Agroforestry Suitability for Planning Site-Specific Interventions Using Machine Learning Approaches

Abstract

Agroforestry in the form of intercropping, boundary plantation, and home garden are parts of traditional land management systems in India. Systematic implementation of agroforestry may help achieve various ecosystem benefits, such as reducing soil erosion, maintaining biodiversity and microclimates, mitigating climate change, and providing food fodder and livelihood. The current study collected ground data for agroforestry patches in the Belpada block, Bolangir district, Odisha state, India. The agroforestry site-suitability analysis employed 15 variables on climate, soil, topography, and proximity, wherein the land use land cover (LULC) map was referred to prescribe the appropriate interventions. The random forest (RF) machine learning model was applied to estimate the relative weight of the determinant variables. The results indicated high accuracy (average suitability >0.87 as indicated by the validation data) and highlighted the dominant influence of the socioeconomic variables compared to soil and climate variables. The results show that >90% of the agricultural land in the study area is suitable for various agroforestry interventions, such as bund plantation and intercropping, based on the cropping intensity. The settlement and wastelands were found to be ideal for home gardens and bamboo block plantations, respectively. The spatially explicit data on agroforestry suitability may provide a baseline map and help the managers and planners. Moreover, the adopted approach can be hosted in cloud-based platforms and applied in the different agro-ecological zones of India, employing the local ground data on various agroforestry interventions. The regional and national scale agroforestry suitability and appropriate interventions map would help the agriculture managers to implement and develop policies.

1. Introduction

Agroforestry is a combination of woody plants (timber, clump, palm, bamboo etc.) and agricultural crops (annual or perennial) and/or livestock, which are arranged in both temporal and geographical patterns on the same land-management unit [1,2,3,4]. Agroforestry helps to alleviate poverty [5], slow land degradation [6], improve food security [7,8], enhance soil fertility [9], improve the quality of the agroecosystems [10], protect biodiversity, improve air and water quality, etc. [5,9,10,11,12]. Executing agroforestry ventures through a landscape perspective has a strong capability for consolidating nature conservation objectives into farming systems [13,14]. According to the Food and Agricultural Organization (FAO) [15], 1.3 billion people are involved in agriculture, and most of them are from developing countries. The Global Agroforestry Network (GAN) is a new initiative of scientists working on cocoa, coffee, and rubber agroforestry systems to understand the ecological, social, and economic consequences of production decisions (www.globalagroforestrynetwork.org/ accessed on 20 September 2021). A global assessment revealed that presently, 10.23 million sq. km of land is used for different agroforestry purposes. Recent studies have estimated that agroforestry areas could be increased by adding 6.3 million sq. km of currently degraded land, grassland, and unproductive land [16,17].

Despite the increased yield in the agricultural crop, relative income per farmer needs special attention, as over 70% of the country’s population is dependent on agriculture for their livelihood. Again, 90% of the small and marginal farmers depend on monsoon, wherein 55% of India’s cultivated lands are rainfed. Previous studies have indicated that most of the agricultural lands in India are prone to various natural disasters [18]. India has committed to increasing the agroforestry area to 53 million ha by 2050 by restoring pastures, groves, fallows, cultivable fallows, and soil rehabilitation [19]. Agroforestry can help achieve several SDGs. One of the significant contributions of agroforestry to the rural economy is the livelihood impact, both income and employment generation [20]. Agroforestry employs relatively lower investment, including for the unskilled rural areas. According to Dhyani et al. [21], agroforestry systems have the potential to generate employment opportunities of 450 person-days per hectare per year in India.

