Introduction

Forests cover change has a pivotal role in global ecosystem services and environmental sustainability (Mori et al., 2017). Improved forest cover ensures the supply of terrestrial ecosystem services: provisioning, regulating, cultural and supportive services (Alfonso et al., 2017) as it delivers goods and services, sequesters carbon, improves habitat quality, and natural environment at local to global level (Paudyal et al., 2017a; Rijal et al., 2021). Meanwhile, it helps to mitigate the adverse impacts of climate change, while aiding the conservation of biodiversity with ensuring socioeconomic benefits and ecosystem services including food security (Borrelli et al., 2020). In sharp contrast, deforestation or forest degradation results in direct adverse effects, ranging from the supply of native food, fuel-wood, construction materials, loss of biodiversity as well as other indirect impacts such as the depletion of water resources and increased carbon emissions (Giam, 2017).

Change in forest cover is largely associated with several anthropogenic and natural factors (Sharma et al., 2020). Anthropogenic determinants of forest cover change include agriculturral land expansion (Solomon et al., 2018), urbanization (Browder, 2002; Nguyen et al., 2020), population pressure (Cueva Ortiz et al., 2019) deforestation, forest degradation, over grazing and forest fires (Chaudhary et al., 2016; Cueva Ortiz et al., 2019), mining (Tsai et al., 2019), government policies (Li et al., 2013) and natural factors including natural regeneration, landslides, soil erosion and flood events (Rickli & Graf, 2009) and other natural disaster (Rifat & Liu, 2020). Such factors have contributed to global forest cover decline in the recent decades: forests occupied 31.6% of Earth’s terrestrial surface in 1990 which declining to 30.6% area by 2015 (FAO, 2018).

Regeneration and reforestation programs have become a top priority globally to help maintain a sustainable environment (Löf et al., 2019). The UN Environment’s sixth Global Outlook lists 17 Sustainable Development Goals (SDG) which were developed to help sustain the planet (Flinzberger et al., 2020; Menton et al., 2020) with the 15th Goal including the sustainable management of forest resources (UN, 2015). Similarly, Forest Landscape Restoration (FLR) is widely recognized as a key goal of a range of programs including the Bonn Challenge 2011 (www.bonnchallenge.org), New York Declaration 2014 (Dave et al., 2018), REDD program (reducing emissions from deforestation and forest degradation), land degradation-neutral world (LDN), and UN decade 2021–2031 for ecosystem restoration (UN, 2019).

Despite an overall global decline in forest cover, Asia is witnessing a gradual increase in the recent decades. Regional forest increase was the highest for Asia (+ 1.2 million ha) compared to Africa (−3.9 million ha), Europe (+ 0.3 million ha), North America (−0.1 million ha), South America (−2.6 million ha), and Australia (+ 0.4 million ha) between 2010–2020 (FAO, 2020). Forest cover increase in Asia is an outcome of the collective efforts to conserve forest through various national plans and programs such as LDN and Bonn Challenge commitments upon where India, for example aims to restore 13 million ha of degraded land by 2020 and an additional 8 million ha by 2030 (Borah et al., 2018). Additionally,the reforestation initiatives of South Korea, Vietnam, and China (Choi et al., 2019); landscape restoration program of Indonesia (Van Oosten et al., 2014); Grain to Green Program (GTGP) launched in China (Feng et al., 2013); and the community forest management program in Nepal (Agrawal & Chhatre, 2006) have contributed considerably in forest restoration.

