Rakesh Paul · Kakoli Banerjee

Abstract Tropical forest cover change along with increasing fragmentation has detrimental Effects on the global biodiversity. In the current study change in both forest cover and fragmentation of Koraput district have been assessed in the past three decades (1987 − 2017) and future decade(2017 − 2027), which has been modelled using logistic regression showing a gradual decrease in the forest cover and increase in fragmentation. The long term deforestation rates from 1987 to 2017 (current period) and from 1987 to 2027 (predicted period) were found to be -0.018 and -0.012,respectively. Out of the total geographical area, 2027 number of grids (1 km 2 ) out of 8856 grids were found to have shown extinction of forest in the study period. The conversion of forested lands into other land uses has been one of the major causes of deforestation in Koraput, especially because of the increasing mining activities and establishment of three major industries namely National Aluminium Company (NALCO),Damanjodi, Hindustan Aeronautics Limited (HAL), Sunabeda and Ballarpur Industries Limited (BILT). The forest fragmentation reveals a negative trend, recording highest conversion from large core fragments to edge (191.33 km 2 )and the predicted period has also shown the same trend of negative change, which poses serious danger to the structure of the forests. Out of all the landscape matrices calculated,number of patches will increase to 214 in 2027 from 93 in 1987. In the test between geographically weighted regression (GWR) and ordinary least square regression (OLS),GWR was the better f it model for drawing a spatial relationship between forest cover and fragmentation changes. The study conf irmed that the forest cover change has impacted the forest fragmentation in the study area. The programmes like REDD + should be implemented along with the experiences of Community Forest Management and the joint forest management should be intensif ied at community level in order to develop better management practices to conserve habitats in biodiversity rich areas.

Keywords Forest cover · Forest ecosystem · Predictive model · Forest fragmentation · Grid based change ·Geographically weighted regression

Introduction

Forest ecosystems are vital for all the life forms on the planet, degradation of which poses direct threat to diff erent ecosystem services in various dimensions. Tropical forest ecosystems harboring more than 60% of the global terrestrial species have been under tremendous pressure facing loss of its biodiversity in the past decades (Gardner et al. 2009)leading to extinction of many plants and animal species along with the forest cover itself (Fahrig 2013). The rate of deforestation has been increasing drastically and reportedly there has been a global forest loss of 2.3 million km 2 from 2000 to 2012. In the recent past, the tropical forests alone have faced 7 Mha of annual forest loss between 2000 and 2010 (Hansen et al. 2013). The important driving forces of deforestation may be attributed to and can be def ined based on their location in relationship to the cultivated lands, settlements and need for fuel woods (Kindu et al. 2013). Tropical deforestation and biodiversity loss induced from land use(LU) changes coupled with pollution and overexploitation of natural resources (Kobayashi et al. 2019) is one of the most potent causes for the addition of anthropogenic greenhouse gases to the atmosphere (Lemke et al. 2007).

Forest fragmentation is increasing in an alarming rate which has been reported worldwide to be a key process causing progressive subdivisions of forested lands and converting them into dispersed and isolated patches (Gibson et al. 1988). Forest fragmentation and forest loss are critical for conserving species diversity and have been rigorously studied in the last decades (Fardila et al. 2017), but there is a great lack of evidence on the dependency of forest fragmentation on change in forest cover (Kozak et al. 2018). Fragmentation process increases the edge of the forested landscape, which can be responsible for dramatic modif ication of ecological conditions and carbon accumulation capacity as the forest edge has 50% lesser carbon storage capacity than the core forests (Gibson et al. 2013). Considering the ecosystem richness and their services, long term change in forest fragmentation, forest cover and forest types have been analyzed and summarized in several parts of India including Odisha and Eastern Ghats which have concluded large scale deforestation and high increasing rate of forest fragmentation (Roy et al. 2013; Krishna et al. 2014).

Spatiotemporal changes of forest cover and other environmental processes can be simulated using the combination of remote sensing with multiple complex modelling systems such as cellular automata (CA), logistic regression (LR),MARKOV chain (MC) as well as artif icial neural networking (ANN) for predicting future land cover and have proven to be relevant and reliable tools for forest ecosystems analyses (Tiné et al. 2019). Along with the anthropogenic causes of forest loss, there are certain other spatial variables like slope, elevation and aspects which constitute the topography of the locality may also help improving the classif ication of the disturbance in the forest landscapes (Schroeder et al.2019).