Therefore, assessing the potential land for agroforestry is essential for planning and interventions. The land suitability assessment requires both qualitative and quantitative datasets [22], wherein the qualitative approaches rely on farmers’ understanding [23], and the quantitative approaches include parametric methods and statistical analysis [24]. Several indicators of environmental and socioeconomic conditions are employed for agroforestry suitability analysis. The geographical information system (GIS) platform has been proven as an effective data integration tool for developing an agroforestry potential map using various parameters, such as soil, climatic, topographic and socioeconomic factors [25]. Ahmad et al. [26] followed the FAO guidelines to assess agroforestry suitability in a district of Jharkhand state, India, where they employed various parameters on climate, topography and soil nutrient availability. A more recent study by Ahmad et al. [27] assessed agroforestry suitability in South Asia, including various environmental data layers on climatic, edaphic, topography, and ecological indicators. Moreover, the land use land cover (LULC) map is critical in the agroforestry potential zone analysis. According to Nath et al. [17], compared to the generic national level LULC classifications scheme, the regional/local classifications are more accurate and useful. The GIS-aided multi-criteria evaluation (MCE) technique is used extensively in agroforestry suitability evaluation [28,29,30]. Nath et al. [17], in their study in the eastern Indian Himalayan region, applied MCE to estimate the agroforestry suitability. One of the critical components of suitability analysis is the variable selection and their weight estimation. Various approaches have been tested in weight estimation, starting from a pair-wise comparison matrix to the parametric method [31], ordered weighted mean [32], and membership approach [33,34]. Researchers have also applied regression-based analysis and principal component analysis (PCA) for weight estimation [35]. A few studies have used machine learning methods for site-specific suitability assessment [36,37,38,39]. Ahmad et al. [40] evaluated agroforestry suitability by utilizing the weights and ranks approach following the FAO guidelines integrating soil nutrient parameters (N, P, K, C, pH, S), climate (wetness) and topography (slope and elevation). Singh et al. [41] used the MCE approach and GIS-based simple water balance model (SWBM) to assess and project the suitability zones for orchards employing the projected climate data.

Both the categorical and numerical data are used in the site suitability analysis. The categorical data generates a rigid boundary, wherein the impact of numerical data is continuous, i.e., the suitability either increases or decreases with increasing value or vice versa. Categorizing the numerical or continuous data is an important step in the multi-criteria analysis. Zadeh developed fuzzy logic in 1965, which is widely used for data generalization [42] and simplification by applying classic logic with rigid boundaries [43]. It applies a membership function that assigns a grade ranging between zero and one [42], representing the non-membership and full membership, respectively, and the intermediate values measure the degree of closeness of the unity [44]. The membership function uses various mathematical models, such as parabolic, sigmoid, inverted sigmoid, and linear functions. Fuzzy logic has been applied in various studies, including water quality model development and land suitability analysis [45,46,47,48,49,50], agricultural suitability [47,51,52,53], etc.

The RF machine learning algorithm is a non-parametric ensemble decision tree classifier. This algorithm randomly selects a subset of input features to grow a decision tree and generates new training sets by randomizing the original training set. The algorithm employs ensembles of trees (by generating random vectors for each tree), and a vote cast by each tree is used to define the class, resulting in considerable gains in classification accuracy. In this approach, a subset of the training data is picked randomly to construct a decision tree, and the remaining data are used to estimate the out-of-bag (OOB) error for each tree. The split is determined by randomly selecting a set of predictor variables at each tree node. Hundreds of trees are constructed similarly, and lastly, fresh data are projected by aggregating all of the trees’ forecasts [54]. RF ranks the factors and gives an independent measure of prediction error [55]. RF remains unaffected by noise in the data and overfitting errors [56]. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them [56]. RF is widely used in both classification and regression problems, including site suitability analysis. For example, Senagi et al. [57] and Taghizadeh-Mehrjardi et al. [58] employed the RF algorithm to assess land suitability for agricultural production. Ogunde et al. [59] used the RF algorithm to develop a web-based decision support system for evaluating soil suitability for selected cassava cultivation farmlands in Nigeria. Several studies have demonstrated the applicability of various machine learning algorithms for site suitability analysis, including afforestation planning [36,60]. Lahssini et al. [61] applied the RF algorithm to predict Cork Oak suitability and develop a relationship between species richness and environmental and social factors. However, agroforestry suitability analysis employing machine learning models in India is unattempted to date to the best of our knowledge.