Historically, Nepal's forest were managed under- a state owned centralized system until the 1970s. However, the approach failed to sustainably manage forest resources (Wakiyama, 2004) resulting in widespread deforestation (Khatri et al., 2018). Forests occupied 45% of total land cover in 1964 (MoPE, 2001) declining to 38% by 1978/1979 (Land Resources Mapping Project (LRMP)), 35.9% by 1984 (National Remote Sensing Center (NRSC)), and to 29% by 1994 (National Forest Inventory (NFI)) and 40.36 % in 2015 (DFRS, 2015). Hence, the government changed its forest policy (Agrawal & Chhatre, 2006) by introducing a community-based forest management program in the 1980s. Forest Act, 1993 (HMGN, 1993) and the Forest Regulation Act, 1995 (HMGN, 1995) were the legal documents to legitimize it. Since 1993, Nepal has gradually handed over portions of national forest to local communities (Paudel et al., 2018) and community-based forest management plans have been successfully restoring the forest resource. Deforestation rate, which was 1.31% during 1930–1975 under the centralized management approach decreased to 0.01% during 2005–2014, under community-based management (Reddy et al., 2018). Forest cover in Nepal plays an important role in mitigating the adverse impacts of climate change and offering diversified livelihood options (Bhattarai & Conway, 2021a). Combined the aforementioned strategies and plans have been remarkably successful in preserving and growing forest resources in recent decades with forest degradation rate dropping and a gradual regeneration being observed. Community forestry has proved successful particularly in the mid-hills (Baral et al., 2018b; Niraula et al., 2013; Tripathi et al., 2020) due to the collective conservation practices (Bhattarai & Conway, 2021a). However, in the case of Tarai, (the fertile lowland plains in the south), forest cover is under extreme pressure due to population concentration and urbanization (Rijal et al., 2020), illegal logging and smuggling of high valued timber and weaker management of community forest (Gautam et al., 2004). In the Tarai region of Nepal, the population was 8.62 million in 1991 (accounting 46.7% of the national population) increased to 13.31 million ( 50.3% of national population) by 2011 (CBS, 2014). Meanwhile, the urban area which was 71.36 km2 during 1989 expanded within 327.26 km2 by 2016 and expansion has mainly occurred in the outskirts of major city centers and adjacent to major road networks in western tarai of Nepal (Rimal et al., 2020a). In addition to these factors, shifting cultivation, overgrazing, poaching and rampant excavation of sand and gravel and subsequent soil erosion and landslide events have massively degraded the forest cover of Churia, the region with young and fragile topography (DSCWM, 2012).

Forest cover is the important natural resource of Nepal and comprises 112 forest ecosystem of the total 118 ecosystems in the country, of which 12 are centered in Tarai, 14 in Churia, and 53 in Middle Mountains and High Himal/ High Mountains 38 and one others ( Kharal and Dhungana, 2018; DFRS, 2015). The forests of our study area—the Gandaki Province—are particularly remarkable in terms of touristic, environmental, ecological significance. Some form the habitats of various endangered animal and bird species are recognized as the important floristic regions where high valued herbs plants are found (PPC, 2019). The province includes major national tourist destinations globally recognized for their religious, cultural, adventurous and ecotourism significance (PPC, 2019). In Particular, the Panchase Conservation area, Annapurna Trekking route, Manaslu Circuit Trail and Dhorpatan Hunting reserve are well known destinations for forest-based ecotourism. The land cover variation of the basin ranges from the highest elevation glaciers and snow cover, grasslands, shrub and coniferous forest to sub-tropical broad-leaved forest in the Tarai region through steep slopes, rugged terrain and deep gorges in the mid-hills (Dandekhya et al., 2017). Since monitoring forest condition at different scales and forest types aids the government in improving its performance in national and international initiatives (Armenteras et al., 2017), exploring forest changes in a geographically complex and biodiversity-rich landscape (Goodin et al., 2015) is imperative.

The historical examination of forest cover at ecological and physiographic scales helps to identify the elevation-wise distribution of forest resource, vegetation composition, ecosystem characteristics, anthropogenic pressure upon vegetation and overall influence of LULC upon the environment (Gerhardt & Foster, 2002). Nepal is characterized by acomplex physiography extending from snowcapped high-mountain ranges in the north to Mahabharat Mountain range, Siwalik region to Tarai, the flat plain in the south (Bhattarai & Conway, 2021b). Several studies (Baral et al., 2018a; Oli & Shrestha, 2009; Paudyal et al., 2017b; Rimal et al., 2018) have assessed the forest cover change of some small areas within the study area and Bhattarai et al. (2009) investigated the forest cover change scenario of central Nepal during 1975–2000 using Landsat satellite data. However, wider assessment of forest resource over a longer time span at the Province level is lacking. Hence, we aim to analyze the spatiotemporal, physiographic level of changes in forest cover for the years 1996, 2006, and 2016 using Landsat time series images. We anticipate our outputs will be useful to the planners, policy makers and researchers in their formulation of effective plans and policies which can ensure the protection of basin-wide biodiversity and ecosystem function.

Methodology

Study area

Nepal is physiographically divided into five classes on the base of land form (Bhuju et al., 2007).The study area, High-Mountain (more than 4500 m covering 24.9%), Mid-Mountain (2500–4500 m covering 28.3%), Mid-Hill (1000–2500 m covering 24.3%), Chure and Mahabharat (500–1000 m covering 15.2%), and Flat Plain (hereinafter Tarai) (less than 500 m covering 7.3% area of the Gandaki River Basin ); physicographical division of the study area is done on the base of elevation.