For the quantif ication of fragmentation, landscape matrices play a vital role to understand its spatial conf iguration and distribution by computing the parameters like patch sizes, edge length and area of the patches (Betts et al. 2003)which helps in temporal comparison of them in any geographical area. The relationships of forest change with forest fragmentation and vice-versa through spatial correlation can provide a wide dimension of management practices which is essential for sustainable conservation of forest resources.Spatial correlation provides the better representation of statistical signif icance of models (Salas et al. 2010), which can be Effectively utilized for forest ecosystem studies. The geographically weighted regression (GWR) for spatial correlation has yielded various dimensions of explanation of relationships between diff erent ecological variables using hierarchical modeling (Eiserhardt et al. 2011; Xu et al.2016). This method can be utilized for understanding the spatial distribution of the forest changes in relation to the forest fragmentation process (Zhang and Gove 2005).

Though the overall forest cover change and forest fragmentation of the locality has been studies by few other scientists, still the critical aspect in micro dimensions of these parameters hasn’t been taken into consideration. In addition the futuristic changes of forest resources and its fragmentation based on the present scenario haven’t been evaluated in Koraput district. There is also a lack of scientif ic evidence on the spatial correlation of deforestation with forest fragmentation in many parts of India including Koraput district.Therefore, based on the past researches the current study is based on the hypothesis that deforestation and forest fragmentation in the study area increases at a considerable rate during the present and future scenarios and the two processes are spatially correlated.

The present study has addressed several research questions related to forest degradation encompassing the complex modeling approaches in a bio-geographically important regional landscape which has not been studied in minute details. Thus the present research programme focuses on addressing the following research questions: (1) What is the rate of deforestation and its driving forces during the period 1987-2017 and what are its future projections till 2027?(2) What is the present change in forest fragmentation and what are its future projections till 2027? (3) How the forest fragmentation and deforestation are spatially correlated?

Materials and methods

Study area

Koraput district of Odisha, an integral part of Eastern Ghats:one of the four Biodiversity Hotspots of India (Myers et al.2000) is an Agro-Biodiversity hotspot of the country ( https://www.plant autho rity.gov.in/hotsp ots.htm). The study area,having a geographical area of 8379 km 2 , extending from 18°13′ N to 19°10′ N and 82°5′ E to 83°13′ E in the peninsular India and elevation ranging from 123 to 1672 m above MSL (mean sea level) has a global signif icance as one of the important biodiversity hotspots (Adhikary et al. 2019)(Fig. 1). It is located in the relatively higher altitude region of the Eastern Ghats compared to the other parts and experiences summer, monsoon (Southwest monsoon) and winterseasons throughout the year. The district gets an average annual rainfall of ～ 2278 mm whereas the average minimum and maximum temperature varies from 14 to 41.5 °C,respectively, and it has three distinct topographical ecotypes:Hilltops, Foothills and Plains (Sahu et al. 2013). Three major industries namely National Aluminium Company (NALCO),Damanjodi, Hindustan Aeronautics Limited (HAL), Sunabeda and Ballarpur Industries Limited (BILT), Jeypore along with two major bauxite mines Panchpatmali and Ampavalli are currently operating in the study area. Tribes of Koraput are dependent on agriculture and adopt the practices like shifting cultivation (slash-and-burn), use forest and forest products for shelter, fuel and also for non-timber forest products (NTFPs) which cause forest loss. However, being a part of community-managed forest, village communities in the district have been involved in aff orestation programmes with their own system of protection for forest regeneration which enhances the importance of the area under study.

