The study was carried out in arid and semi-arid landscapes of Northern Karnataka, located between 13° 34' and 18° 28' N and 74° 59' and 77° 41' E across districts Vijayapura, Chitradurga, Bagalkot, Koppal, Bellary, Raichur, Kalaburagi, Yadgir, and Bidar covering an area of 71149.04 km² Figure 1. The study area is a part of the Krishna Basin, situated on the Deccan Plateau at an elevation between 300 and 730 meters. The landscape is predominantly black and red soils, categorized as shallow, medium-deep, and deep, which supports the cultivation of key crops like green gram, pearl millet, sunflower, pigeon pea, sorghum, chickpea, and rabi sorghum. LU in the region is dominated by agriculture, fallow areas, wastelands, and degraded forests, with most of the terrain exhibiting slopes of less than 5%.

North Karnataka's hydrological network consists of the Krishna River Basin (Krishna River and tributaries, Bhima, Ghataprabha, Malaprabha, Vedavathi, and Tungabhadra), and the Godavari River Basin (Manjira and Karanja). These rivers serve as vital water resources for agriculture and support riparian ecosystems throughout the region.

This ecoregion extends northward into eastern Maharashtra, highlighting the ecological interconnectedness of the area. The region receives most rainfall during the monsoon season from June to September, ranging from 370 to 4200 mm annually. The region is also characterized by high temperatures, with summers often exceeding 40°C. Rising temperatures during March-June, especially in recent decades, have exposed North Karnataka to increasingly frequent heatwaves, posing a significant challenge to human and animal health. This region is prone to severe floods in the Krishna River basin.

The region possesses a rich historical legacy, evidenced by powerful dynasties (Kadamba, Rashtrakuta, Chalukya) and flourishing literary figures (Pampa, Ponna, Ranna). Extreme climatic events, high rates of anemia (50% in women, 65.5% in children), and malnutrition, particularly in districts like Kalburgi, Raichur, Yadgir, Koppala, Ballari, Bidar, and Gadag, further exacerbate the challenges. It is divided into two distinct sub-regions, Hyderabad-Karnataka (Bidar, Kalaburagi, Raichur, Yadgir, Bellary, and Koppal) and Mumbai-Karnataka (Vijayapura, Bagalkote), with lower socio-economic development.

Spatial analyses were carried out using remote sensing (RS) data and collateral data. Temporal remote sensing data of Landsat MSS, TM, OLI-1, and OLI-2 were acquired from the spatial data portal of the United States Geological Survey (USGS - https://earthexplorer.usgs.gov/) for the 1970s, 1980s, 1990s, 2000s, 2010s, and 2020s, detailed in Supplementary Table 1. The Landsat program has been operational since 1972 and offers the most extensive medium spatial resolution satellite data collection. It has been extensively used in the assessment of LULC. The datasets were carefully chosen to ensure minimal cloud coverage (<10%). The data were pre-processed to rectify geometrical and radiometric discrepancies in the Google Earth Engine (GEE) Platform (https://earthengine.google.com/). Region-specific taluk and district administrative boundary maps were obtained from the K-GIS portal (https://kgis.gok.in).

Training data for LU classification were gathered from various locations within the study area using a handheld pre-calibrated global positioning system (GPS), online spatial portals (Google Earth -https://earth.google.com), and Bhuvan (https://bhuvan.nrsc.gov.in) with high-resolution remote sensing data. All these datasets corresponding to the study area were reprojected to a common geodetic datum, the World Geodetic System 1984 (WGS84), and Universal Transverse Mercator (UTM) within 43N zones, ensuring consistency in mapping. Road networks were extracted from Survey of India topographic maps at scales of 1:50,000 and 1:250,000 (https://www.surveyofindia.gov.in). The study considered Virtual online spatial maps such as Bhuvan (http://bhuvan.nrsc.gov.in) and high-resolution Google Earth (http://earth.google.com) to validate classified thematic maps.

