Multi-Decadal Mapping and Climate Modelling Indicates Eastward Rubber Plantation Expansion in India

Abstract

Automated long-term mapping and climate niche modeling are important for developing adaptation and management strategies for rubber plantations (RP). Landsat imageries at the defoliation and refoliation stages were employed for RP mapping in the Indian state of Tripura. A decision tree classifier was applied to Landsat image-derived vegetation indices (Normalized Difference Vegetation Index and Difference Vegetation Index) for mapping RPs at two-three years intervals from 1990 to 2017. A comparison with actual plantation data indicated more than 91% mapping accuracy, with most RPs able to be identified within six years of plantation, while several patches were detected after six years of plantations. The RP patches identified in 1990 and before 2000 were used for training the Maxent species distribution model, wherein bioclimatic variables for 1960–1990 and 1970–2000 were used as predictor variables, respectively. The model-estimated suitability maps were validated using the successive plantation sites. Moreover, the RPs identified before 2017 and the Shared Socioeconomic Pathways (SSP) climate projections (SSP126 and SSP245) were used to predict the habitat suitability for 2041–2060. The past climatic changes (decrease in temperature and a minor reduction in precipitation) and identified RP patches indicated an eastward expansion in the Indian state of Tripura. The projected increase in temperature and a minor reduction in the driest quarter precipitation will contribute to more energy and sufficient water availability, which may facilitate the further eastward expansion of RPs. Systematic multi-temporal stand age mapping would help to identify less productive RP patches, and accurate monitoring could help to develop improved management practices. In addition, the existing RP patches, their expansion, and the projected habitat suitability maps could benefit resource managers in adapting climate change measures and better landscape management.

1. Introduction

Hevea brasiliensis (rubber) is the major source of natural rubber; it is indigenous to the Amazon rain forest, where the annual rainfall is more than 2000 mm without a marked dry season, and there are 125–150 rainy days annually, with a monthly mean temperature of 25–28 °C [1]. Due to its economic importance, it is commercially cultivated in various parts of the world in warm and humid tropical regions, even in sub-optimal growth conditions. H. brasiliensis is mostly grown in monoculture, replacing the old and native species; intercropping with various shade-tolerant species has been found to be both environmentally and economically sustainable [2]. Effective rubber harvesting and management relies on the efficacy of latex production. The maximum production is reported between six and thirty years of plantation, depending on the climate, soil, topography, and management practices [3]. H. brasiliensis plays an important role in climate change mitigation by influencing the microclimate conditions through carbon sequestration and achieving various sustainable development goals [4,5]. The ecosystem structure, species composition, and diversity largely depend on the species-specific ecological niche provided by the local hydro-climatic and physiographic settings [6,7]. Furthermore, the competitive water and energy use mechanisms of various species in a local environmental setting determine the productivity, dominance, and spread in species distribution. In addition, fire and other anthropogenic disturbances trigger modifications in species distribution and migration [8,9]. The increasing global mean temperature and anthropogenic disturbances impose severe threats of invasive species expansion and risk of environmental marginality for many native species [10]. While water stress often enables adaptation to drier conditions, temperature stress causes significant yield reduction [5]. Moreover, climate change impacts vary substantially across climate zones. Therefore, it is imperative to study how the impacts of projected climate change might drive the management practices and land use planning of economically important rubber plantations (RP).