Fig. 1 Study area map with major sampling stations (for forest types analysis), ground control points (GCPs) and land use land cover (LULC) information of the current period (2017)

Quantif ication of variables and data analyses

Evaluation of rate of deforestation and fragmentation

The satellite imageries were downloaded from the United States Geological Survey (USGS), Earth Explorer ( https://earth explo rer.usgs.gov/ ) over a period of 40 years viz.1987,1997, 2007, and 2017 of 30 m spatial resolution. The uniformity of resolution (30 m) and the acquired periods of the satellite imageries (November to January) were maintained throughout the study. Landsat-5 TM (Thematic Mapper)imageries were used for the years 1987 (path/row: 142/047,cloud cover: 0%) and 1997 (path/row: 142/047, cloud cover:0%), Landsat-7 ETM + (Enhanced Thematic Mapper plus)for 2007 (path/row: 142/047, cloud cover: 2%) and Landsat-8 OLI (Operational Land Imager) for 2017 (path/row:142/047, cloud cover: 0%). The SRTM (Shuttle Radar Topography Mission) DEM (Digital Elevation Model, 30 m)was used for the elevation and slope information. Toposheets(1:50,000 scale) procured from Survey of India, Bhubaneswar (year of publication: 2010) were georectified and used for the digitization of the major roads and used as the basemap during the accuracy assessment process along with Google Earth Pro. All the imageries were projected in the WGS 1984/UTM zone 44 N projected coordinate system.

The rate of deforestation and forest fragmentation along with future projections were done using ERDAS IMAGINE, QGIS and ArcMap. The image processing was performed in ERDAS IMAGINE-2015 using image to image rectification method. Maximum likelihood supervised classif ication method (Lillesand et al. 2014) was followed for the land use land cover (LULC) classif ication of six diff erent land use classes for successive year (1987, 1997,2007 and 2017). The classes are namely water bodies(WB), agricultural lands (AL), human settlements (HS),open forest (OF), moderately dense forest (MDF) and dense forest (DF). For fragmentation analysis the LULC maps were converted into binary maps of non-forest and forest classes by keeping WB, AL and HS under non-forest class and OF, MDF and DF under the forest class.

Both natural (Elevation, slope and distance to water bodies (DWB)) and anthropogenic factors (Distance to major roads (DMR) and distance to human settlements(DHS)) were taken as the potential drivers of deforestation in the study area. Transition matrix of the diff erent LULC classes from 1987 to 2017 was generated in order to explain potential drivers of deforestation along with all the spatial variables taken during the future simulation of forest cover in the district.

Decadal forest cover change (1987-1997, 1997-2007,2007-2017 and 2017-2027) in the study area were analyzed over the said periods. ArcMap 10.2.1 and QGIS 2.8.4 were used for all the geospatial analyses in the present study.

The rate of deforestation was calculated as per the formula given by Puyravaud ( 2003)

where,T1andT2are the initial (time 1) and f inal time (time 2) (years),A1 andA2 are the areas under forest in time 1 and time 2.

Evaluation of forest fragmentation

Landscape Fragmentation Tool (LFT) (version 2.0) extension in ArcGIS was used for classifying the diff erent fragment classes using the binary rasters (class 1 = non-forest,class 2 = forest) assigning 100 m as edge width (Vogt et al.2007). The classif ied fragments are patch (small forest areas without core pixels and enclosed by the non-forest area), edge (transition zone between core forests pixels and non-forest pixels), perforated (buff er between the core forest pixels and non-forested pixels present inside the core)and core (unaff ected from any of LUs and edge lengths)based upon their vulnerability towards different LUs.Further, core forests were classif ied into of small core,medium core and large core which have areas < 100 ha,100-200 ha and > 200 ha respectively (Sulieman 2018).

Change in different fragmentation classes over the four decades (1987-1997, 1997-2007, 2007-2017, and 2017-2027) were analyzed and conversion matrices were generated. The conversion of any of the three core fragments (small, medium and large core) to patch, edge or perforated fragments, perforated to edge, edge to patch along with the conversion of larger cores to smaller ones were taken as negative change and the vice-versa was considered under positive change.

Grid based forest cover and fragmentation change analysis

To critically analyze the changes in forest cover and forest fragmentation, grids of 1 km 2 (1 km × 1 km) were laid over the total geographical area and areas under each categories of change were spatially joined to the grids. Grids under both the decrease and increase in case forest cover and negative change and positive change in case of fragments were assigned class breaks of < 5% (< 0.05 km 2 ), 5-10%(0.05-0.1 km 2 ), 10-20% (0.1-0.2 km 2 ), 20-40% (0.2-0.4 km 2 ), 40-80% (0.4-0.8 km 2 ), and > 80% (> 0.8 km 2 ) (of each 1 km 2 grid). The long term forest cover changes and fragmentation changes from 1987 to 2017 and 1987 to 2027 were analyzed through the grid based approach.