The past studies on agroforestry suitability in India mostly used MCDA-based approaches wherein weight was assigned by experts to accommodate the influence of various indicators. Such an approach could incorporate bias owing to the experts consulted or the interpreter’s knowledge. This study included ground data on agroforestry interventions (that are based on local knowledge) to identify the impact of various parameters. The current study used the agroforestry sites (geolocation) to develop a machine learning model for site-suitability analysis employing 15 sub-criteria. The determinants include soil (type, depth, texture, drainage, pH), topography (slope), total rainfall, temperature (annual mean, maximum and minimum), and proximity parameters (distance to settlement, road, water, cropland, and forest). The RF model was applied to estimate agroforestry suitability, which remains unbiased to correlated multiple input variables. The developed machine learning is applied to estimate agroforestry suitability in the entire study area.

2. Study Area

The current study area, the Belpada block of Bolangir district, Odisha state, is geographically located between 20.46° N–20.77° N latitude and 82.81° E–283.13° E longitude. The total area of the block is 3810.63 km², and the average altitude is 383 m (Figure 1). The region’s average temperature varies between 21 °C and 34.9 °C, and the average annual rainfall is ~1600 mm (https://balangir.nic.in/agriculture/ accessed on 20 September 2021) [62]. The Belpada block is an agricultural land dominated landscape [63]. The majority of the population in the block is dependent on agriculture and allied sectors, and paddy is the major crop in the Kharif season [64]. Several initiatives are now underway in various regions of India, particularly the state of Odisha, to enhance agroforestry. The World Agroforestry (ICRAF) project on “Enabling smallholders in Odisha to produce and consume more nutritious food through agroforestry systems” funded by the government of Odisha has successfully implemented various agroforestry interventions in the Belpada block [65].

3. Materials and Methods

3.1. Satellite and Field Data

The soil resource database of the National Bureau of Soil Survey and Land Use Planning (NBSS & LUP), Govt. of India, was used as the base soil layer, which includes various soil physical and nutrients parameters. The climate data were taken from the Worldclim database (www.worldclim.org accessed on 20 January 2021), which provides gridded data at 30 arcsec (~1 km) resolution, representing the climate. The long-term minimum temperature and maximum temperature data were used in the current analysis. The high-resolution Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) rainfall data were employed, wherein the annual precipitation was considered. The slope map was derived using the Alos PalSAR digital elevation model (DEM) at 12.5 m spatial resolution. The LULC map at 10 m spatial resolution generated by Singh et al. [63] was used in the current study. All the data were brought to 10 m resolution using the nearest neighborhood resampling technique.

The ground data on various agroforestry interventions were collected during the field visits (Figure 2 and Figure S1). The majority of the agroforestry interventions included wadi (agro-horti-forestry), bund and boundary plantation, and home gardens. The field data from 200 samples included the species information and geolocation of single (142) species (Eucalyptus, Mango, Mahua, Teak) and mixed (58) species (Teak, Gamahari, Khajur, Mahua, Chahar) (

Table 1. Ground data.

Class	No. of Points	Species
Agroforestry Single species	142	Eucalyptus, Mango, Mahua, Teak
Agroforestry Mixed species	58	Teak, Gamahari, Khajur, Mahua, Chahar, etc.

, Figure 2 and Figure S1). The LULC map was converted into point data, and the random sampling tool in ArcGIS software was used to collect 200 non-agroforestry data points.

3.2. Selection of Variables for the Agroforestry Land Suitability Analysis

The first step in tracing suitable areas for agroforestry is to identify influencing variables. The local environmental factors that regulate the growth and production of agroforestry in this region were chosen, including soil, topography, climate and socioeconomic [46,66]. The variables used in the current study consist of 15 parameters under the four broad criteria. The list of variables, their nature and source are given in

Table 2. List of variables used in the study.