Both periodic and systematic RP mapping and monitoring are crucial for developing appropriate management strategies in assessing forest conservation and biodiversity, evaluating ecosystem services, determining climate change impacts, etc. [11]. Moreover, long-term RP mapping and modelling for local environmental conditions can help to forecast potential shifts in the niche under projected climate change scenarios. Multi-temporal satellite imagery offers a rapid and consistent assessment of the stand-age and spatial extent of RPs [12]. RPs are mostly monoculture, resulting in spatially homogenous vegetation cover with distinct boundaries. This enables accurate identification of RPs in satellite images through visual image interpretation owing to their smooth texture, shape, and phenology [13,14]. Previous studies have employed semi-automatic classifiers such as the maximum likelihood [15,16], a neural network with Mahalanobis typicalities [17], decision tree [18], and random forest [11] in RP mapping. The accurate identification of RPs relies on their unique phenological characteristics, wherein a sharp canopy cover change is observed during the defoliation and refoliation stages [19]. The defoliation stage, or leafless period, is indicated by the low canopy cover at the beginning of the dry season, resulting in lower greenness. In contrast, the refoliation stage has a higher canopy density during the growing period [20]. Previous studies have used satellite-based vegetation, water, and fire-burned indices (e.g., Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Atmospherically Resistant Vegetation Index (ARVI), Normalized Difference Moisture Index (NDMI), difference vegetation index (DVI), Normalized Burned Ratio (NBR), etc.) with decision tree classifiers to indicate the refoliation and defoliation stages of RPs [20,21,22,23]. Chen et al. [11] adopted an integrated pixel- and object-based classification approach along with a tree growth model for mapping RP areas and stand age. Ye et al. [24] developed a time-series model for long-term RP monitoring by capturing the deforestation and afforestation stages.

In India, RP development is being practiced in the warm and humid tropical climate regimes of the Western Ghats (WG) and the northeast (NE) [25]. RP mapping in the biodiversity-rich WG and NE India is essential for examining the impact of anthropogenic disturbances and native landscape management practices such as shifting cultivation [26,27]. Few studies have attempted to map the extent of RPs in WG and NE India. Porwal and Roy [14] and Chakraborty et al. [13] each used the visual image interpretation technique to identify RPs in Kerala and a few watersheds of three NE Indian states. They used Landsat TM imagery, high-resolution aerial photographs, and LISS-III imagery for visual identification of RP patches. Meti et al. [15] used LISS-III and employed the maximum likelihood classifier to identify RPs in Kottayam district, Kerala, India. Ranganath et al. [16] used multi-temporal imageries to segregate diseased RP areas in Karnataka, India. A long-term and automated RP mapping process is called for in India.

The mathematical modelling of species distribution primarily depends on current species distribution and associated biogeographic conditions that indicate potential biotic interactions. Species distribution models (SDMs) offer robust and simple methods of assessing habitat suitability. SDM spatially examines the climate–species relationship and estimates the probability of species occurrence in each grid cell as a function of the environmental variables. Disturbances or anthropogenic activities have been reported as significant contributors, especially the distribution of species with high economic importance, such as natural rubber [28]. Various studies have used the maximum entropy (Maxent) model for species distribution modelling, including the use of both climatic and non-climatic variables in India [29,30,31]. The Maxent model allows both continuous and categorical data to be used as input variables, and provides a comparative measure of importance for each predictor variable. Tan et al. [32] employed a coupled model integrating system dynamics, Maxent, and cellular automata to simulate natural forest degradation due to the expansion of farmland, tea gardens, and RPs. Zomer et al. [33] employed a multi-model approach to examine land use and climate change impacts on the distribution of RPs. Static (e.g., Hevea) and process-based models (e.g., Land Use Change Impact Assessment tool, or LUCIA) have been used to project climate impact and estimate RP growth [34,35]. Ray et al. [28] assessed the contribution of climate factors in regulating the RPs in WG and NE India using the Maxent model. They identified suitable areas for RP expansion in NE India using the Special Report on Emissions Scenarios (SRES) climate projections. They further reported the key contribution of warmest quarter temperature variables such as annual range, mean of coldest months, seasonality, and rainfall in the warmest quarter in regulating the RP in NE India.