Future forest cover simulation

MOLUSCE (Module for Land Use Change Evaluation)plug-in of QGIS 2.8.4 was used for the simulation of future forest cover in the year 2027 based on the forest covers of 2007 and 2017. Elevation (in m), slope (in degree (°))(obtained from DEM) and the Euclidean distances from water bodies (DWB) (in km), major roads (DMR) (in km)(obtained from digitization of toposheets) and settlements(DHS) (in km) were taken as explanatory spatial variables for the simulation. The Euclidean distances were obtained in ArcMap 10.2.1. using the distance tool. Pearson correlation was established between all the spatial variables in order to evaluate the strength of their interdependencies. Artif icial neural network (ANN) and logistics regression (LR) models were tested with 100 iterations for 1000 number of random samples through transition potential modelling to f ind out the best f it model (taking 80% samples for training, 20% for validating).

In the present study logistic regression (LR) proved to be the better f it model. This method is used worldwide for LULC change modeling which uses continuous predictor variables within a predef ined time interval and can be Effectively used for binary responses (Tiné et al. 2019). The mathematical expression for the model is given below:

where,P iis probability of a pixel for occurrence of the response variable,X n,iare spatial variables, andβis regression parameter.

Two step validation process was followed for the model,where in the f irst step a simulated map for the year 2017 was generated by taking 1997 and 2007 as the initial and f inal years. In the next step the simulated map of 2017 was validated taking the classif ied map of 2017 as the reference map by checking the persistence classes in f ive validation iterations. Percentage of correctness, Kappaoverall, Kappahistogramand Kappalocalwere the four diff erent parameters used for accuracy assessment of the model.

Ground truth verif ication for accuracy assessment

The f inal forest cover maps of the years 1987, 1997, 2007 after merging of diff erent land use classes were validated by taking 272, 386 and 419 number of ground control points(GCPs), respectively. Field visits were carried out during 2017, where 442 geo-tagged GCPs were taken in the study area by hand held GPS. The vegetation types were monitored in four diff erent sampling locations having GPS coordinates 18°43′15.6′′ N 82°30′28.1′′ E, 19°07′30.33′′ N 82°48′42.6′′E, 18°56′00.5′′ N 83°06′02.7′′ E, and 18°46′02.3′′ N 82°08′22.07″ E viz.stations 1, 2, 3 and 4 respectively taking square quadrats of 1 ha (250 m × 250 m). The bigger quadrat was further sub-divided into quadrats of 0.1 ha(31.6 m × 31.6 m) for the forest type analysis. Commission error, omission error, producer’s accuracy, user’s accuracy and Kappa coeffi cients were the parameters for the validation (Pontius 2000).

Analysis of spatial correlation of forest fragmentation and deforestation

Landscape Ecology Statistics (LecoS) version 2.0.7 plug-in of QGIS was used for the landscape matrices calculation,which has been developed using FRAGSTATS used as a stand-alone tool or in combination with complex ecological models (Jung 2016). Grids of 5 km 2 (5 km × 5 km) were taken inside the study area where edge length (EL), number of patches (NP), greatest patch area (GPA), mean patch area (MPA), largest patch index (LPI) and overall core area(OCA) in each grid were calculated. In case of complete correlation (r= 1), one matrix was selected over the other depending on its performance in explaining the fragmentation. The grids cutting or touching the district boundary and those falling in non-forest areas were discarded in order to minimize the statistical error. For spatial correlation analysis only the actual long term change (1987 − 2017) was taken into consideration. Ordinary least squared (OLS) (global model) and geographical weighted regression (GWR) (local model) were tested taking the forest cover change (FCC)as the explanatory variable and fragment matrices as the dependent variables in each case. We used a f ixed kernel type with akaike information criterion (AICc) method of bandwidth selection. OLS regression takes the whole area into consideration and computes the overall relationship between variables whereas GWR identif ies spatial relationships at each location by calculating locally-specifying coeff icients (Tripathi et al. 2019). The statistical analyses and graphs were prepared in MATLAB 2015a and R-Studio.