Ecological, biological, geo-climatic, and resource data were compiled through field investigations, review of published literature, and reports. Elevation and slope maps were derived from the Shuttle Radar Topography Mission (SRTM) data with a 30-meter resolution (https://earthdata.nasa.gov).

Figure 2 outlines the protocol for delineating NRRRs at disaggregated levels across the arid regions of North Karnataka. This entails (i) division of the study region into grids of 5’× 5’ (or 9 km × 9 km), (ii) land cover and land use analyses, (iii) assessment of the condition of forests through fragmentation metrics, (iv) prediction of likely LUs, (iv) delineation of NRRRs at disaggregated levels (grids).

The Normalized Difference Vegetation Index (NDVI) characterizes vegetation cover by assessing the difference in reflectance in the visible and near-infrared spectrum (land cover). It is widely employed for monitoring vegetation dynamics on various scales (Tucker, 1979)(Ashok et al., 2021). NDVI is exceptionally responsive to red reflectance, strongly influenced by density and green cover, whereas NIR reflectance is impacted by density alone, not green cover (Bremer et al., 2011). Utilizing the red and NIR bands of Landsat data, NDVI values are computed, ranging from -1 to 1; values below zero signify dormant seasons (e.g., bare land, open land, cloud cover, snow, water bodies), while values above zero indicate vegetation cover during the growing season. Equation 1, detailed below, is used for the computation of NDVI.

(1)

NIR and RED denote the electromagnetic spectrum corresponding to near-infrared and red wavelengths. The vegetation and non-vegetation have been categorized based on a threshold value Supplementary Table 2. LU analysis involves generating FCC (false color composites) from remotely acquired data bands (NIR, Red, and Green) to identify heterogeneous landscape patches. The current study collected training polygons from the field using pre-calibrated handheld global positioning systems (GPS) and high spatial resolution data from Google Earth. Chosen training polygons represent all LU classes, covering 15% of the study, and are uniformly distributed throughout the study area. The attribute information for these training polygons was collected from the field using precalibrated handheld GPS devices and high spatial resolution data from Google Earth.

70% of training polygons are used for supervised classification, while the remaining (30%) are used for testing (Nguyen et al., 2021). Spatial data (RS) were classified using a supervised machine learning algorithm, RF Supplementary Figure 1a. RF is a novel technique employing a set of classifiers or a collection of multiple decision tree predictors. Each tree is constructed based on the randomly sampled feature vectors with replacement. It is independently generated with a uniform distribution shared across all decision trees to acquire high training data accuracy and enhance generalization accuracy as their complexity increases Supplementary Figure 1b. These multiple classifiers are typically aggregated through a plurality voting scheme known as bagging (Breiman, 1996).

RF can effectively handle multi-dimensional data while employing a substantial number of trees within the ensemble ((Ramachandra et al., 2022);(Ramachandra et al., 2023)). RF requires a significant amount of memory due to the storage of an N by ntree matrix in memory, and it is not computationally intensive; the trees are constructed without pruning (Gislason et al., 2006)(Rodriguez-Galiano et al., 2012). The computational time for RF is computed as per Equation 2.

(2)

Where c is the constant, ntr represents the number of trees, M is the number of features, and N is the number of samples. A majority vote among the trees is employed in the prediction of the class of observation in the RF model. As ntree and M are two main hyperparameters in random forest, optimization of these parameters' aids in an increase in model accuracy. Increasing these parameters generally improves model performance but also increases computation time. The current study considered the ntr based on the iterative method, ranging from 50 to 500 with an interval of 50, and prioritized 300 trees for best performance. M was considered as its default value (√M). Classified LU is validated using training data (30%) through computation of overall accuracy, producer accuracy, user accuracy, and kappa statistics.