Several studies have been carried out on mapping RP and species distribution modeling in order to identify potential regions for expansion and climate change impacts in the Indian state of Tripura [13,15,28,31]. However, there is a lack of studies involving systematic RP monitoring and the use of such multi-temporal data in species distribution modeling. The current study maps the expansion of RPs and stand age at two- to three-year intervals for the past three decades in the Indian state of Tripura. Therefore, this study’s contribution is two-fold: first, we used Landsat imagery and the decision tree model to identify defoliation and refoliation stages and to map stand age. Such information can help land managers to identify those patches that require stand replacement in order to ensure maximum productivity. Second, we used the spatio-temporal RP sites and associated climatic variables to assess RP distribution using the Maxent model. Finally, we predicted the potential habitat suitability of RPs under two Shared Socioeconomic Pathways (SSP) climate projections. The projected habitat suitability map generated under future climate scenarios can help resource managers to develop improved land management plans.

2. Materials and Methods

2.1. Study Area

The current study was conducted in the state of Tripura, India (91°09′ E to 91°20′ E and 22°56′ N to 24°31′ N), which has a geographic area of 1,049,000 ha. The study area is dominated by forests, which occupy nearly 74% of the total area (Figure 1) [36]. The topography is characterized by rugged terrain, with an elevation ranging from 15 m to 975 m above the mean sea level. The mean monthly winter and summer temperature varies between 10 °C and 33 °C, and the annual mean precipitation (1981–2010) is 2393 mm (agartala.imd.gov.in/Tripura-Climatology/, accessed on 5 May 2022). Tripura is part of the Indo-Burmese hotspot region and has a tropical wet and humid climate supporting a rich diversity of species [37]. Cropland expansion, dynamic jhum cultivation, human disturbances, etc., have been reported as major factors of forest cover loss in this region [38,39].

2.2. Models and Datasets Used

The first step was to identify RPs over the past three decades at two- to three-years interval using a decision tree (DT) classifier. The images available from 1990 to 2017 were used for RP mapping. The images for 1990, 1996, 1999, 2002,1, 2005, 2009, 2011, 2014 and 2017, were used based on the cloud-free and striping error-free scenes data availability. Next, the identified RP patches were used in species distribution modelling to identify the suitable zones. The flowchart of the overall methodology is shown in Figure 2.

2.2.1. Decision Tree Classifier for Rubber Plantation Mapping

During the defoliation and refoliation phases, leaf fall and regrowth cause variations in green cover (low to high) in RPs. Leaf regrowth, variations in chlorophyll content, and rapid changes in leaf area during these phases create a distinct spectral signature for RPs compared to other vegetation (Figure S1). Previous studies show that the maximum leaf fall and regrowth of an RP in the tropical climate of Tripura occur during the driest period of the year [13,14].

Landsat multi-spectral satellite data from various sensors, such as MSS, TM, ETM+ and OLI, were used to delineate RP sites at two- to three-year intervals depending on cloud-free scene availability. The images for the first and second week of February indicated a defoliation stage, followed by a refoliation stage in the latter half of February. Field data (geolocation and plantations years) on RP sites in the study area were collected using a handheld global positioning system (GPS) device. In situ data included 165 observations collected from 1975 to 2017 (

Table 1. Accuracy in classifying rubber plantation sites in different years.

Actual Year of Plantation	Number of Observations	Correctly Classified within 6 Years of Actual Plantation	Correctly Classified with 6–10 Years of Actual Plantation	Correctly Classified after 10 Years of Actual Plantation	Total Correctly Classified	Misclassified	Not Identified Due to Limited Data
1975 & 1980	5	No data	4	1	5
1990	15	6	2	6	14	1
2000	6	6			6
2002	9	6		2	8	1
2005	30	10	15	3	28	2
2006	11	1	7	1	9	2
2007	18	6	10		16	2
2008	10	6	4		10
2009	13	9	3		12	1
2010	16	11	3		14	2
2014	11	5			5		6
2015	11	2			2		9
2016	6	2			2		4
2017	4	1			1		3
Column total	165	71	48	13	132	11	22