The schematic diagram of detailed methodology is given in Fig S1.

Results

Rate of deforestation and its futuristic perspective

The area under forest were calculated to be 3170.42 km 2(37.84%), 2733.68 km 2 (32.62%), 2220.93 km 2 (26.51%),and 1858.36 km 2 (22.18%) for the years of 1987, 1997, 2007 and 2017, respectively. The predicted forest cover for 2027 was found to be 1681.78 km 2 (20.07%) out of the total geographical area (TGA) (Table 1; Fig. S2). The Kappa statistics obtained from the training of ANN predictive model was 0.762 which was much lower than the pseudoR2 value of 0.920 in LR model proving to be the better f it model for simulation. The percentage of correctness, Kappaoverall,Kappahistogramand Kappalocalbetween the simulated and the classif ied forest cover maps of 2017 were found to be 99.993%, 0.998, 0.998 and 1.000, respectively, proving the predicted map of 2027 to be highly accurate. The rate of deforestation varied as − 0.015, − 0.021, − 0.018, and − 0.010 in the decades 1987 − 1997, 1997 − 2007, 2007 − 2017,and 2017 − 2027, respectively, showing the highest rate in 1997 − 2007 accounting for 512.75 km 2 (6.12%) out of the TGA (Table 1). The long term deforestations were estimated to be 1312.06 km 2 (15.66%) from 1987 to 2017(current period) and 1527.58 km 2 (16.69%) from 1987 to2027 (predicted period) at the rates of − 0.018 and − 0.012,respectively (Table 1; Fig. 2 ).

Table 1 Forest cover and rate of deforestation over four decades in the study area

The long term changes of 1987 − 2017 and 1987 − 2027 showed 1849 (20.88%) and 1609 (18.17%) number of grids in the decreasing category and 345 (3.90%) and 358 (4.04%)in the increasing category out of total 8856 grids, but in the later period the grids under the greater break values of 40% − 80% and > 80% showed more grids indicating the intensity and localization of the deforested areas to certain regions of the study area (Table 2; Fig. 2). In the period of past 30 years a lot of forested grids have shown extinction(100% loss) of forest cover and got converted to non-forest lands. A total of 1653 number of grids shown extinction of forest in the period and the predicted decade will face the extinction of 374 more grids till 2027 (Fig. 2).

Pearson’s correlation showed obvious interrelationships between most spatial variables (rSlope×DWB= 0.998,rDMR×DWB= 0.981,rDHS×DWB= 0.999,rDMR×Slope= 0.980,rDHS×Slope= 0.998 andrDHS×DMR= 0.981) showing the roads,human settlements and rivers inf luencing the forest loss, but elevation and slope don’t determine the rate of forest loss.The transition matrix of land use land cover classes indicated the highest probability in the conversion of dense forest into the human settlements (72.66%), followed by moderately dense forest into open forest (46.96%), open forest into agricultural lands (35.85%), moderately dense forest into agricultural lands (30.99%) and open forest into human settlements (27.3%), apart from which all the transitions have shown weaker probabilities which clearly indicates that transition of the healthy forested lands into the human habitations and agricultural lands acted as the primary causative factor for deforestation in the district over the years.

The above results have been justif ied through ground data obtained during the f ield studies in 2017. The forest types found were North dry mixed deciduous forest (Group.5,Sub-group.5B) in Stations 1 and 2 and Moist peninsular high level Sal forest (Group.3, Sub-group.3C) in Stations 3 and 4, respectively (as per the Champion and Seth’s revised classif ication of diff erent forest types on India, 1968). The forest maps for the years 1987, 1997, 2007 and 2017 were classified with overall accuracies of 99.63% (k= 0.99),99.48% (k= 0.99), 98.81% (k= 0.98) and 98.64% (k= 0.97)respectively with very less commission and omission errors.