Criteria	Variables	Unit	Agroforestry Classes/Fuzzy Scale for Continuous Data				Data Sources
			Very High	High	Moderate	Low
Soil	Soil type	Code	Fine loamy, clayey	Fine mixed	Course loamy	Clay sandy, Sandy	NBSS & LUP
					Sandy loam
	Depth		Mod deep	Mod shallow	Shallow	-
	Texture		Fine	Fine to course	Course	-
	Soil drainage		Very well	Well	Mod well	-
	pH	−log(H+)	6.5–7.5	5.5–6.5	7.5–8.5	-
Topography	Slope	Degree	Sigmoidal decreasing				Alos PalSAR
Climate	Annual Mean Temperature	°C	Monotonically increasing linear				Worldclim
	Annual Min. Temperature
	Annual Max. Temperature
	Rainfall	mm	Sigmoidal increasing				CHIRPS
Socioeconomic	Distance from forest	meter	Monotonically decreasing linear				LULC
	Distance from settlement
	Distance from waterbody
	Distance from cropland
	Distance from road						Odisha Space Application Center (ORSAC)

.

Five variables were chosen from the soil data, such as soil type, depth, texture drainage, and pH (Figure S2). Each soil variable was reclassified and assigned new fuzzy values to each category based on their importance. The topographic parameters, such as surface elevation, slope and aspect, play an essential role in regulating microclimatic conditions and regulate the hydrological conditions, which influence plant growth and productivity. The ALOS PALSAR DEM was used to derive the topographic slope (Figure S3). Similarly, four variables were chosen from the climate data, such as annual mean, min, max, and rainfall (Figure S4).

The socioeconomic variables include proximity factors, such as distance to settlement, road, waterbody, forest and cropland (Figure S5) [32]. The Euclidean distance tool of ArcGIS software was used to create the proximity layers. The LULC map (Figure 1) was employed to create proximity layers of the forest, cropland, waterbody and built-up areas.

3.3. Standardization and Ranking of Each Variable of Criteria

The variables under different criteria are diverse with different dimensions. Thus, the variables were classified to establish consistency for comparison and integration for land suitability evaluation [67]. All the variables were transformed to a fuzzy scale. In contrast to the classical set with rigid boundary, fuzzy set modeling allows computing a transitional or partial membership in a set measured between 0 and 1 through the membership function [42]. The membership function is denoted with a value ranging between 0 and 1, representing the non-membership and full membership, respectively, and the intermediate values measure the degree of closeness of the unity [44]. The membership function can be assessed using various mathematical models as parabolic, sigmoid, inverted sigmoid, and linear functions. Fuzzy evaluates the possibility that each pixel belongs to a fuzzy set by evaluating a series of fuzzy set membership functions. The fuzzy method ranked the soil variables based on their importance, wherein “0” has the lowest suitability and “1” has the highest suitability for agroforestry. The topography variable (slope) was reclassified based on the fuzzy function “sigmoidal decreasing” which means a lower slope leads to the highest agroforestry suitability. The climate variables (Tavg, Tmax and Tmin) were reclassified using the fuzzy function “Monotonically increasing linear”, and rainfall was classified using the “Sigmoidal increasing” function. Similarly, the socioeconomic variables were reclassified using the fuzzy function “Monotonically decreasing linear”, which means a decreasing value leads to high importance in agroforestry suitability (Table 2). The appropriate scores classes for different variables under each criterion were determined based on the relevant available research and8 discussions with domain specialists.

3.4. Weight Calculation and Assignment to the Variables

Each input layer (variable) under each criterion has a unique attribute that defines its value for agroforestry. Several objective and subjective methods are used to calculate the weights of each variable/criterion based on their relative importance [68]. In the current study, the weight of each variable was calculated based on their variable importance derived by RF regression analysis, wherein higher weights were assigned to the variables of greater importance. The agroforestry location points collected on the ground were utilized as input to estimate the weight of each variable for assessing agroforestry suitability. The reference ground observations were randomly segregated into training (70%) and validation (30%) data points (

Table 3. Criteria table for various agroforestry interventions.