Markov chain (MC) analysis represents a heuristic modeling approach, which has been extensively used to examine LU change dynamics across various spatial scales (Halmy et al., 2015). A Markov chain operates based on the principles of the probability of a system assuming a particular state at a given time can be ascertained based on its known prior state (Rimal et al., 2018). Markov chain analysis involves the development of a transition probability matrix that accounts for LU change between two distinct periods (Fu et al., 2018). The Markov chain model does not account for changes in the spatial distribution. The cellular automata model, which is a spatially explicit model, can overcome these shortcomings by representing spatial attributes in mapping LU change compared to non-spatial models (Guan et al., 2011). An integration of CA and Markov Chain (CA-Markov) model aids in predicting likely LU changes (Rimal et al., 2017), based on the transition probability. A Markov chain determines the distribution of the LU class to another from time t to t+1 (Setturu & Ramachandra, 2021). The CA model detects changes in the spatial distribution at the cell level and captures interactions with neighboring cells. The Markov chain model enables the prediction of future spatiotemporal modifications (Equation 3) (Tariq et al., 2023)(Wang et al., 2022).

(3)

Where S represents the LU status at time t, S(t+1) denotes the LU status at time t+1, P ij stands for the transition probability matrix within a specific state.

This matrix (Equation 4) is computed as described in previous studies (Sahu et al., 2021).

Where P represents the transition probability, where signifies the probability of transitioning from state i to another state j in the subsequent period. denotes the state probability at any given time. In this context, states with a low transition rate tend to have a probability close to 0, whereas high-growth states tend to have probabilities approaching 1 (Mumtaz et al., 2020). Model validation: The accuracy of the model was validated by comparing current LU map of 2022 (reference map) with the simulated map of 2022. This validation process utilized an integrated VALIDATE module within IDRISI Selva 17.02 software (https://idrisi-selva.software.informer.com/) to assess the level of agreement between the classified and the simulated maps. The agreement metrics are based on the widely recognized Kappa Index of Agreement (KIA), which includes various metrics such as Kappa for location Strata ( K locationStrata ), Kappa for location (K location ), and Kappa for no information (K no ). K no was employed to assess the overall agreement between the proportions of the reference and modeled maps, which evaluated the precision of the spatial attributes, quantity, and locations of grid cells within specific LULC class categories (Ozturk, 2015).

An analysis of forest fragmentation quantifies the condition of forests, which determines the extent of structural and compositional changes in the forest ecosystem. The condition of forests in the study region is assessed through the computation of fragmentation indices, P f , representing the proportion of forest pixels to non-water pixels ( P f ) and P ff represents the proportion of cardinal pixel pairs (both forest pixels) to pairs with at least one forest pixel (Riitters et al., 2000)(Riitters et al., 2004)(Ramachandra et al., 2016).

This aided in assessing the condition of forests through pixel categorization based on the type of fragmentation (details are provided in Supplementary Table 3), as interior forest (Pf = Pff =1), transition (pertaining to pixels with Pf <0.6 and Pf >0.4), patch forest (for pixels with Pf < 0.4), perforated forest (applicable to pixels with Pf > 0.6 and (Pf – Pff ) < 0), edge forest (relevant for pixels with Pf > 0.6 and (Pf – Pff ) > 0), non-forest pixels encompass all pixels not classified as forest cover. This classification scheme serves as a structured framework for analyzing different types of forest fragmentation, providing a nuanced understanding of the diverse spatial patterns within the study area.

Analyzed hydrological, biological, geo-climatic, and socio-economic details at disaggregated levels in the arid region of Northern Karnataka for identifying Natural Resource Rich Regions (NRRRs). The region was divided into grids of 5′ × 5′ equivalent to approximately (9 × 9) km² (Ramachandra et al., 2018), comparable to grids in the 1:50000 scale topographic maps (the Survey of India, Government of India). The spatial extent and occurrence of features for each variable have been assessed at the grid level, and the variable is assigned a weight based on the relative worth. This approach aided in combining multiple datasets and their significance in the landscape details in Supplementary Table 1 Weights for variables were assigned as per Supplementary Figure 2 and aggregated for each grid, as per Equation 5. The study area was grouped into four zones considering aggregated weights, which also highlights the ecosystem condition based on the availability and vulnerability of natural resources:

where n is the number of factors or variables, W i is the weight associated with criterion i, V i is the associated value with that criterion.