Fig. 2 Grid based long term forest cover change ( a, b) and forest extinction ( c) in the current (1987 − 2017) and predicted period (2017 − 2027)

Table 2 Number of 1 km2 grids under diff erent class breaks of forest cover and fragmentation changes for current and predicted periods

3.1. Change in forest fragmentation and its future projections

Overall result of forest fragmentation suggests that most share of area was under the edge fragments class (901.70 km2) followed by the large core fragments (803.85 km2)in 1987 which was gradually degraded by diff erent land uses giving rise to more patch fragments in the later years,apart from which all other fragments have shown gradual decrease in area (Fig. S3). The large core has decreased in the trend − 246.51 km 2 (− 2.94%), − 227.27 (− 2.71%),and − 219.18 (− 2.62%) in the intervals of 1987 − 1997,2007 − 2017, and 1997 − 2007, respectively, whereas the medium core and small core fragments have shown better stability with less negative change.

The overall change in the forest fragments were calculated for 1987 − 1997, 1997 − 2007 and 2007 − 2017 to be − 436.75 km 2 (− 5.21%), − 512.75 km 2 (− 6.12%)and − 362.57 km 2 (− 4.33%) respectively in 1997 − 2007 registering the highest negative change in fragments(Fig. A3). The predicted decade has shown a greater decrease of − 176.59 km 2 (− 2.11%) from 2017 predicting more threat to the structure of the forest in the study area which was supported by the increase in the patch fragment to 54.08 km 2 (0.65%) matching the trend of the past three decades. The conversion matrices for the long term changes have shown the highest conversion from the large core fragments to edge (191.33 km 2 ), than to perforated fragments(157.84 km 2 ) and small core fragments (92.63 km 2 ) apart from which large conversions were recorded only in case of edge fragments converting to patch (98.49 km 2 ) (Table S1).Similar observations in the trend were observed in case of the predicted period.

In the long term change of fragmentation, the period of 1987 − 2027 has shown more number of negatively changing grids (2355) contributing 26.59% than the past 30 years(2347) contributing to 26.50% of the total number of grids with less number of unchanged grids which pertains to the scenario of more negative fragmentation compared to the past showing greater threat to the forest structure (Table 2;Fig. 3). Neither the current period nor the predicted period showed positive change in the break values greater than 20%. The break values of < 5%, 5% − 10%, 10% − 20%,20% − 40%, 40% − 80%, and > 80% have shown similar trends having the highest number of grids under the smaller break values with 90% of the changing grids concentrated in the < 5%, 5% − 10%, and 10% − 20% ranges and gradually decreased towards the greater break values of 20% − 40%,40% − 80%, and > 80%, which implies that mass destruction of the forests has lesser probability in degrading the forest structure in the study area compared to the slower processes like expansion of the industrial areas and increase incidents of invasion of human settlements into the forested lands(Table 2; Fig. 3).

3.2. Interrelationship between forest fragmentation and deforestation

An exemplary grid (grid no. 140) out of 238 grids has been taken where the forest change FC, number of patches change(NPC) and edge length change (ELC) have shown the least collinearity among all the other matrices whereas the complete positive correlation (r= 1) between greatest patch area change (GPAC) and largest patch index change (LPIC)compelled us to choose LPIC over GPAC as it provides the measure of dominance of fragment size (Figs. S4 and S5).

The equivalent box plots for the diff erent years have shown diff erences in case of all the fragment matrices. The medians from 1987 to 2027 has decreased from 5.84 km 2 to 4.10 km 2 and the year 2007 showed the highest skewness of 1.15 and most number of outliers indicating the dispersedistribution of forest cover (Fig. S6). In our results edge length has shown gradual increase from 1987 (118.41 km)to 2017 (127.32 km) with the maximum length of 329.52 km in 2017 and further increase to 142.74 km in the predicted period with the maximum length of 343.02 km showing the risk to forest structure as it is one of the most important reasons for the forest fragmentation (Fig. S6). In the present study the number of patches has drastically increased throughout the current and future periods showing an increase from 93 in 1987 to 189 in 2017 and a predicted increase up to 214 in 2027 which can be a major threat to the forest ecosystem (Fig. S6). The mean patch area, largest patch index and overall core area has shown the similar trend over the years showing the highest values in the year 1987 which has gradually decreased towards the later period showing greater upper quartile values in the inter-quartile range and increased number of outliers beyond the whiskers (Fig. S6)

