorial Trust, New Delhi to NIAS to host Chair Professor Scheme. Authors also thank the editors and reviewer of the journal for spending their valuable time in reviewing this paper.

Ethics approval and consent to participate

Authors declare that this work is original and has not been published elsewhere nor is it currently under consideration for publication elsewhere.

Availability of data and material

Supplementary files are attached with the manuscript as Annexure I. Other data will be made available to users on request.

Competing Interests

There is no conflict of interest regarding this publication.

Funding

This work was supported by ‘Indira Gandhi Memorial Trust (IGMT)’ to host the IGMT chair professor for Environmental Sciences. There is no funding available for publication for this research.

Author Contribution

First Author : Dr. Kripa M K : Data collection, Analysis, Interpretation, Writing

Second Author : Prof. Vinay Kumar Dadhwal : Conceptualization, Interpretation, Editing, Writing

Third Author : Atul Kaushik : Analysis, writing

Ahirwal J, Saha P, Nath A, Nath A. J, Deb S, and Sahoo U. K. 2021. “Forests Litter Dynamics and Environmental Patterns in the Indian Himalayan Region.” Forest Ecology and Management 499 (November):119612.https://doi.org/10.1016/j.foreco.2021.119612.
Ajtay, G., Ketner P, and Duvigneaud P. 1979. “Terrestrial primary production and phytomass, in The Global Carbon Cycle”, SCOPE 13, edited by B. Bolin, E. T. Degens, S. Kempe and P. Ketner, pp. 129–181, John Wiley, New York.
Bray, J. R., and Eville Gorham. 1964. “Litter Production in Forests of the World.” In Advances in Ecological Research, 2:101–57. Elsevier.
Breiman, L. 2001. “Random forests”. Machine learning, 45(1), 5–32.
Buchhorn, M; Lesiv, M; Tsendbazar, N. E; Herold, M; Bertels, L; Smets, B. Copernicus 2020. Global Land Cover Layers-Collection 2. Remote Sensing, 12, 108, 1044. doi:10.3390/rs12061044
Champion, H. G. and Seth, S. K. 1968. “A Revised Forest Types of India”. Manager of Publications, Government of India, Delhi.
Chen, X and Chen, H. 2018. “Global effects of plant litter alterations on soil CO2 to the atmosphere”. Glob. Change Biol. 24, 3462–71
Chhabra, A. & Dadhwal, V. K. 2004. “Assessment of Major Pools and Fluxes of Carbon in Indian Forests”. Climatic Change. 64. 341–360. 10.1023/B:CLIM.0000025740.50082.e7.
Dadhwal, V. K and Nayak S. R. 1993. “A preliminary estimate of biogeochemical cycle of carbon for India”. Science & Culture, 59(1/2): 9–13.
Dadhwal, V. K., Shukla, N., and Vora, A. B. 1997. “Forest litter fall in India: A review and an estimate of litter fall carbon flux”. The Indian Forester, 123, 45–52.
Domke, G., Perry C, Walters, B et al. 2016. “Estimating litter carbon stocks on forest land in the United States”. Sci Total Environ 557:469–478.
Fick, S.E. and Hijmans R.J. 2017. “WorldClim 2: new 1km spatial resolution climate surfaces for global land areas”. International Journal of Climatology 37 (12): 4302–4315.
Fararoda, R., Reddy, R. S., Rajashekar, G., Chand, T. K., Jha, C. S. and Dadhwal, V. K. 2021. Improving forest above ground biomass estimates over Indian forests using multi source data sets with machine learning algorithm. Ecological Informatics, 65, p.101392.
Farr, T. G., Rosen, P. A., Caro, E., Crippen, R., et al. 2007. “The shuttle radar topography mission”, Reviews of Geophysics,45(2) RG2004, https://doi.org/10.1029/2005RG000183.
FSI: 2021, ‘The Indian State of Forest Report’, Govt. of India, Forest Survey of India, Ministry of Environment & Forest, DehraDun, Vol 17, 2019–2020.
Gairola S., Rawal R. S., & Dhar U. 2009. “Patterns of litterfall and return of nutrients across anthropogenic disturbance gradients in three subalpine forests of west Himalaya, India”. Journal of Forest Research, 14(2), 73–80. https://doi.org/10.1007/S10310-008-0104-6