Based on the aggregate weightage matrix, the study region is classified into four zones (NRRR 1 to 4). NRRR 1 represents natural resources rich region, requiring strict conservation and protection measures, NRRR 2 is less sensitive than NRRR 1, except for the degradation of some natural resource patches. NRRR 3 represents a moderate resource region, and NRRR 4 represents lower sensitivity with erosions in the ecosystem conditions.

The long-term analyses of LC changes using NDVI of the northern arid regions of Karnataka have been done to delineate the spatial extent of vegetation. The area under vegetation has shown an increasing trend, as depicted in Figure 3, increasing from 32.79% (in 1973) to 53.35% (in 2022), which suggests the intensification of agricultural and horticultural practices with increased water availability due to the construction of multiple reservoirs. The area under non-vegetation has shown a consistent decrease over the decades, from 67.21% (in 1973) to 46.65% (in 2022), as open spaces, including fallow land, were converted into croplands, horticultural lands, agroforestry, and forest plantations. The area under non-vegetation would decrease as detailed in Supplementary Table 4.

Temporal LUs (1973 to 2022) of the Northern arid regions of Karnataka illustrate that a predominant agrarian landscape has undergone intense anthropogenic changes since 1996 due to globalization and liberalization. The overall accuracy of the remote sensing analysis is 96% and the kappa coefficient is 0.91 Supplementary Table 5. Highly elevated hills and plateaus are covered with dry deciduous forest and scrub forest in the lower regions of Chitradurga, Bellary, Koppal, and Bagalkote districts. Dry deciduous forest extent has shown a continuous trend of decrease (depicted in Figure 4), from 2,154.20 sq. km (1973) to 1,096.34 sq. km (2022), and details are provided in Supplementary Table 6. Scrub land, prevalent in semi-arid ecosystems, has shown a similar reduction, declining from 5,650.74 sq. km (7.94%) in 1973 to 2,260.19 sq. km (3.18%) in 2022. Bellary district has a reasonable spatial extent of dry deciduous forests in the Sandur forest range of Sandur taluk, and rampant iron ore mining in the Sandur taluk has impaired the integrity of forest ecosystems. These declines are attributed to land conversion for agriculture (witnessed in the Yadgir Reserved Forest area) and urban development. Some areas of degraded deciduous forest have been converted into scrubland.

The study region is principally agrarian, and Bellary, Raichur, and Koppal districts are popularly known as the “rice bowl of Karnataka”. The extent of agricultural land increased from 78 % (in 1973) to 83.11% in 2022. The study area is enriched with the Krishna, Tungabhadra, Bhima, and Godavari rivers. The study region has witnessed a shift toward wet cultivation with enhanced water security due to multiple irrigation projects with the implementation of reservoirs like Almatti Reservoir, Basava Sagar Reservoir (Narayanpura), Karanja Reservoir, Jurala Reservoir, Vani Vilas Sagar in this region. The spatial extent of water bodies has increased markedly from 494.93 (1973) to 1,397.68 (2022) sq. km. The increased water security has expanded horticultural land from 1,089.35 sq. km (1.53%) in 1973 to 4,198.58 sq. km (5.90%) in 2022. The expansion of paved surfaces (built-up) from 186.22 sq. km (in 1973) to 1085.12 sq. km (in 2022) in the district reflects urbanization and infrastructure development. Cities are expanding due to urbanization in the core area and sprawl in the peri-urban area, with good connectivity of road networks. The growth of various industrial layouts in the core area aided as a catalyst for the expansion of the urban centers. Paved surfaces (built-up) in rural areas have been increasing at a constant rate.

Mining of iron ore is rampant in the Sandur-Hospet region of Bellary district, which is rich in iron ore reserves, and has increased post-2005, covering around 27.31 sq. km (in 2022). Plantation of exotic species like Acacia auriculiformis, Acacia catechu, Tectona grandis, Eucalyptus globulus, Casuarina equisetifolia L., and others have been increasing, reaching 131.08 sq. km (in 2022).