In total, 10 models (5 for OLS and 5 for GWR) were run in order to verify the spatial correlation of forest fragment change with the forest cover change (Table S2). The coeffi cient of determinationR2 (overall) were stronger in the GWR models showing values of 0.601 for M 12 (FCC vs ELC), 0.402 for M22(FCC vs NPC), 0.469 for M32(FCC vs MPAC), 0.836 for M42(FCC vs LPIC) and 0.935 for M52(FCC vs OCAC) with respect to 0.051, 0.028, 0.188,0.743, and 0.839 in the corresponding OLS models proving the former to be the better f it model for all the combinations (Table S2; Fig. 4 a, b, c, d and e). The lower z-scores supported by the less spatial autocorrelation with lower Moran’s I values of the residuals in GWR indicated lesser probability of clustering and hence proving the spatial heterogeneity among the variables which was also indicated by lower AICc values for GWR models (− 916.683,2520.670, − 175.847, 1575.233 and 574.876) compared to OLS models (− 756.227, 2590.408, − 120.659, 1636.692 and 745.164) in each respective combination signifying the better performance of the local model over the global model.

In case of FCC vs ELC theR2 values showed strong spatial correlation in the southern part of the district and least towards the north-eastern part, whereas the north-eastern part showed higher correlation for FCC vs NPC which reveals that patch numbers have increased irrespective of the increase or decrease edge length (Fig. 4 a and b). The relationship of forest cover change with change in mean patch area and overall core area indicates that deforestation contributes to degradation of core areas irrespective of the area of the patch (Fig. 4 c and e). Forest cover change has shown a stronger spatial relationship with largest patch area index in the western part and northern part of the district resembling with the results of change in mean patch area with overall core area except for its weaker correlation towards the southern part of the district (Fig. 4 c, d and e). All the landscape matrices showed relatively stronger spatial correlation with forest cover change towards most part of the western region and some part of northern region of the district except the change in number of patches (NPC).

Discussion

Understanding the forest change pattern in the past and future are inevitable for formulation of management planning and conservation policies. The LR simulation model has been proven to be the best f it model for the simulation of static processes (Mas et al. 2014) and in our study we have successfully simulated a predicted map with high accuracy using the same. We found a gradual loss of forested lands in the past decades which will continue in the future which complements the results of Reddy et al. ( 2019) in Myanmar and Ahammad et al. ( 2019) in Bangladesh. In 2007-2017,the forest cover has increased in comparison to the previous two decades which has also been stated in the Odisha government state forest report. The overall historical forest loss of the study area has previously been analyzed by Dash et al.( 2018) where they have pointed out the shifting cultivation and dam construction as the major cause for forest loss. The current study reveals that along with the expansion of agricultural lands, the conversion of dense forest to moderately dense forest to open forest and f inally to human settlements.The establishment of industries like HAL, NALCO and BILT has also caused instability of the forest lands.

Habitat fragmentation poses serious threat to the species richness, their biogeographical environment and lead to their endangerment (Thiene et al. 2012). Our results showed that the size of the core forest has drastically reduced in the past three decades which can alter the species composition and frequency of many species (Forman and Alexander 1998)and it can degrade considerable share of endemic species of the Koraput such asThemeda saxicolaBor,Themeda mooneyiBor,Habenaria grandif loriformisBlatt & McCann,Habenaria panigrahianaS. Misra andDimeria mooneyiRaizada (Saxena and Brahman 1996; Dash et al. 2017). Isolation of the forest areas has been noticed in the present study with the increase of patch fragments and decrease in the large core fragments which complements the study of Li et al. ( 2010). High conversion rate from the large core to perforated and edge fragments in both the current and predicted periods have added to the vigor of fragmentation.The greater number of grids (1 km 2 ) in the lower break values and their gradual decrease towards greater break values indicates the slower and gradual forest loss and fragmentation which complements the results of Reddy et al. ( 2013),which can be rehabilitated through quick plantation actions by the forest department of the district. However the predicted period has shown increased share of patch fragments and more grids with negative change, which implies that the structure of the forest can be degraded hugely if present scenario of land uses for agricultural lands and human settlements continue. The increasing transition probability of forested lands into other land-uses indicates the anthropogenic activities like industrial establishments, mining and shifting cultivation have posed more threat on the forest fragmentation rather than the natural phenomenon such as river course change, altitude and slopes (Sulieman 2018).