). In addition, several RP patches were visually identified from Landsat images and verified using ground reference data points. More than 100 training data points were identified from RP patches in 2017 using a visual image interpretation approach. The reference data points (a minimum of 100) for other land use and land cover (LCLU) classes, such as forest (FR), built-up (BU), cropland (CL), and water bodies (WB), were identified by employing visual image interpretation for the year 2017. The DT classifier was used for RP mapping, utilizing the phenological characteristics. Two vegetation indices, NDVI and DVI, were employed along with the temporal difference between defoliation and refoliation periods as determinant variables (Equations (1) and (2)):

$$\mathrm{NDVI}=\frac{\mathsf{\rho }\mathrm{NIR}-\mathsf{\rho }\mathrm{RED}}{\mathsf{\rho }\mathrm{NIR}+ \mathsf{\rho } \mathrm{RED}}$$

(1)

$$\mathrm{DVI}=\mathsf{\rho }\mathrm{NIR}-\mathsf{\rho }\mathrm{SWIR}1$$

(2)

NDVI Change = NDVI (Refoliation − Defoliation)

(3)

DVI Change = DVI (Refoliation − Defoliation)

(4)

where ρ_SWIR1, ρ_red, and ρ_NIR are the surface reflectance values of the short-wave infrared, red, and near-infrared bands, respectively.

The vegetation indices for various LULC classes were extracted and used to generate the box plots to compare their differential ranges (Figure 3). The threshold values of the DT classifier to exclude water bodies and built-up areas were derived using the box plots. However, the index range values for RP, forest, and croplands indicated wide overlaps. Thus, the Classification and Regression Tool (CART) was applied to segregate RPs from cropland and forests. The training data points for RP, cropland, and forests were imported into the R platform in order to develop the CART classification model, i.e., the threshold values to segregate RP from forests and croplands.

The developed DT classifier was executed in the Google Earth Engine (GEE) platform for rapid data analysis and automated RP mapping [40]. The RP patches identified in different years were verified using the ground data on RP years. The multi-temporal RP layers were then stacked to generate a stand age map with the reference baseline of 2021, i.e., patches identified in 2017 and 2014 had stand ages of four and seven years, respectively, etc.

2.2.2. Species Distribution Modelling (SDM) Using Maxent Model

The Maxent model was employed to identify the potential areas for RP expansion using a set of climatic determinants. The RP patches identified in different years were converted into point vector data and were considered as the observed species occurrence. The Maxent model was employed to assess the present climate niche and to predict the habitat under projected SSP climate scenarios. The climate variables used in the study were obtained from the Worldclim database (www.worldclim.org/bioclim, accessed on 5 May 2022), which contains nineteen bioclimatic layers at a spatial resolution of 30 arcsec (~1 km). Worldclim versions 1.4 and 2.1 were used for the years 1960 to 1990 and 1970 to 2000, respectively. Details concerning data preparation, validation, and accuracy can be found in [41]. In addition to climate variables, Shuttle Radar Topographic Mission (SRTM) data-derived digital elevation model (DEM) data (30 m spatial resolution) were used as an additional variable in the Maxent modelling. The SRTM DEM data were obtained from GEE.

2.3. Rubber Plantation Suitability Prediction

Principal component analysis (PCA) was performed in order to reduce the dimensionality of the data and to reduce the overfitting error. PCA indicated that the first four principal components contributed >96% of the total variability, wherein the first two components contributed >77% (Table S1). The factor loading of the various input variables in the first four principal components is shown in Table S2. The biolcim variables with a high factor loading value (≥0.9) were selected, as this indicated significantly high contributions in the first three principal components, explaining >90% of the total variability. These included annual mean temperature (bio_1), mean diurnal range (bio_2), temperature seasonality (bio_4), minimum temperature (bio_6), annual temperature range (bio_7), mean temperature (bio_9), mean temperature of the coldest month (bio_11), precipitation seasonality (bio_15), precipitation of wettest quarter (bio_16), and precipitation of driest quarter (bio_17).