Forest ecosystems in the arid region of North Karnataka are undergoing fragmentation due to anthropogenic activity, with LU changes leading to land degradation. Fragmentation analysis emphasizes the loss of intact forest cover. The study area comprises degraded forest patches, and the results of fragmentation metrics also reveal that the interior/intact forest has declined from 3,252.39 (in 1973) to 1,508.29 sq. km (in 2022), and details are provided in Supplementary Table 7. Figure 5 highlights that the spatial extent of non-forests has increased from 63,819 (1973) to 66,543 sq. km (2022).

Predictions of likely LU indicate the impact of the current rate of LU transitions in the next two decades with the help of CA-Markov techniques. Modeling is validated by comparing the simulated LU with the actual LU of 2022 and the computation of Kappa statistics. Kappa of 0.9 to 0.95 suggests agreement between the predicted and actual LU with high efficiency. The simulated LU showed a very minimal overestimation of water bodies in 2022.

The predicted LUs depicted in Figure 6 and details given in Supplementary Table 8 show a likely increase in built-up to the extent of 6.23% (in 2030) and 7.17% (in 2038). The likely built-up increase will be due to the rise in food processing, the food and beverage sector, and the expansion of roads or highways. The decline of scrubland to 11.63% and dry deciduous to 1.04% (in 2038) in a business-as-usual scenario highlights the likely continuation of forests and scrublands.

Prioritization of NRRRs in the northern arid regions of Karnataka at disaggregated levels (grids and villages) was done through integrated assessment considering bio-geo-climatic, land, ecology, energy, environmental, and social variables compiled through field investigations and supplemented with the review of published literature. Weights were assigned to these variables at grid levels based on the relative significance at disaggregated levels.

Dry deciduous forests and scrubland in the hills of Chitradurga, Bellary, Bagalkolte, and Raichur account to 15% to 60% Supplementary Figure 3a. Forest degradation in Vijaypura, Kalaburgi, and Yadgir is due to anthropogenic pressures with agricultural expansion. The intact/interior forests Supplementary Figure 3b are confined to higher-elevation regions and protected areas.

The ecological variables like endemic flora, fauna, forest biomass, species abundance, species diversity (Shannon diversity), and the presence of conservation reserves in the Northern Arid Karnataka were assessed, and Supplementary Figure 4a and Supplementary Figure 4b give the spatial distribution of flora and fauna, respectively. Shannon’s diversity ranges from 1 to 2.5 Supplementary Figure 4c, and Supplementary Figure 4d depicts species that range from 50 to 200. The region has dense, dry deciduous forests with a carbon sequestration potential up to 300 Gg Supplementary Figure 4e. The protected areas in the study area are Malaksamudra Bird Sanctuary (Koppal), Yadahalli Chinkara Wild Life Sanctuary (Bagalkote), Bonal Bird Sanctuary (Yadgir), Yadgir reserved forest (Yadgir), Gudekote and Daroji Sloth Bear Sanctuary (Bellary), Ankasamudra Bird Sanctuary (Bellary), and Chincholi Wildlife Sanctuary (Kalaburagi) depicted in Supplementary Figure 4f.

The study area has the highest elevation of 750 m in Chitradurga and Bellary; Elevation in Koppal, Bagalkote, Vijaypura, and Bidar ranges from 500 to 750m, and elevation in Raichur, Yadgir, and Kalaburagi is at 250 to 500 m Supplementary Figure 5a. The slope is less than 15% in the study regionSupplementary Figure 5b. The northern part of the study area (Bidar, Kalaburagi, Vijaypura, Yadgir, Bagalkote, Raichur, Koppal and partly Bellary) receives 1200 to 600 mm of rainfall, whereas part of Bellary and Chitradurga receives <600 mm of rainfall Supplementary Figure 5c. Supplementary Figure 5d shows that Bidar, Kalaburagi, Vijaypura, Yadgir, Bagalkote, and Koppal have coarse loamy soil; Bidar, Kalaburagi, Vijaypura, Yadgir, and Bagalkote have sandy or sandy skeletal soil; Bagalkote, Raichur, Bellary, and Chitradurga have rocky outcrops or Fragmental soil; Kalaburagi, Vijaypura, Yadgir, and Bagalkote have clayey loamy or clayey skeletal soil; Koppal, Raichur, and Chitradurga have loamy or clayey soil. The middle part of Bagalkote is composed of Charnokities; Chitradurga, Bellary, Koppal, Raichur, and Yadgir are primarily composed of Peninsular Gneiss; the hills of the study area are composed of Dharwars or Granite; Bagalkote, Vijaypura, Kalaburagi, and Bidar are part of the Deccan trap Supplementary Figure 5e. Bidar, Kalaburag, and Yadgir are in the arid zone; Vijaypura, Yadgir, Bagalkote, Raichur, and Chitradurga are majorly in the hot-dry semi-arid zone; and part of Vijaypura, Raichur, Koppal, Bellary, and north Chitradurga are in the hot-dry arid zone Supplementary Figure 5f.