The study area, the Gandaki River Basin is located in the Gandaki Province of Nepal, sharing its eastern boundary with Bagmati Province, western boundary with Lumbini Province and Karnali Province , northern boundary with Tibet autonomous region of China, and the southern boundary with Bardaghat Susta -East of Nawalpur and with India. Administratively, the study area integrates 11 districts (Myagdi, Mustang, Parbat, Baglung, Gorkha, Lamjung, Manang, Syangja, Kaski, Tanahun, and Nawalparasi East) (Fig. 1a, b), one (1) metropolitan city (Pokhara), 26 municipalities and 58 rural municipalities. The population of the basin was 2.19 million in 1991, 2.61 million in 2001 and 2.74 million in 2011 (Fig. 1c). Major population concentrations are in Nawalparasi (23.52% (including east and west), Kaski (17.99%), and Tanahun (11.82%) districts (CBS, 2014).

Fig. 1
figure 1

Location of the study area

Full size image

Geographically, the study area is enclosed between 27.441667 and 29.330556 N latitude to 82.878333″ to 85.201389 E longitude covering about 22,000 km2 (approximately 15% area of the country) with complex topography extending from the Tarai ± 93 masl up to the High Mountain region with a maximum of 8167 masl. The basin is characterized by multiple land use features (Pant et al., 2020), variations of hydrogeochemistry and ecology (Pant et al., 2018). The northern part of the study area consists of the Annapurna mountain range integrating the snowcapped mountain peaks (Mahhapuchhre, Annapurna first, Annaapurna second, Dhaulagiri, Nilgiri, Manaslu, Himchuli, and Lamjung Himal). Ramsar listed lakes of Nepal (60,561 ha ): study area includes (Phewa, Begnas, Rupa, Dipang, Khaste, Maidi, Nyureni, Kamalpokhari and Gunde ( 172.83 Km2 area ) as well as Tilicho—the highest lake in the world also located ). Additonally, it has incorporated 96 km2 core and 151 km2 buffer area of Chitwan National Park, 501 km2 of Dhorpatan Hunting Reserve, the entire area of Annapurna Conservation Area (7629 km2), and Manaslu Conservation Area (1663 km2) (PPC, 2019). Gandaki is the major river of the watershed with 368 small to large sub/watersheds. Regarding LULC, forest is the dominant land cover of the area and 29% of the forest cover is community managed with 3844 Community Forest Users Groups (CFUGs) , 1073 leasehold forests, 18 religious forests, one collaborative forest, one protected forest, (Panchase Conservation area), and 377 private forests (PPC, 2019).

Data

In this study, freely available terrain corrected (LIT) Landsat images (Landsat 5, hereinafter Thematic Mapper (TM); Landsat 8, hereafter Operational Land Image (OLI)) for the years 1996, 2006, and 2016 were used for the land cover analysis, and all images were collected from the United State Geological Survey (USGS) Web site https://earthexplorer.usgs.gov (USGS) (Table 1). The images were verified for accuracy. The FLAASH atmospheric model was used for image processing using ENVI environment and eight land cover classes were extracted from 93 to 8167 masl. A 30-m resolution digital elevation model (DEM) was acquired from Shuttle Radar Topographic Mission (SRTM). Furthermore, topographical data for the scale 1:25,000 and 1:50,000 were used from the Survey Department, Government of Nepal (GoN, 1998). High-resolution Google Earth images and land cover data 2010 (Uddin et al., 2015) were additional data sources.

Table 1 Dates of the Landsat time series 5 and 8 images
Full size table

Extraction of LULC

There are multiple advanced parametric and nonparametric technologies available for land cover classification (Steiner, 1970) including parametric classifiers such as maximum likelihood (ML), minimum distance (MD), and Bayesian classifiers (BCs) and nonparametric classifiers such as support vector machine (SVM), artificial neural network (ANN), and decision tree. ML classifier, MD, BC, ANN, and fuzzy classification (FC) are further described by Campbell (Campbell & Wynne, 1996). Nonparametric approaches are considered to be most appropriate for LULC analysis (Rodriguez-Galiano et al., 2012). However, thorough-going training samples or region of interest (ROI) is essential to achieve high levels of accuracy in land cover classification (Campbell, 1981; Hixson & Fuhs, 1980; Scholz & Hixson, 1979). In this study, we used SVM algorithms to extract the major land cover categories in the study area using Landsat images. To extract the land cover data of the study area, topographical data developed by Survey Department and Google Earth images were used as reference data. SVM is supervised, nonlinear, nonparametric classification method which is widely applied for the extraction of land cover change of the study area as it has higher accuracy compared with ML (Kavzoglu & Colkesen, 2009; Rimal et al. 2020b). SVM approaches was evaluated by Ortega Adarme et al. (Ortega Adarme et al., 2020) for deforestation mapping, and this approach was used for land cover change and urban monitoring (Karimi et al., 2019).