Species distribution suitability is expressed in the Maxent model as a probability value ranging between 0 and 1, with higher values indicating more suitable areas and vice versa. Based on the availability of climate and species occurrence data, we performed SDM during three phases: (a) 1960 to 1990; (b) 1970 to 2000; and (c) SSP projected climate data for 2041–2060. A subsampling approach was adopted wherein 70% of the total reference data were used for model development, and 30% of the data were used for validation. During the first phase, the mean climate from 1960 to 1990 and the RPs identified before 1990 were employed to generate a suitability map. The RP sites identified in 1996 and 1999 were overlaid on the output suitability map to validate the output. Similarly, during the second phase, the mean climate data for 1970 to 2000 and the RPs identified before 2000 were employed (i.e., ≤1999). The plantation sites identified after the year 2000 (i.e., 2001–2017) were overlaid on the output map for validation. The model simulations were further continued to predict the suitable zones under SSP126 and SSP245 climate scenarios for 2041–2060 using the RP occurrence sites identified in 2017. Simulated climate projections from the Centre National de Recherches Meteorologiques (CNRM) were used, as this has been reported to be one of the best simulations covering South Asia, including India [42,43].

3. Results

3.1. Rubber Plantation Mapping

The spectral reflectance curves indicated prominent differences between RPs and other LULC features during the defoliation and refoliation stages (Figure S1). The spectral signature curve of RPs was similar to that of cropland (fallow on 11 February 2017) in the defoliation stage, while they had higher similarity with forests in the refoliation stage. This can be attributed to the dominance of soil background during the leaf-less conditions (defoliation stage) and high canopy cover in the refoliation stage. A lower NDVI, varying between 0.23 and 0.43, was observed in RPs during the defoliation stage (early February); the majority of the NDVI values (first and third quartile) ranged between 0.31 and 0.37, with a median value of 0.34 (Figure 3i). The NDVI for forests varied between 0.37 and 0.59, wherein the majority of the NDVI values ranged between 0.44 and 0.5, with a median value of 0.47. The NDVI values for bodies of water were observed to be negative, whereas they were below 0.1 for built-up areas. The range of NDVI values for cropland (fallow) ranged between 0.04 and 0.39, wherein the majority of the NDVI values varied between 0.15 and 0.24, with a median value of 0.2. A few outliers were observed for cropland at a higher NDVI range. RPs had a significantly higher NDVI value (ranging between 0.47 and 0.75 with a median value of 0.69) in the refoliation stage than in the defoliation stage (Figure 3ii). Moreover, the majority of the NDVI values (0.62 and 0.72 for the first and third quartile, respectively) in the refoliation stage were significantly higher than that of forests (0.55 and 0.62 for the first and third quartile). In addition, the range in NDVI values for cropland, built-up areas, and bodies of water were observed to be significantly lower than that of RPs. Furthermore, the range in NDVI values (0.16 to 0.45) for RPs showed an overlap with cropland at the lower quartile (–0.02 to 0.2), which was below 0.16 for forests, built-up areas, and bodies of water. The DVI range for RPs showed a wide overlap with the rest of the classes in the defoliation stage. The majority DVI value for RP (0 to 0.02) shows a high overlap with cropland (–0.02 to 0.08), followed by built-up areas (0 to 0.01), bodies of water (narrow range ~0.04), and forests (0.08 to 0.11) (Figure 3iv). In addition, the DVI of RP shows high overlap with forest and cropland, and the least with built-up areas and bodies of water in the refoliation stage (Figure 3v). However, the majority (first and third quartile) range for RPs shows no overlap with other classes. Moreover, the DVI change for RPs shows high overlap with cropland and forest, and no significant overlap with built-up areas and bodies of water.

In summary, the NDVI range for RPs during the defoliation and refoliation stages showed the least overlap with built-up areas and bodies of water. An NDVI threshold value of 0.2 in defoliation and refoliation stages and an NDVI change value >0.1 accurately removed/ excluded the built-up areas and bodies of water from the images. However, such an approach is not feasible for removing cropland and forests due to their higher overlaps. Thus, a multi-criteria threshold value was derived by applying the CART model using the R platform. The developed CART model identified cascaded criteria using NDVI and DVI layers (shown below) for RP mapping, with more than 95% accuracy.