LULC Classes	Criteria Table
	LULC	Interventions	Species	Examples
Agricultural land	(a) Double crop: Paddy-Pulses	Bund plantation	Fruit and wood (teak) plantation	Apple Ber Teak and Bamboo
	(b) Single Crop: Paddy-Fallow	Intercropping and Bund plantation	Bund plantation-Wood plantation	Apple Ber Teak and Bamboo
	(c) Single Crop: Cotton-Fallow	Intercropping and Bund plantation	Intercropping-Orchard plantation	Intercropping with Mango with Cotton
Built-up	Settlement areas	Home garden	Fruit and multipurpose plantation	Orchard, Amla, Lemon
Wasteland	Wasteland	Block/bulk plantation	Bamboo plantation	Bamboo

). The training data points were used in model development and weight estimation, while the validation data points were used to evaluate the accuracy of the agroforestry suitability map. Further, the training data points were randomly segregated into training (70%) and testing (30%), to train the RF model and to test the model accuracy. The weight assigned to each variable was specified so that the sum of all the weights is equal to 100%.

3.5. Agroforestry Suitability Model Generation and Validation

A multi-criteria evaluation approach was used for agroforestry suitability mapping integrating the input variables of soil, topography, climate and socioeconomic by applying the weighted linear combination (WLC) method. The standardized variables were integrated into a GIS environment using the multi-criteria approach to create an agroforestry suitability map. The final suitability index (SI) map was computed as follows:

$$\mathrm{SI}={\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}}·{\mathsf{\mu }}_{\mathrm{i}}\left(\mathrm{x}\right)$$

where n is the number of factors, w_i is the weight of factor i, and µ_i is the membership grade for factor or variable i. The value of the suitability index varies between 0 and 1, where 0 represents least suitable and 1 indicates most suitable.

The final agroforestry suitability map was validated using validation data points containing 30% of the total ground observations. The ground agroforestry sites were superimposed on the suitability map, and the values were extracted for validation.

3.6. Generating Suitability Maps for Various Agroforestry Interventions

The agroforestry suitability map was further overlaid on the LULC and slope maps to demarcate suitable zones for various agroforestry interventions. The criteria used in this case with suitable species are shown in Table 3. The agroforestry interventions are categorized into three main components: (i) bund and boundary plantation (planting trees on bunds and boundaries), (ii) intercropping (putting trees with crops) and (iii) backyard (kitchen garden/nutri-garden) or home garden (Table 3).

Bund plantations are carried out in wide bunds between the agricultural lands, indicating close association with croplands, whereas boundary plantations are carried out along the boundary of the crop fields. This intervention is also practiced in the contours in sloppy terrain and bunds in the water conservation structures, thus reducing soil erosion and runoff and helping in retaining the soil and moisture. Backyard or home gardens are mostly complex, and a mixed cropping system contains fruit trees or sub-canopy vegetables planted around the home. Such interventions are dissimilar from orchards with a significant difference in the patch area, house proximity and utilization or management practices. Intercropping indicates trees are planted inside crop fields following a defined spacing or geometry, maintaining the tree canopy and intermediate crop cover area. The bund and boundary plantation and intercropping were prescribed in agricultural areas depending on the cropping intensity. The home gardens are prescribed in and around built-up areas, wherein block plantations in fallow land with a slope of more than 15 degrees. The data processing overall methodology is shown in Figure 3.

4. Results

This study considered 15 variables from four broad criteria to determine agroforestry suitability (Table 1). The developed RF machine learning regression model indicated 83% overall accuracy compared with the testing data. The RF model derived variable importance for each criterion was converted into relative weights (

Table 4. Random forest (RF) estimated weights for different variables.