Krishna, Tungabhadra, Bhima, and Godavari Rivers flow, and the duration of water flow in streams Supplementary Figure 6a varies from 3 to 6 months in this region. The drainage density is higher (>2.5) in Raichur and Bellary Supplementary Figure 6b. The major reservoirs of this district are Almatti, Basava Sagar (Narayanpura), Karanja, Jurala, and Vani Vilas Sagar Supplementary Figure 6c.

Northern Arid Karnataka has the potential of more than 6 kWh/sq. m of solar energy Supplementary Figure 7a. Multiple solar parks have been established in Chitradurga, Bellary, Bagalkote, and other districts. Kalaburagi, Raichur, and Yadgir have a high potential for wind energy with wind speeds of more than 3.5 to 4 m/sec throughout the year, and windmills are present in the hills of the districts Supplementary Figure 7b. Also, there is scope for bioenergy Supplementary Figure 7c of 200-400 MKcal in Bidar, Kalaburagi, Vijaypura, Yadgir, Bagalkote, Raichur, Koppal, and Bellary; 200-600 MKcal in Chitradurga and Kalaburagi.

The population density is presented grid-wise in Supplementary Figure 8a, and livestock density is in Supplementary Figure 8b. The forest dwellers' settlements are mapped in Supplementary Figure 8c. in all districts of Northern arid Karnataka except Vijayapura.

The aggregated weightage metric score is computed for each grid, considering bio-geo-climatic, ecological, hydrological, energy, and social factors. Grids are grouped into four levels and presented in Figure 7 depending on the frequency of occurrences of aggregated scores. The NRRR1 (67 grids) and NRRR 2 (127 grids) are considered highly rich regions of natural resources, NRRR3 (304 grids) is moderate, and NRRR4 (522 grids) is less sensitive. Figure 7 shows, grid-wise and at village levels, NRRRs in the northern arid regions (districts) of Karnataka state, India. Policy recommendations are:

NRRR1 zones include the protected forest areas and interior forests, where the integrity of forests must be maintained without large-scale development projects such as mining. This region is highly fragile, and prudent management of natural resources through monitoring by regulatory authorities is required by including Village Forest Committees (VFCs) and Biodiversity Management Committee (BMC) at village Panchayath. NRRR1 regions are to be protected without any alterations in topography due to the linear projects (new / expansion) such as roads, and railway lines. Degraded forest patches to be revitalized with native species and regulation of monoculture plantations. Existing exotics and non-endemic plantations are to be replaced with native species. Needs to promote locally available renewable energy sources such as bioresources, solar, and wind. NRRR2 characterizes a zone of higher conservation as being a transition zone between NRRR1 and NRRR3, moderate conservation regions.

A regulated sustainable development path may be allowed in NRRR3 with stringent environmental norms and location-specific environmental management plans (EMP) to mitigate the impacts. Small-scale industries, like agro-based industries, are permitted to stimulate the rural economy. Incentives should be provided to youth and women self-help groups to encourage rural entrepreneurship and establishing agro-processing industries based on local resources.

NRRR4 represents the least diverse areas, where moderate developmental activities are allowed as per the requirement with the stringent regulatory norms.