NDVI (Refoliation) > 0.35 and NDVI change > 0.16 and –0.06 < DVI (Defoliation) < 0.08 and DVI Change > 0.6.

The identified threshold values were implemented in GEE in order to automate RP mapping (Figure S2) and download the classified layers. The RP stand age was derived by referring to data for 2021. A stand age of four years indicates RP patches identified in 2017 (w.r.t 2021), wherein seven years stand-age indicates patches identified in 2014, etc. (Figure 4i). The stand age map was verified based on in situ data containing the geolocation of the plantation site and the year of the actual plantation (Table 1). Out of a total of 165 reference data patches, only five data points were available on or before 1980, 15 data points on or before 1990, six data points on or before 2000, and the rest of the 139 datapoints after 2000. Four RP sites (out of five sites on or before 1980) were identified in the 1990 classified map (within 6–10 years after actual plantation), and one patch was identified in the 1996 classified map (after 10 years of actual plantation). Out of the 15 plantation sites from 1990, six patches were identified in 1996 (within 6 years of actual plantation), while only two patches were identified in 1999 (within 6–10 years of plantation), and 6 patches were identified in 2001 classified map (after 10 years of actual plantation). However, one patch was misclassified, wherein RPs could not be identified even after 10 years of the actual plantation. Similarly, the in situ data on plantation sites were verified using the identified patches from the classified images.

The NDVI and DVI change layers were created by subtracting defoliation (leaf-off condition) from refoliation (leaf-on condition), ensuring higher positive values for RP layers than for other features. However, the opposite change, i.e., subtracting refoliation from defoliation, could misclassify RPs with deforestation or shifting cultivation in northeastern India [26,27]. This approach ensured accurate differentiation of RPs from natural or anthropogenic deforestation, such as shifting cultivation.

We found that most of the patches could be identified within 6 and 6–10 years of their plantation year. However, several RP sites from plantations that were carried out during 2005–2007 could be identified within 6–10 years of actual plantation, i.e., in 2014. Such patches could not be captured in 2011 (within 6 years of actual plantation). Although most of the actual RPs were planted on or before 2005, the majority of them could be easily identified in 2014 and 2017. Moreover, it should be noted that eleven data points out of 165 in situ references were misclassified, wherein 22 recent plantation sites could not be identified in 2017. We found that the total area under RP was 819 km² in 2017. The majority of RPs were identified in the western and south-western districts of Tripura. The maximum area was identified in Sepahijala (31.12%) and South Tripura (30.46%), followed by Gomati (14.71%) and West Tripura (10.13%) districts. The areas where RPs were found least were North Tripura (2.18%), Unokoti (3.04%), Dhalai (3.33%), and Khowai districts (5.04%).

3.2. Species Distribution Modelling

The identified RP patches were used as input for Maxent species distribution modelling. The RP patches identified during different periods (on or before 1990, 1999, and during 2001–2017) were converted into point data. A random sample of points was used as input in the Maxent modelling. Due to the lower number of patches and in order to avoid spatial autocorrelation (or bias), 150 points were randomly selected from the patches identified in 1990 (Figure 4ii). In comparison, 250 and 300 randomly selected points were considered for the patches identified in 1999 and 2017, respectively (Figure 4ii). The bias files were created for different periods in order to accommodate any uneven distribution or sampling of RPs within the study site [44]. The ‘ENMevaluate’ package in the R platform was employed to generate the bias files. The bias files created for 1990 and 1999 identified the clustered and higher spatial autocorrelations in the western parts, which were comparatively low in 2017 owing to higher spatial variability in the sample locations (Figure S3). A higher spatial autocorrelation was observed in Sepahijala and South Tripura districts compared to other regions (Figure S3). Out of nineteen bioclimatic variables, along with elevation, ten variables with a factor loading ≥ 0.9 were chosen, as this indicated a higher contribution in the first three principal components. Moreover, Maxent modelling indicated a negative contribution from mean diurnal range (bio_2), temperature seasonality (bio_4), annual temperature range (bio_7), precipitation seasonality (bio_15), and precipitation of wettest quarter (bio_16). Thus, the final Maxent modelling was performed, including annual mean temperature (bio_1), minimum temperature (bio_6), mean temperature (bio_9), mean temperature of the coldest month (bio_11), and precipitation of the driest quarter (bio_17). The Maxent-derived mean area under the curve (AUC) value was 0.78 during phase I with mean climate data from 1960–1990 and RPs identified in 1990 (Figure 5). A similar mean AUC value was estimated (0.76) for phase II, including the mean climate data from 1970–2000 and RPs identified in 1999 (Figure 5). The mean AUC values were 0.63 and 0.62 for the predicted SSP climate scenarios of 126 and 245, respectively, with those RP sites identified in 2017.

