Soil Fertility Assessment Through the Integration of Satellite Imagery and Spatial Analysis: Application to Arabica Coffee Cultivation in Lonya Grande, Peruvian Amazon

Abstract

Soil fertility assessment is fundamental for improving agricultural productivity and promoting sustainable land management. This study proposes an integrated methodological framework that combines Sentinel-2 satellite imagery, spatial analysis techniques, and field-based soil data to evaluate soil fertility in Arabica coffee plantations in the Lonya Grande district, Peruvian Amazon. The framework involves three analytical phases: (i) spatial interpolation of soil macronutrients using Inverse Distance Weighting (IDW), (ii) local modeling through Geographically Weighted Regression (GWR), and (iii) spectral correlation analysis between field-measured soil properties and Sentinel-2 reflectance bands. The SWIR2 (Band 12) data were identified as the most sensitive predictor of soil moisture-related properties, with the strongest relationship observed for soil saturation (R² = 0.40). Field validation revealed pronounced spatial heterogeneity, particularly for macronutrients such as nitrogen, phosphorus, and potassium. The study also found that soils exhibited moderately acidic pH values (5.1–6.8), favorable for coffee cultivation. Despite adequate water retention, nutrient deficiencies highlight the need for site-specific soil management strategies. Overall, spatial analysis confirmed consistent relationships between remote sensing data and soil parameters, demonstrating the feasibility and cost-effectiveness of this approach under data-limited tropical conditions. The proposed framework offers a scalable basis for regional soil fertility monitoring, and future research should incorporate machine learning and expanded sampling networks to further enhance predictive performance.

1. Introduction

Soil fertility is a critical factor for agricultural productivity, directly influencing crop yield, quality, and the economic stability of farming communities [1,2,3,4]. According to the Food and Agriculture Organization (FAO) [5], approximately 33% of the world’s arable land is degraded, severely affecting soil fertility and productive capacity. This issue is particularly acute in developing regions, where agriculture underpins local economies. The growing global demand for food, driven by population growth and changing consumption patterns, further intensifies pressure on soil resources [6,7,8,9,10].

In the context of Latin America and the Caribbean, soil degradation represents an urgent and complex challenge. The Economic Commission for Latin America and the Caribbean (ECLAC) [6] reports that nearly 48% of soils in the Caribbean, 50% in Mesoamerica, and 18% in South America are classified as severely degraded. Such declines limit agricultural productivity and threaten the sustainability of food systems, posing a significant risk to regional food security [6,11,12].

At the local scale, in the Amazonas region of Peru, soil fertility is fundamental to agricultural development, serving as a key economic driver and accounting for about 25% of the economically active population [13,14,15]. In the province of Utcubamba—particularly in the district of Lonya Grande, a major coffee-producing area—soil degradation poses a serious challenge, mainly resulting from monoculture practices. The persistent reliance on monoculture leads to nutrient depletion and accelerates erosion, thereby jeopardizing the long-term sustainability of coffee farming in the region [16,17]. According to the World Reference Base for Soil Resources [18], the dominant soils in the study area are classified as Dystric Cambisols and Haplic Acrisols, characterized by moderate weathering, acidic pH, and limited natural fertility—conditions consistent with tropical highland zones in the Peruvian Amazon.

To address the limitations of conventional field-based soil fertility assessments, remote sensing technologies have emerged as a powerful alternative. Among the available satellite platforms, Sentinel-2 was selected due to its high spatial resolution (10–20 m), frequent revisit cycle (5 days), and the inclusion of the SWIR2 band, which has demonstrated particular sensitivity to soil moisture and organic matter variability. These attributes make Sentinel-2 exceptionally well-suited for fine-scale and data-driven assessments of soil fertility in heterogeneous agricultural landscapes, such as those found in the Peruvian Amazon.

Traditional methods for assessing soil fertility, such as extensive soil sampling and laboratory analysis, have several limitations, including high cost, labor intensity, and limited capacity to capture spatial variability over large areas. These constraints restrict the ability of farmers and agricultural planners to implement evidence-based soil management strategies aimed at mitigating degradation [18,19,20,21,22,23,24]. Consequently, there is an urgent need for more efficient, scalable, and spatially explicit approaches to evaluate soil fertility, providing critical information for sustainable agricultural planning and decision-making.

Remote sensing technologies have emerged as a promising solution to overcome the limitations of conventional soil fertility assessments. Satellite imagery—particularly from the Sentinel-2 mission—enables the collection of high-resolution spatial data, offering a comprehensive and up-to-date representation of soil conditions without the need for direct field sampling [25,26,27,28,29,30,31,32,33,34,35,36]. This platform employs a multispectral sensor capable of capturing 13 spectral bands, including visible, near-infrared (NIR), and shortwave infrared (SWIR) wavelengths. These spectral capabilities allow the estimation of key soil properties such as moisture content, nutrient concentration, and organic matter, with the SWIR2 band showing strong correlations with principal soil fertility indicators [28,34,37,38,39,40,41,42,43,44].