SVM approach is generally arranged into four major kernel function, such as polynominal, linear, radial function, and sigmode. In this study, the radial basic function (RBF) kernel was chosen as it usually provides better results compared with other kernels. The penalty parameter of the error 100 was assigned using ENVI software.

$$\begin{array}{ll}\left(i\right)& Linear:K\left({x}_{i},{y}_{i}\right)={x}_{i}^{T}\cdot {x}_{j},\\ \left(ii\right)& Polynomial:K\left({x}_{i},{y}_{i}\right)={{(g.x}_{i}^{T}\cdot {x}_{j}+r)}^{d},g>0,\\ \left(iii\right)& Radial\ basis\ function:K\left({x}_{i},{y}_{i}\right)={e}^{-g{\left({x}_{i}-{x}_{j}\right)}^{2}},g>0,\\ \left(iv\right)& Sigmoid:K\left({x}_{i},{y}_{i}\right)=\mathrm{tanh}{(g.x}_{i}^{T}\cdot {x}_{j}+r)\end{array}$$
(1)

where xi, yi are training vectors, g, d, and r are user-controlled parameters of kernel function.

Land cover analyses for the years 1996, 2006, and 2016 were based on the classification scheme developed by Anderson (1976) and eight major LULC classes were identified: urban/built-up, agriculture, forest, shrub, grass land, barren land, water body, ice, and snow cover (Table 2).

Table 2 Land use/cover classification schemes
Full size table

Land cover change trajectories

A transition matrix of the land cover map of period 1 and period 2 was prepared using Land Change Modeller of IDRISI software (https://clarklabs.org/). The LULC transition statistics show the change of attributes from period 1 to period 2. This LULC change considers two paths (Yadav & Ghosh, 2019) and change analysis shows the pixel based change amount from one class to another during the study period.

Physiographic zone and forest type

First, elevation data was prepared using SRTM DEM (30 m resolution), and this data was used to prepare the elevation-wise physiography-level land cover change data for the year 1996, 2006, and 2016. Furthermore, forest cover within the study area was broadly categorized into 21 forest types. The forest type data (GIS vector data) of the study area was collected from ICIMOD (https://servir.icimod.org/datasets), and change for each forest type was assessed for the study period using modified data. We extracted the forest class of our classified images of different years (1996–2016) using collected forest type layer and analyzed the difference regarding the changes in forest type. Forest-related information at provincial, district, and local level were acquired from various non/governmental sources. In addition, various key government policies were evaluated.

Accuracy and uncertainty

The land cover classification accuracy assessment is significiant when land cover maps are prepared using satellite images (Feng et al., 2017; Jensen, 1996; Sexton et al., 2013). The confusion matrix method is widely used for the assessment of land cover classification (Foody, 2002) as is user’s accuracy (UA), producer’s accuracy (PA), and overall accuracy (OA). Here, accuracy assessment was prepared based on the GPS points collected from field verification conducted on 2016 and 2018. Accuracy assessments were aided by the use of topographical maps developed by the Survey Department of 1998 (scale 1:25,000 and 1:50,000) (GoN, 1998), land cover maps of 1990 and 2013 (Rimal et al., 2015) of the Seti river watershed, and high-resolution Google Earth images (http://earth.google.com). For the accuracy assessment, a total of 2382 stratified random sample points for each year were developed in the already classified land cover maps for 1996, 2006 and 2016 and were further verified in high-resolution Google Earth images and land cover data 2010 (Uddin et al., 2015). The available Google Earth images were printed in A1 size for field verification. Additionally, hand GPS was used for further collection of the sample points. The acquired sample points were used for accuracy assessment. Classified images were overlaid in Google Earth and observed for verification.

At least 200 sample points for each LULC class were represented with many LULC that covered larger areas having many more sample points (see Table 6). Accuracy assessment was conducted on sample points using GPS during field verification conducted in 2016 and 2018. The confusion matrix was generated based on the existing ground truth reference data and classified images.

Producer’s accruracy

$${N}_{ii}+{N}_{+i}n\ 100\%$$
(2)

where Nii = total number correct cells in a class and N+i = sum of cell values in the column.

User’s accruracy

$${N}_{ii}+{N}_{i+}n\ 100\%$$
(3)

where Nii = total number correct cells in a class and Ni+ = sum of cell values in the row.

Overall accruracy

$$X/Nx\ 100\%$$
(4)

where X = total number correct cells as summed along the major diagonal and N = total number of cells in the error matrix.