Fig. 4 GWR models between FCC and change in fragment matrices ( a- e) with ranges of coeffi cient of determination

The primary driver of deforestation has been the transition of forest areas to other land uses along with the spatial variables which will lead to anthropogenic climate change.Forest fragmentation has been seen increasing in the past three decades and it will continue in the same manner unless major steps are taken to suppress the process through the acts like restriction of illegal activities of land encroachments, timber harvesting, shifting cultivation and mining.The grid based analysis has provided critical information about the micro level spatial change of forest cover and forest fragmentation which can be integrated with micro scale climatic parameters such as temperature and precipitation to know their spatial variation and response towards natural climate change. The study area showed a gradual increase in the number of patches and edge length which are two of the major parameters for quantifying fragmentation (Lausch and Herzog 2002) (Fig. S6). The northern region of the study area is under major threat as most of the landscape matrices have shown negative change along with considerable amount of deforestation making it the priority zone for conservation.Such studies on forest fragmentation have been carried out in some of the other districts of India revealing the similar results (Reddy et al. 2010; Pattanaik et al. 2011).

The local model (GWR model) has proven to be the better f it model for the study compared to the global model(OLS model) which has also been the case in the study of Tripathi et al. ( 2019). In contrast to Kozak et al. ( 2018), we conf irmed the considerable impact of forest cover change on the increase of forest fragmentation through spatial correlation. The spatial pattern of correlation has been diff erent for diff erent matrices in diff erent regions. The forest cover loss and fragmentation show strong correlation (R2 values) in the western and northern part of the district. These regions are relatively f lat plains with mostly dominated by the agricultural lands and the spatial correlation has gone weaker as the distance increased from these areas towards the eastern part of the district (Fig. 4). In the eastern zone several pockets of forested lands has shown strong spatial correlation where the major bauxite mines are operating which implies the mining activities have caused both forest loss and fragmentation simultaneously. The overall results suggest that deforestation has impacted forest fragmentation in the study area up to a considerable degree. However, forest cover change has comparatively lesser impact on NP and MPA which are two important parameters to quantify forest fragmentation,and thus more studies can be done in diff erent geographical regions for its generalization.

Conclusion

The study revealed that the rate of deforestation has been found on the higher side with the decade of 1997-2007 being the most critical year in terms of both deforestation and forest fragmentation showing the highest rate of deforestation and fragmentation in the past 30 years period. The ten years of predicted period has also shown a negative trend in case of both forest cover and fragmentation. In the studies area, the major drivers of deforestation has been the anthropogenic activities such as increase in the industrial areas,mining practices, encroachment of human settlements in the forested area and also the increase in the agricultural land uses in the past three decades. From the current study it has been conf irmed that the process of deforestation and forest fragmentation are spatially correlated with most of the landscape matrices showing positive interrelationship with the forest cover change. The northern part of the district with considerable amount of forest loss and well as fragmentation has been identif ied as the conservation priority zone.

These noble f indings of the present study using Geographical Information System (GIS) can be introduced in the forest working plans in order to locate the conservation priority areas for formulating management strategies for future plantation programmes. However being a part of community-managed forest, diminishing forest areas of the district can be regenerated at micro scales using the f indings of the current study by the active local communities managed by the district level federation. The present study also recommends that proper management strategies should be taken up in order to balance this natural resource with the social forestry probably referred as Corporate Social Responsibility (CSR), which will directly or indirectly benef it the fringe population, in terms of their social and economic stability. The national strategies such as REDD + in India which aims at reducing carbon dioxide (CO2) emission through increase of sequestered carbon in forest, should intensify its reach towards involvement of tribal regions like Koraput and sharing their indigenous strategies to conserve forest and forest products.

Acknowledgements The authors duly acknowledge the constant support of Forest Department of the district for providing appropriate permissions in the f ield. The f irst author is grateful to DST-INSPIRE for providing fellowship (Sanction No. DST/INSPIRE Fellowship/2015/IF150127 dated 10.04.2015) during the tenure of the research work.

Publisher’s NoteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affi liations.