Sl. No.	Factors	Variable Importance	Relative Weight
1	Soil type	1.06	0.01
2	Depth	3.41	0.05
3	Texture	2.27	0.03
4	Drainage	0.78	0.01
5	pH	2.38	0.03
6	Slope	0.85	0.01
7	Mean temperature	2.74	0.04
8	Minimum temperature	4.36	0.06
9	Maximum temperature	3.14	0.04
10	Rainfall	0.73	0.01
11	Distance from settlement	6.04	0.08
12	Distance from road	21.29	0.28
13	Distance from water	1.57	0.02
14	Distance from cropland	23.05	0.31
15	Distance from forest	1.03	0.01

). The significance of individual parameters was denoted by the weights for agroforestry suitability and utilized in MCE. The validation data points (30% of the total ground observations) were overlaid on the agroforestry suitability map, which indicated a high average value (>0.86 with a 0.71–0.95 range). The agroforestry suitability map (Figure 4) was classified into four groups: low, moderate, high, and very high suitability, based on an equal interval value of 0.25. The agroforestry suitability was assessed based on the existing agroforestry interventions adopted by the farmers on the ground. The regression analysis indicated higher importance of the proximity variables, wherein the highest weight was estimated for distance to cropland, followed by distance to road and settlement (Figure S1). The estimated variable importance indicated the highest importance for distance from cropland (31%) (Table 4), indicating that cropland surrounding regions are mostly utilized for agroforestry as identified in the ground observations. The second most important variable was the distance to roads (28%), followed by distance to settlement (8%). The proximity to the waterbody and forest indicated lower importance as being 2% and 1%, respectively (Table 4, Figure S1). On the other hand, comparatively, lower importance was estimated for forest proximity.

The soil variables show significant importance, varying from soil depth (5%), soil type (1%), texture (3%) to pH (3%) (Table 3, Figure S2). The lower importance of soil variables shows that soil types are mostly uniform and have very lesser variability in the regions. Soil depth is found to be more significant, wherein a higher soil depth leads to higher agroforestry suitability. Drainage density shows less significance (1%), indicating that agroforestry has been practiced in both less to high drainage-density areas. The topography variable of slope shows less importance (1%) in the suitability assessment because the study area is almost uniform with lesser elevation variations (Figure S3). The climatic variables show moderate importance, whereas temperature variables show higher importance than rainfall (Figure S4). The maximum importance was estimated for the minimum temperature (6%), followed by annual mean temperature (4%) and maximum temperature (4%). On the contrary, lower importance was estimated for rainfall (1%), indicating lesser variations (mean annual rainfall varies between 1200 mm and 1685 mm) or sufficient rainfall that supports all kinds of observed interventions in this region.

4.1. Agroforestry Interventions

4.1.1. Bund and Boundary Plantation and Intercropping

Double cropping system (paddy-pulses): the double-cropping agricultural lands mainly cultivate paddy in lowland areas in the monsoon/Kharif season and pulses in the Rabi season, depending on the availability of irrigation facilities. Such low laying agricultural lands depend on water in the monsoon season for primary cropping, and the second crop is grown using the residual moisture. Such croplands are only suitable for bund and boundary plantations (Figure 5 and Figure S6a), identified in the 7311.40 ha area. The double-cropping land overlaid on the agroforestry suitability map indicates very high suitability in more than 90% (6602 ha) of the total area, followed by high suitability in 692.47 ha (9.47%). Only 0.23% area is found in the moderate and less suitable category. The suitable areas (

Table 5. Area suitability statistics under the various agroforestry interventions.

Classes	Interventions	LULC Area (ha)	Class Area (ha)				Class Area (%)
			Low	Mod	High	Very High	Low	Mod	High	Very High
Double crop: Paddy-pulses	Bund and boundary plantation	7311.4	6.56	10.31	692.47	6602.06	0.09	0.14	9.47	90.3
Single crop: Paddy-fallow	Intercropping and Bund and boundary Plantation	20,702.76	12.25	17.42	1650.21	19,022.88	0.08	0.06	7.97	91.89
Single crop: Cotton-fallow	Intercropping and Bund and boundary plantation	9970.14	10.05	7.18	927.08	9025.83	0.1	0.07	9.3	90.53
Settlement	Home garden	915.43	0.13	1.51	85.21	828.58	0.01	0.17	9.31	89.52
Wasteland	Block/bulk plantation	2823.98	1.91	4.74	289.27	2528.06	0.07	0.17	10.24	89.96

) are mostly estimated for agricultural lands along with the stream network.