Land cover classification accuracy can be determined by training sample points developed during the land cover classification process (Prestele et al., 2016). Various factors can play a role in this error. For the accuracy assessment, collection of high-resolution real-time data can be a major challenge (Bhattarai & Conway, 2021c). Sample size affects the magnitude of margin of error. The larger the sample, the smaller the margin of error. On the other hand, the low Producer’s accuracy for urban areas can be one of the reasons for the increase margin of error in this class. We recognized that in this study, there are a few limitations. First, we limited our classification to only eight major LULC classes which limits the mapping of the major ecosystems of the basin. Second, lack of updated reference data and the very complex landscape also limits the classification accuracy. For example, in the northern part of the project area, there were some seasonal fluctuations between the areas of grass land, barren land and snow cover which we did not assess. Land cover classification were also impacted by regional weather and climate (Ge et al., 2019). Thirdly, we could not assess urban areas into sub-categories such as garden and urban forest due to the limited resolution of the Landsat images (Rimal et al., 2019).

In the current study, we performed stratified random sampling to determine the accuracy of classification. In this way, it is possible to estimate the area of the LULC classes by adjusting the area for the mapping error (Gallaun et al., 2015). Confidence intervals were extracted to assess the uncertainty of the accuracy measures and the area estimates of all classes. This unbiased estimator of the area proportion covers the area of map omission error and eliminates the area of commission error (Costa et al., 2018). Using the information directly provided by the error matrix and the Eqs. 5, 6, 7 and 8 (Olofsson et al., 2013), we adjusted the biased results.

$${\widehat{A}}_{j}={A}_{tot}\times {\sum }_{i}{W}_{i}\frac{{n}_{ij}}{{n}_{i+}}$$
(5)

where Aj is unbiased estimator of the total area, nij is the number of points of category j which mapped as category i, ni+ is the total number of points related to category i, Wi is the proportion of the area mapped as category i, and Atot is the total mapped area. The estimated standard error of the estimated area proportion is:

$$\left({\widehat{P}}_{j}\right)=\sqrt{{\sum }_{i=1}^{q}{{W}_{i}}^{2}\frac{\frac{{n}_{ij}}{{n}_{i}}\left(1-\frac{{n}_{ij}}{{n}_{i}}\right)}{{n}_{i+}-1}}$$
(6)

The standard error of the error-adjusted estimated area is:

$$S({\widehat{A}}_{j})={A}_{tot}\times S({\widehat{P}}_{j})$$
(7)

A 95% confidence interval for Aj is:

$${\widehat{A}}_{j}\pm 1.96\times S({\widehat{A}}_{j})$$
(8)

To assess the accuracy for the post classification change analysis of forest class, we overlaid two maps to produce the 1996–2016 forest change map. Accordingly, a stratified random sample was selected as the reference of land cover for forest class to assessing the accuracy using Google Earth images, TM, and Landsat images. Then, the abovementioned equations were applied to evaluate the error-adjusted area of deforestation.

Results

Accuracy and uncertainty assessment

An assessment of the accuracy measures and estimated areas suggests that the classification of most classes was highly accurate (Tables 3, 4, 5, 6, and 7). For example, the mapped area of forest class for years 1996, 2006, and 2016 km2 is 7571, 7673 km2, and 7778 km2, respectively; whereas the stratified error-adjusted area estimate of forest area is only slightly less (i.e. 7404 km2, 7581 km2, and 7577 km2) (Tables 3, 4, and 5). The confidence interval quantifies the uncertainty associated with the sample-based estimate of the area of different classes. Accordingly, the true area of forest for 2016 could be as low as 7410 km2 or as high as 7744 km2 at the 95% level of confidence (Tables 3, 4, 5, 6, and 7).

Table 3 The mapped and estimated adjusted areas with a margin of error (95% confidence interval) for 1996
Full size table
Table 4 The mapped and estimated adjusted areas with a margin of error (95% confidence interval) for 2006
Full size table
Table 5 The mapped and estimated adjusted areas with a margin of error (95% confidence interval) for 2016
Full size table
Table 6 Accuracy measures based on error matrix of sample counts for 1996, 2006, and 2016
Full size table
Table 7 Accuracy measures based on error matrix of estimated area proportions (with a 95% confidence interval) for 1996, 2006, and 2016
Full size table
Table 8 Error matrix for the 1996–2016 change map of forest class. Accuracy measures are presented with a 95% confidence interval
Full size table

The accuracy assessment based on the stratified estimators showed that the forest area had an overall accuracy of 91%, 93%, and 96% and a user’s accuracy of 96%, 97%, and 96% for years 1996, 2006, and 2016, respectively (Table 7). Also, the producer’s accuracy is above 95% for forest class, except for 1996 which is 90%. So, the omission error of forest class does not have a strong influence on the estimated area of forest. Accordingly, map error is small and the area mapped of forest class is close to the true area. If the producer’s accuracy was too low, it would have alerated to the problem of omission error associated with the forest category. Totally, the difference between the biased and unbiased overall accuracies was less than 0.03 in all years (Tables 6 and 7).