Geng, A, Qingshi T, Jiaxin C, Weifeng W, and Hongqiang Y. 2022. “Improving Litterfall Production Prediction in China under Variable Environmental Conditions Using Machine Learning Algorithms.” Journal of Environmental Management, 306 (March): 114515. https://doi.org/10.1016/j.jenvman.2022.114515.
Holland, E. A., Post W.M., Matthews E.G., Sulzman J.M., R. Staufer, and Krankina O.N. 2015. “A Global Database of Litterfall Mass and Litter Pool Carbon and Nutrients.” Data Set. Tennessee, USA: ORNL. http://daac.ornl.gov/cgi-bin/dsviewer.pl?ds_id=1244
Hu, X, Xiao-Yan L, Pei W, Yong L, Xiu-Chen W, Zong-Chao L, Yun-Duo Zhao et al. 2019. "Influence of exclosure on CT-measured soil macropores and root architecture in a shrub-encroached grassland in northern China." Soil and Tillage Research. 187: 21–30.
Jia, B., Xu, Z., Zhou, G., Yin, X., 2018. “Statistical characteristics of forest litterfall in China”. Sci. China Life Sci. 61 (3), 358–360.
Jia, B, Hongru S, Wenying Y, and Guangsheng Z. 2020. “Quantifying the Interannual Litterfall Variations in China’s Forest Ecosystems.” Edited by Osbert Jianxin Sun. Journal of Plant Ecology 13(3): 266–72.https://doi.org/10.1093/jpe/rtaa010.
Li, Shihua, Wenping Yuan, Philippe Ciais, Nicolas Viovy, Akihiko Ito, Bingrui Jia, and Dan Zhu. 2019. “Benchmark Estimates for Aboveground Litterfall Data Derived from Ecosystem Models.” Environmental Research Letters 14 (8): 084020 https://doi.org/10.1088/1748-9326/ab2ee4
Liu, X., Zhou, T., Luo, H., Xu, P., Gao, S., Liu, J. 2019. “Models ignoring spatial heterogeneities of forest age will significantly overestimate the climate effects on litterfall in China”. Sci. Total Environ. 661, 492–503.
Malhi, Y., Doughty, C., & Galbraith, D. 2011. The allocation of ecosystem net primary productivity in tropical forests. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 366(1582), 3225–3245. https://doi.org/10.1098/rstb.2011.0062
Matala, J, S. Kellomaki and T. Nuutinen 2008. “Litterfall in relation to volume growth of trees: analysis based on literature”. Scandinavian Journal of Forest Research 23: 194–202.
Matthews, Elaine. 1997. “Global Litter Production, Pools, and Turnover Times: Estimates from Measurement Data and Regression Models.” Journal of Geophysical Research: Atmospheres 102 (D15): 18771–800. https://doi.org/10.1029/97JD02956.
Meentemeyer, V, Elgene O. Box, and Richard T. 1982. "World patterns and amounts of terrestrial plant litter production”. BioScience 32, no. 2: 125–128.
Mishra, A. P., Rai, I. D., Pangtey, D. and Padalia, H., 2020. Vegetation Characterization at Community Level Using Sentinel-2 Satellite Data and Random Forest Classifier in Western Himalayan Foothills, Uttarakhand. Journal of the Indian Society of Remote Sensing, pp.1–13.
Neumann, M., Liisa U., James J., Sue B., Lars V., Radek N., Arne V., et al. 2018. “Quantifying Carbon and Nutrient Input From Litterfall in European Forests Using Field Observations and Modeling.” Global Biogeochemical Cycles 32 (5): 784–98. https://doi.org/10.1029/2017GB005825.
NRSC (2016). Indian Soil Carbon Data set. NICES/DS(L)/SC/2014/Jan2016.
Pietsch, S. A., Hasenauer, H., & Thornton, P. E. 2005. “BGC-model parameters for tree species growing in central European forests”. Forest Ecology and Management, 211(3), 264–295. https://doi.org/10.1016/j.foreco.2005.02.046
Reddy, C. Sudhakar, Jha, C. S., Diwakar, P. G., Dadhwal, V. K. 2015. “Nationwide classification of forest types of India using remote sensing and GIS”. Environmental Monitoring and Assessment.187, 777. https://doi.org/10.1007/s10661-015-4990-8
Roy, P. S., P. Meiyappan, P. K. Joshi, et al. (2016). Decadal Land Use and Land Cover Classifications across India, 1985, 1995, 2005. ORNL DAAC, Oak Ridge, Tennessee, USA. https://doi.org/10.3334/ORNLDAAC/1336