Single cropping system (paddy-fallow): the single crop paddy is cultivated in lowland areas in the Kharif season and remains fallow in the other seasons due to a lack of irrigation facilities. Such areas are suitable for both bund and boundary plantation and intercropping (Figure 6 and Figure S6b). The single crop paddy is the dominant cropping practice in the study site, wherein most of the area is estimated to be suitable for this category. The total area for bund and boundary plantation with conditionally intercropping is estimated in 20,702.76 ha, wherein ~92% of the total area is found to be under the very high suitable category, followed by ~8% under the high and 0.14% under the moderate and less suitable category.

Single cropping system (cotton-fallow): the single long-duration crop of cotton is cultivated in high land areas from the Kharif to Rabi seasons. Such areas are suitable for bund, boundary and intercropping plantations. Several farmers plant mango or other fruit plants via intercropping in the cotton croplands in the study area (Figure 7 and Figure S6b). Out of the total area of 9970 ha, very high suitability is estimated in 90.53% (9025.83 ha) area, followed by high suitability in 9.3% (927.08 ha) area, whereas only 0.17% area shows less suitability.

4.1.2. Home Gardens

Home garden is a popular agroforestry model compared to other practices and is a common practice in Odisha. Home garden is practiced as a mixture of crops (mostly vegetables, herbs, medicinal plants), trees (fruit and or fodders trees), and provides diversified products. Nearly 99.82% of the total settlement area has been identified as suitable for home gardens, while only 0.18% has been identified as less and moderately suitable (Figure 8 and Figure S6c). A few settlement patches in the western, southern, and northern parts indicate less suitability, which could be due to the high slope, low soil depth, coarse texture, sandy and highly alkaline soil type.

4.1.3. Block/Bulk Plantation

The block/bulk plantations in this region are mostly seen in permanent fallow or wastelands (Figure 9 and Figure S6d). Out of the total area of 2823 ha, very high suitability is estimated in ~89.96% (2528.06 ha) area, followed by high suitability in 10.24% (289.27 ha) area, whereas only 0.07% area shows less suitability.

5. Discussion

The site-specific suitability intervention plans for agroforestry implementation were successfully assessed using machine learning techniques and the MCE approach. Compared to previous studies that relied on perspective-based analysis, the present study used extensive agroforestry field data for suitability assessment for on-ground interventions. The RF machine learning algorithm well estimated the variable weights based on the successful agroforestry models implemented in the study site. This approach better estimated the relative weights of the factors compared to the method where the weights are derived based on perception. The estimated weights indicated higher importance of socioeconomic (proximity) variables, wherein the maximum value was observed for distance to cropland (31%), followed by distance to the road (28%) and distance to settlement (8%). The higher importance of the proximity variables, wherein the highest weight was estimated for distance to cropland, followed by distance to road and settlements, could be attributed to the preference of the local people in selecting suitable areas for agroforestry interventions. The proximity variables combinedly contribute to 70% of the total weight. In comparison, the lower weights were estimated for climate and soil parameters. Again, the topography indicated less importance in agroforestry suitability assessment. The high importance of distance to roads could indicate accessibility and connectivity. The results showed that most agroforestry intervention is carried out along the road or in proximity area to better transport agroforestry products. The importance of distance to settlement could be attributed to the farmer’s preference for agroforestry interventions in and around home premises for easy access, regular monitoring, and better management. Moreover, the home gardens could lead to the high importance of distance to settlement. The proximity to the waterbody and forest indicated lower importance. This shows that agroforestry plantations are carried out both near rivers, canals or dams and distant to water sources. Ground visits show that several water irrigation structures or systems have been installed to ensure the survival of saplings during dry summer. Similarly, a lower weight was also estimated for the distance to forest. It could be noted that there are no contiguous forest patches inside the study area; instead, multiple fragmented forest patches are seen, which leads to less significance of distance to the forest.