The obtained user’s accuracy of the deforestation class through change detection procedure was about 84% which shows the forest change obtained by post classification is acceptable. Also, forest/non-forest maps were highly accurate with user’s accuracy more than 90% (Table 8).

Details of LULC changes and the spatial distribution of LULC classes are presented in Table 9 and Fig. 2. Major changes observed during 1996–2016 include increases in urban/built-up, forest cover and barren land, sharp declines in cultivated land, and fluctuations for grassland, water bodies, barren land and shrub areas. Built-up land expanded in this period, increasing from 53 to 141 km2. Grass land was converted into forest and forests increased by 207 km2 over the 20-year period. Cultivated land faced intense pressure due to the rapid urbanization below 999.99 m elevation zone. However, forest area decreased in the < 500 m elevation zone. Forest encroachment was observed mainly in the city outskirts (Appendix Fig. 6).

Table 9 Land use/land cover change of the study area during 1996–2016 (in km2)
Full size table
Fig. 2
figure 2

Land use/land cover changes in the Gandaki Basin during 1996–2016

Full size image
Fig. 3
figure 3

Transition map. (a) Land cover transition from all land cover to forest cover 1996–2006 and 2006–2016. (b) Land cover change 1996–2006. (c) Land cover change 2006–2016

Full size image

During 1996–2006, forest cover increased by 102 km2, from 7571 to 7673 km2 with losses of 37 km2 cultivated land, 19 km2 shrub, 38 km2 barren land, 11 km2 water body, 142 km2 grass land, and 16 km2 snow/ice cover. Shrub area increased from 624 to 727 km2 as 60 km2 was converted from cultivated land and 89 km2 from forest (Appendix Table 12). A transition map of the year 1996 and 2006 and 2006 and 2016 highlights these changes (Fig. 3a–c).

Based on the land use transition matrix grass land declined by 84 km2 (from 2434 to 2349 km2), mainly due to the conversions into cultivated land (21 km2), forest (142 km2), barren land (52 km2), water body (25 km2), and ice/snow cover (5 km2). The land cover change matrix during 2006–2016 (Appendix Table 13) shows that forest area increased by 105 km2 (from 7673 to 7778 km2). Major factors contributing to the increase in forest cover are conversions from cultivated land (277 km2), shrub (75 km2), barren land (66 km2), water body (15 km2), grass (115 km2), and snow/ice (19 km2) into forest cover. Meanwhile, barren land increased by 292 km2 (from 3941 to 4233 km2) due to the conversions from snow/ice (628 km2), grass (310 km2), and forest (40 km2) (Fig. 3a–c). Shrub area declined by 92 km2 (from 727 to 635 km2) due to the conversion of 60 km2 shrub area into cultivated land, 75 km2 into forest cover, and 42 km2 into grass.

Forest cover change based on physiographic zone

The Tarai and the Mid-Hill zones are mostly cultivated and forest land. Forest, barren land, and grass are the dominant land uses in the mid-mountain zone and most of the high-mountain zone is covered by barren land and snow. Most importantly, extensive deforestation occurred during 1996–2006 which contributed in the decline of forest area from 625 to 558 km2 and doubled the shrub area by 78 to 155 km2 (Appendix Table 14) in Tarai < 500 meters. Forest encroachment and forest decline was observed mainly in areas where there was infrastructure development and urban development. However, during 2006–2016, forest cover decline was limited to 1 km2 because of the strong contribution of community-based forest management plans and Chure Conservation program of the Government of Nepal to combat deforestation and degradation.

In Chure and Mahabharat (Siwalik), increases in urban area and forest cover and decline in cultivated land area were the notable transformations during 1996–2016. Forest cover area increased by 1 km2, from 1399 to 1400 km2 during 1996–2006 and a further 49 km2 during 2006–2016 (Table 10). We observed that the major changes were driven by the community forestry program and increases in private forest in abandoned cultivated land in the rural areas due to out-migration. Decline in cultivated land and increases in urban/built-up and forest area were the major changes for the Mid-Hill zone. Urban area increased in both time periods 1996–2006 and 2006–2016 (Appendix Table 14). Forest cover increased by 14 km2 during 1996–2006 and 56 km2 during 2006–2016 (Table 10, Fig. 4) in mid-hill. In the mid-mountain zone, the majority of the area is occupied by forest, barren land, and grass. According to our analysis, forest cover had occupied 2436 km2 in 1996 which increased by 154 km2 (2590 km2)in 2006 but slightly dropped (−3 km2) and totaled 2587 km2 by 2016 (Fig. 5a–o). The High mountain zone is mainly occupied by barren land, snow cover and grass land. However, the northern-most part of the study area is mostly periodically covered by snow. Seasonal snowfall means that coverage of ice/snow, grass land, and barren land fluctuate.