The Maxent estimated probability maps for the 1960–1990 and 1970–2000 periods are shown in Figure 6. The phase I model results (climate data 1960–1990) predicted higher suitability in regions in and around the existing RP sites. In addition, we found that the higher altitude and ridges in the eastern regions are estimated to be less suitable. In contrast, the phase II results (climate data of 1970–2000) indicated comparatively higher suitability for high-altitude areas. We found that the probability values of RP sites (in situ data) extracted from the probability maps created during phase I and phase II varied from 0.3 to 0.73 and 0.32 to 0.63, with a mean value of 0.53 and 0.52, respectively (Figure 7). Furthermore, we found that the probability values of randomly selected RP sites (satellite data derived maps) during phase I and phase II had similar values (minimum, maximum, and mean of 0.35 and 0.38, 0.66 and 0.66, and 0.53 and 0.53, respectively). It was observed that >67% of the total RPs after 1990 were established in suitable zones (probability > 0.5), of which >80% of the total RPs after 1999 were established in suitable zones (probability > 0.5). The central and eastern regions indicated lower suitability for RPs under the current climatic patterns. The projected maps under SSP scenarios 126 and 245 showed similar distributions and an overall increase in habitat suitability, mostly in the central and eastern regions, with a maximum value of ~0.52 (Figure 7 and Figure 8). The increase in habitat suitability in these regions could indicate a significant rise in the temperature-related variables, along with a minor reduction in precipitation in the driest quarter (mean: –4 mm) (Figure S4). On the contrary, we found a minor decrease in the probability in the southwestern and a few pockets in the northeastern regions, which are currently dominated by RP sites. However, it should be noted that the resulting probability in such regions is well above 0.5, and as such the present RP sites would be sustainable.

4. Discussion

The NDVI and DVI value ranges of RPs during the defoliation and refoliation stages clearly distinguish water from built-up areas while showing the overlap between forest and cropland (Figure 3). A multi-criteria threshold of CART regression was able to effectively discriminate forest and cropland from RP. Most of the RP sites can be identified within 6 years of actual plantation, whereas several other patches were identified between 6–10 years after actual plantation (Table 1). The comparison with ground reference data indicated a high accuracy, wherein more than 91% of the plantations carried out before 2010 could be identified. Moreover, the RP maps showed high similarity with the LISS-IV data-derived map generated by Ray et al. [29]. The patches identified in different years showed a clustered pattern, highlighting new plantations around existing patches or in adjoining areas (Figure 4). Approximately >61% of the total plantations identified are in two districts, Sipahijala and South Tripura, with the rest in the northern and northeastern areas. The stand age map accurately identified the expansion of RPs in different regions and years, showing an overall trend of eastward expansion.