The results show that most of the agroforestry patches were found in and around the agricultural lands with good transport connectivity. In comparison, the rainfall and temperature are found to be a homogeneous and acceptable range for agroforestry in a sub-humid tropical climate zone. Similarly, the soil and topography are found to be uniform in a smaller study site. Doddabasawa et al. [69] assessed the agroforestry interventions in the six districts of Karnataka, India, and reported six agroforestry systems, wherein bund plantation was reported as the dominant system (43.06%), followed by boundary plantation (19.44%) practiced in this region. They reported the farmer’s preference for bund and boundary plantation over other interventions. Ahmad et al. [40] carried out a national assessment on agroforestry suitability in India. Their study identified 46% suitability out of the total 52% of the total agricultural land in India, wherein the agricultural lands in tropical and sub-tropical climate zones are estimated to be highly suitable, including the current study site. Nath et al. [17] assessed agroforestry suitability in the eastern Indian Himalayan region, integrating different climate, soil, topography, ecology, and socioeconomic variables, wherein they assigned higher weights to climate, topography and soil, and lower weights to socioeconomic and ecological variables. This could indicate the choice of appropriate variables based on landscape characteristics. The distance to the waterbody indicated less influence (2%). Agroforestry patches are found both along the river and distant locations, wherein various artificial water harvesting structures are installed for the initial survival of agroforestry plantations in regions with lesser water availability. The water-filled pipe structure was installed around plants/saplings that supplement plant’s water requirement for 1–2 weeks for their sustainability in the summer season.

The currently adopted approach can be tested in climate zones or landscapes to assess the transferability or applicability with modifications in the list of the indicators. The current method heavily relies on the existing agroforestry interventions or local practices, which may be unavailable for many regions or costly to collect. However, the current study using the existing agroforestry data better assures the suitability compared to approaches that rely on ad hoc or perspective-based weights. The developed approach can be utilized for agroforestry suitability analysis and prescribing appropriate interventions in other areas using the local agroforestry data. The approach can be hosted on cloud-based platforms, such as Google Earth Engine (GEE). The entire analysis framework can be made available, including the independent variables and data processing tools. This proposed system will take minimal input from the users, as the ground observations on existing agroforestry data and will automatically execute the process to generate the agroforestry suitability maps and prescribe suitable interventions. The generated output will assist the planners in implementing suitable agroforestry. The generated data working as a baseline layer could be useful in the stakeholders’ consultation with the local people or beneficiaries as the next step to implementation. The approach can be executed in the various agro-ecological zones of India to generate national agroforestry suitability maps blended with the localized ground data and covariates. The spatially explicit data and information will help to develop suitable agroforestry policies, agriculture and water resource management, etc., that are in the context of location-specific conditions.

6. Conclusions

The site-specific agroforestry suitability was assessed using machine learning, fuzzy, and GIS-MCE techniques. The RF machine learning algorithm was applied for weight estimation. At the same time, fuzzy theory offers an outstanding framework for translating the numerical data of different thematic or membership sub-categories based on their importance to agroforestry suitability. The socioeconomic (proximity) parameters are found to be the most suitable indicators in identifying the suitable areas for agroforestry interventions in a sub-humid tropical climate zone. In comparison, the environmental conditions defined by climate, soil and topography are found to be uniform and have less influence. The fusion of fuzzy and MCE ensures the accurate identification of suitable areas for agroforestry. Moreover, the criteria-based integration with the LULC map well identified the potential areas for appropriate agroforestry interventions. Most of the area in the study site identified as suitable for various agroforestry could be attributed to the dominance of agricultural lands, where bund and boundary plantation and intercropping can be implemented. The adopted approach can be tested in other regions by modifying the list of indicators. The landscape-level agroforestry suitability mapping would allow agriculture managers and policymakers to plan location-specific agroforestry implementations. The implementation of cost-effective and sustainable agroforestry interventions will have multiple benefits, such as reducing soil erosion, improving water retention, addressing climate change, ecosystem restoration, food and nutrition security, and economic improvement.