Shen, G., Xiang, Q., Chen, D., Wu, Y., Liu, C. 2017. “Spatio- temporal distribution characteristics of forest litterfall in China”. Chin. J. Appl. Ecol. 28 (8), 2452–2460.
Shen, G, Dongmei C, Yu W, Lu L, and Chunjiang L. 2019. “Spatial Patterns and Estimates of Global Forest Litterfall.” Ecosphere 10 (2): e02587. https://doi.org/10.1002/ecs2.2587.
Sreenivas, K., Sujatha, G., Sudhir, K., Kiran, D.V., Fyzee, M.A., Ravisankar, T. and Dadhwal, V.K. 2014. Spatial assessment of soil organic carbon density through random forests based imputation. Journal of the Indian Society of Remote Sensing, 42(3), pp.577–587.
Sreenivas, K., Dadhwal, V.K., Kumar, S., Harsha, G., Mitran, T., Sujatha, G., Suresh, G. J. R., Fyzee, M.A., and Ravisankar, T. 2016. Digital mapping of soil organic and inorganic carbon status in India. Geoderma, 269(160). https://doi.org/10.1016/j.geoderma.2016.02.002
Srinet, R., Nandy, S. and Patel, N. R., 2019. Estimating leaf area index and light extinction coefficient using Random Forest regression algorithm in a tropical moist deciduous forest, India. Ecological Informatics, 52, pp.94–102.
Vogt, K.A., C.C. Grier, and D.J. Vogt. 1986. “Production, Turnover, and Nutrient Dynamics of Above- and Belowground Detritus of World Forests.” In Advances in Ecological Research, 15:303–77. Elsevier. https://doi.org/10.1016/S0065-2504(08)60122-1.
Wang, Hsueh-Ching, Kuo-Chuan L and Cho-ying H. 2016. “Temporal and Spatial Patterns of Remotely Sensed Litterfall in Tropical and Subtropical Forests of Taiwan: Patterns of Remotely Sensed Litterfall.” Journal of Geophysical Research: Biogeosciences 121 (2): 509–22. https://doi.org/10.1002/2015JG003113.
Wen, D and He, N. P. 2016. “Spatial patterns of litter density and their controlling factors in forests and grasslands of China”. Acta Ecol. Sin. 36, 2876–2884.
White, M. A., Thornton, P. E., Running, S. W., & Nemani, R. R. 2000. “Parameterization and sensitivity analysis of the BIOME–BGC terrestrial ecosystem model: Net primary production controls”. Earth Interactions, 4(3), 1–85. https://doi.org/10.1175/1087-3562(2000)004
Winkler, R., Boon, S., Zimonick, B., Baleshta, K. 2010. “Assessing the effects of post-pine beetle forest litter on snow albedo”. Hydrol. Process. 24 (6), 803–812.
Yang, X., Szlavecz, K., Pitz, S. L., Langley, J. A., Chang, C. H 2020. “The partitioning of litter carbon fates during decomposition under different rainfall patterns: a laboratory study”. Biogeochemistry 148 (2), 153–168
You, C., Wu, F., Yang, W., Tan, B., Yue, K., Ni, X. 2017. “The National key Forestry Ecology Project has changed the zonal pattern of forest litter production in China”. For. Ecol. Manag. 399, 37–46.
Zhang, H, Wenping Y, Wenjie D, and Shuguang L. 2014. “Seasonal Patterns of Litterfall in Forest Ecosystem Worldwide.” Ecological Complexity, 20 (December): 240–47. https://doi.org/10.1016/j.ecocom.2014.01.003.
Zhang, N., Li, Y., Wubet, T., Bruelheide, H., Liang, Y., Purahong, W., Buscot, F., Ma, K. 2018. “Tree species richness and fungi in freshly fallen leaf litter: unique patterns of fungal species composition and their implications for enzymatic decomposition”. Soil Biol. Biochem., 127, 120–126
Zhao, X, Xiaolu T, Jie D, Xiangjun P, Guo C, and Tingting X. 2022. “A Data-Driven Estimate of Litterfall and Forest Carbon Turnover and the Drivers of Their Inter-Annual Variabilities in Forest Ecosystems across China.” Science of The Total Environment, 821 (May): 153341. https://doi.org/10.1016/j.scitotenv.2022.153341.

No competing interests reported.

Appendix.odt

Spatial Assessment of forest litterfall in Central Himalayas (India): Comparison of geospatial, remote sensing and data-driven estimates

Status:

Version 1

Abstract

Figures

Introduction

Study area

Methods