The total forest area across all zones increased from 7571 km2 in 1996 to 7673 in 2006 and 7778 km2 in 2016 (Appendix Fig. 7 and Table 14). Most of these increases were in the Siwalik, mid-hill, and mid-mountain zones, but forest cover declined in the Tarai region from 8.36% in 1996 to 7.2% to 2016 (Table 10).

Table 10 Physiographic distribution of forest cover change during 1996–2016 (in km2 and percentage)
Full size table
Fig. 4
figure 4

Physiographic distribution of forest cover area in the study area during 1996–2016

Full size image
Fig. 5
figure 5

Physiographic distribution of forest cover and other land uses in the Gandaki basin during 1996–2016 (Tarai (a–c); Chure and Mahabharat (d–f); mid-hill (g–i); high-hill (j–l); and high-mountain (m–o))

Full size image

Ecological distribution of forest resource

We identified 21 different forest types in the study area (Appendix Fig. 8) which Schima-Castanopsis and hill sal forest were the two most widespread. During 1996–2006, there were notable increases in some of the forest classes including birch-fir-blue pine-cypress, birch-rhododendron forest, Schima-Castanopsis, temperate mountain oak, trans-Himalayan forest, fir forest, and upper temperate blue pine forest. In northern part, birch-fir-blue pine-cypress increased from 332 km2 in 1996 to 395 km2 in 2006. Similarly, trans-Himalayan forest expanded by 40 km2, from 51 to 91 km2, and birch-rhododendron expanded by 15 km2, from 245 to 260 km2. Schima-Castanopsis, which had occupied 1949 km2 which expanded by 16 km2 and temperate mountain oak forest increased from 1006 km2 to 1017 km2 by 2006 (Table 11). In contrast, there were declines for hill sal forests, lower tropical sal, and mixed broad leaved forest, Khair Sisoo riverine, mixed blue pine oak forests. Hill sal forest declined from 1299 km2 in 1996 to 1246 km2 by 2006. Similarly, lower tropical sal and mixed broad leaved forest declined by 14 km2 from 162 to 148 km2 and continued to decline from 2006 to 2016 but the rate of decline was low. On the other hand, alder forest and lower temperate oak forest remained almost constant.

Table 11 Changes in areas of different forest types during 1996–2016 (in km2)
Full size table

During 2006–2016, alder forest, hill sal forest, Schima-Castanopsis, and upper temperate blue pine witnessed increase whereas trans-Himalayan zone, fir forest, birch-fir-blue pine-cypress forest experienced some declines.

Alder forest increased from 118 to 132 km2 whereas hill sal forest increased from 1246 to 1276 km2. Schima-Castanopsis increased by 49 from 1966 km2 to 2014 km2. Upper temperate blue pine increased from 156 to 167 km2. Trans-Himalayan forest decreased by 6 km2, from 91 to 85 km2. Birch-rhododendron declined from 260 to 251 km2. Fir forest declined from 500 to 496 km2 while birch-fir-blue pine-cypress had decreased by 10 km2 during 2006 to 2016.

Discussion and conclusion

Many factors could have contributed to the increase in forest cover change in the hill and mountain regions, and these include (a) improved institutional mechanisms (Ministry of forests, departments, regional forest offices and district forest offices); (b) sectoral plans and policies; (c) community, leasehold, and collaborative forestry programs (at least 3,844 CFUG); and (d) other factors such as abandonment of agricultural land, partnership with donor agencies, and collaborative actions with conservation partners such as IUCN, UNDP, WWF, and ICIMOD (Ghimire et al., 2018). REDD + program to reduce deforestation and forest degradation, sustainably managing forests resource while conserving and enhancing forest-based carbon stocks (MoFSC, 2015). Furthermore, the Government of Nepal introduced a forestry decade (2014–2024) with the motto of “one house one tree, one village one forest, one city several gardens” targeting in particular the restoration and planting of at least 26,000 hectares of forest in the Tarai, Siwalik, and Hill regions. Consequently, a scientific forest management program was created aiming the protected forest area by around 0.2 million hectares (DFRS, 2016). Similar successful forest restoration programs were observed in India (Borah et al., 2018), China, South Korea, Vietnam (Choi et al., 2019), Indonesia (Van Oosten et al., 2014), and Africa (Goffner et al., 2019).