The Maxent model was able to capture the climate suitability of RPs. The mean AUC is a well-accepted value of >0.76 for the current period, whereas it was >0.62 in the projected climate scenarios. Furthermore, a comparison with the RPs planted after 1990 indicated high suitability (probability > 0.5) for >67% of the total plantations. Similarly, Maxent predicted a higher proportion of RP plantations (>80%) in suitable climatic zones after 2000. Ahrends et al. [45] studied RP expansion in southeast Asia (China, Laos, Cambodia, Vietnam, Thailand, and Myanmar), and reported that large plantations in sub-optimal or environmentally marginal zones are susceptible to climate stress and other environmental hazards, which causes low RP yield.

The projected changes in temperature (annual mean temperature, minimum temperature, mean temperature, and mean temperature of the coldest month) indicated an increase, with a minor reduction in precipitation in the driest quarter. Moreover, no significant differences were observed between SSP126 and SSP 245 for these climatic variables, leading to a similar RP suitability projection for the period 2041–2060. These findings are in parallel to those of previous studies, which have reported similar variables associated with the growth of RPs in India, including in Tripura state [29,32]. The projected overall increase in temperature with a minor decrease in driest quarter precipitation could indicate higher energy availability and sufficient water availability, leading to higher suitability for RPs in several regions of Tripura. The geographical shift and expansion of RPs indicated growth towards higher altitudinal ranges, with temperature increases reducing existing climatic barriers [43]. A similar increase in suitability has been reported in earlier studies using SRES climate projections [28,29]. One previous study predicts that in the future, RPs may expand eastward due to higher energy availability and sufficient water availability [9]. On the other hand, the projected climate change may increase the risk of environmental marginality and lead to a reduction in yields in the western and southern regions [2]. Hazir et al. [34] employed the Hevea 1.0 static model and reported that climate change positively impacted the rubber plantation potential of Malaysia due to rising temperature and precipitation. Selvalakshmi et al. [46] studied climate change impacts on RP suitability in Xishuangbanna, China, by employing the Maxent model, and reported RP range expansion under the projected climate change scenarios. Yang et al. [47] employed the LUCIA model to assess the impacts of the altitudinal gradient and climate change (Representative Concentration Pathways scenarios) on RP suitability in Xishuangbanna, China. They observed that lowland RPs showed faster growth and higher latex yield than highland RPs, as climate change projections enabled greater yield in highland RPs. Although future climatic conditions may facilitate RP expansion, excessive heat (>28 °C) would be a major limiting factor [47]. Future plantations should consider local and regional environment-tolerant species/clones for yield maximization and climate-resilient agronomic practices [5]. The multi-temporal RP occurrence and habitat suitability maps produced here can help in landscape management, sustainable land use planning, and development of policy guidelines, support decision-making systems, and help to rejuvenate degraded lands [48].

5. Conclusions

The current study used publicly available Landsat multi-spectral data for automated and long-term RP mapping. Phenological and spectral variations of RPs in the defoliation and refoliation stages are observed at the beginning of the dry season (February). The generated decision tree classifier enabled rapid and automated mapping with limited data processing resources, thereby avoiding bulk data downloads of Landsat images from Google Earth Engine. The vegetation indices and CART classifier clearly distinguished water from built-up areas and overlapping regions of forest and cropland from RPs during the defoliation and refoliation stages. Maxent showed good accuracy in predicting RP suitability for both historical and current conditions. The spatiotemporal patterns of RPs indicate expansion around existing plantation sites, then eastwards under both the SSP126 and SSP245 scenarios. Temperature increase with a minor decrease in driest quarter precipitation allocates more energy, with optimum water availability facilitating RP expansion in Tripura under future climate change conditions. The future geographical shift and expansion of RPs indicated growth towards those regions having higher altitudes and temperatures. The generated RP stand age maps enable accurate tracking of RP range expansion. Conservationists and decision-makers can utilize the developed GEE codes for automated mapping and monitoring. The current study considered climatic variables for RP prediction. However, the inclusion of non-climatic variables may further improve the model’s precision. Although the present study aimed to map RPs at regular three-year intervals, a few intervals were not mapped owing to striping errors and the unavailability of cloud-free images.