In this study, we have explored the spatiotemporal, physiographic and ecological changes in forest within the Gandaki river basin during 1996–2016 and identified the increases in forest cover in all the regions except Tarai (< 500 m). A common trend in Nepal as previous reported by Bhattarai and Conway (2021d), in the study area, is the out-migration in the mid-hill and mountain regions,has resulted in previously cultivated lands being left fallow and turning into other forms of vegetation cover (Bhattarai & Conway, 2021e).

The trend in increasing also seen in the national level from 38% in 1978/1979 (LRMP) to 40.36% in 2015 (DFRS, 2015). Furthermore, this trend is also replicated in some smaller clusters within the study area such as Kaski district (Bhandari et al., 2019), Tanahun district (Oli & Shrestha, 2009; Shrestha, 2015), CHAL region (Subedi, 2018), and Phewa watershed (Paudyal et al., 2017b). In Phewa lake watershed, forest cover has increased by 12% since the 1970s mainly due to the community-based forest management and regeneration programs (Besseau et al., 2018). Subedi et al. (2018) found an increase in forest cover by 57.2 km2 in the 12 districts of Chitwan Annapurna Landscape (CHAL) (Baglung, Dhading, Gorkha, Gulmi, Kaski, Lamjung, Manang, Mustang, Myagdi, Parbat, Syangja, and Tanahu districts) during 2000–2010.

The National Biodiversity Strategy and Action Plan provides the strategic roadmap for biodiversity conservation of Nepal (Rai et al., 2016). National level plans prioritize preservation of forest, control of forest fire and invasive species, community-based integrated forest management for water, wildlife, conservation of endangered species, wetland and riverine forest conservation and agroforestry (MoFSC, 2015).

In addition, there has been reduced risks of environmental degradation and watershed destruction and improved landscape regeneration (Paudyal et al., 2017b). Much of this is because of controls on illegal logging, encroachment and forest fires (Pokharel & Nurse, 2004), dissuading farmers from open grazing, reduced pressure upon the community forests (Upreti, 2001), and preservation-oriented forest operational plans (Kimengsi et al., 2019).

Emigration to the Tarai and peri/urban areas due to personal insecurity aroused by political conflict (1996–2006) and people’s quest for better quality of life, economic opportunities, and public service accessibility has resulted in abandoned cultivated land in hill and mountain region of Nepal (Adhikari et al., 2019; Jaquet et al., 2016, 2019; Khanal & Watanabe, 2006; Rai et al., 2019). Of the total cultivated land, 24% was abandoned in the study area where private forest cover developed (PPC, 2019).

Shifts to alternative sources such as liquefied petroleum gas (LPG) and electricity (Paudyal et al., 2019) for cooking has reduced people’s dependency upon forest aiding the increase in forest cover. The consumption of LPG gas was increased and imported 77,594 ton in 2004/2005 and 258,299 ton in 2014/2015 from India (Bhandari & Pandit, 2018).

The national-level plantation program of rehabilitation and conservation in degraded and denuded areas is still ongoing with trained and motivated NGOs and community-based organizations (CBOs) contributing. Funding has come through a number of bilateral and multilateral institutional arrangements including DFID/USAID, GIZ, and ADB. UK-funded DFID has largely contributed to the restoration of 1.56 million ha in the denuded hills and degraded forests through the Plantation and Forest Management Research and Extension Program (Tamrakar & Mohans,2013; DFRS, 1999). New plantations have been developed in 204.28 ha in the barren land in the study area during 2019 (PPC, 2019).

Protecting primary forests is essential as they maintain ecological functions, carbon storage, and environmental equilibrium. The increasing forest cover observed for the study area go a long way to help meet to the national and global agenda of reforestation along with SDG 15. However, utilization of the forest ecosystem for poverty alleviation and local livelihood enhancement still remains as a challenge. For this, forest management, forest-based ecotourism and other enterprises map provide new opportunities.

Despite widely reported global forest losses and degradation, other evidence of forest restoration at the global, regional, and national levels indicates that the historical momentum is moving toward balanced ecosystems and the natural environment (Jacobs et al., 2015). The novelty of our study lies in the fact that we have used regional LULC data and forest cover data on ecological and physiographic zones to provide the details of changes in a mountain to lowland watershed, and we suggest it can be replicated to provide a scenario analysis of forest restoration elsewhere in the world at regional and local levels.