Integrating satellite data with advanced spatial analysis techniques further enhances the accuracy and applicability of soil fertility assessments. In this study, two spatial analysis methods—Inverse Distance Weighting (IDW) and Geographically Weighted Regression (GWR) [45,46,47,48]—were employed to analyze soil properties in Lonya Grande. IDW serves as an interpolation technique that generates spatial distribution maps of soil nutrients and physicochemical parameters, providing insight into spatial variability, whereas GWR allows for the exploration of localized relationships between soil properties and spectral data, capturing the inherent heterogeneity of the landscape [45,46,47,48,49].

Recent research has demonstrated the potential of remote sensing data to estimate key soil characteristics such as organic carbon, pH, and nutrient content [50,51,52,53,54,55,56]. Advances in multispectral sensors, particularly Sentinel-2 and Landsat 8, have enabled large-scale soil property mapping through reflectance in the visible, near-infrared, and shortwave infrared regions. Studies have successfully used these data to model soil fertility in crops like soybean and sugarcane [3,4,42,50,51,52,53,54], and to assess geothermal patterns [25,55,56,57,58,59,60]. However, existing research still faces limitations such as the heterogeneity of soil types, reduced model transferability due to regional calibration needs, and the weak integration between field observations and remote predictions. These gaps highlight the need for integrated approaches that combine spectral information with spatial interpolation techniques to improve reliability and scalability in tropical, data-scarce environments.

The present study aims to assess soil fertility in Lonya Grande, Amazonas, through a hybrid approach that integrates Sentinel-2 satellite imagery with spatial modeling techniques. Specifically, it evaluates the correlation between spectral indices (with emphasis on the SWIR2 band) and soil properties, and applies spatial interpolation using the Inverse Distance Weighting (IDW) method in areas or parameters where spectral relationships are weak or non-significant. This integrative approach enables a more comprehensive, consistent, and scalable assessment of soil fertility variability.

This study introduces an innovative, three-phase methodological framework that addresses the specific challenges of soil fertility assessment in rural Arabica coffee-growing regions. The framework covers (1) identification of optimal study areas, (2) soil sampling and laboratory analysis, and (3) model development and prediction of the Soil Fertility Index (SFI). Unlike previous studies that focused on single-crop systems or lacked spatial detail, this research integrates the SWIR2 band of Sentinel-2 imagery with localized GWR modeling, achieving high-resolution and field-scale mapping of soil fertility. Moreover, this approach provides a scalable and cost-effective solution for soil assessment, bridging the gap between technological innovation and practical field applications in sustainable agriculture.

The primary novelty of this work lies in the development of a Soil Fertility Index (SFI) based on spatial analysis modeling derived from satellite imagery, validated through field surveys of soil parameters. The proposed method was applied to a 365 km² area of Arabica coffee cultivation in the Lonya Grande district, Amazonas region, Peru. The analysis considered both physical and chemical soil properties, including key nutrients essential for Arabica coffee growth. The results demonstrated that macronutrient levels are strongly related to the SWIR2 spectral data, and that the fertility index and spatial models effectively identify nutrient concentration patterns, supporting targeted soil management practices to improve crop performance and sustainability.

2. Methodology

The proposed methodology focuses on predicting soil fertility zones for Arabica coffee cultivation through three main phases, integrating remote sensing data, spatial modeling, and field-based soil analyses to compute a Soil Fertility Index (SFI). This framework considers key environmental, physical, and chemical parameters, as well as macro- and micronutrient concentrations. The first phase delimits optimal regions within the study area based on environmental factors. The second phase identifies the soil properties associated with high Arabica coffee productivity and determines the relevant remote sensing datasets from an agricultural perspective.

In addition, a comprehensive field survey and soil sampling campaign were conducted to establish the relationship between in situ data and satellite imagery. Finally, the third phase quantifies soil fertility by integrating multiple parameters—physical and chemical soil properties and nutrient contents—to generate an SFI raster map and database. These outputs serve as decision-support tools for farmers and agricultural planners. Figure 1 presents the overall workflow of the proposed methodology.

2.1. Phase 1: Optimal Region Delimitation

This phase is subdivided into three steps. First, the analysis area is defined, and environmental variables are extracted from satellite imagery datasets accessible through public databases such as NASA and NOAA [61,62,63]. Depending on the level of analysis, two methodological alternatives can be applied: (i) a simplified approach considering elevation, rainfall, and mean annual temperature; or (ii) a detailed approach including up to 22 environmental variables.

The next step involves data preprocessing and curation, including filling missing values using spatial regression techniques, extracting relevant data for the study area, and converting them into raster formats. Pixel resolution is a key factor influencing subsequent analyses, as it depends on the source data and can be refined through local surveys to improve model accuracy—though the size of the study area and project resources may constrain it.

Finally, optimal regions are identified through a reclassification process based on the selected environmental criteria, followed by combining the reclassified layers. This procedure employs the Maximum Entropy (MaxEnt) model, a machine learning algorithm widely used for species distribution modeling [64,65]. Model performance is evaluated using the Area Under the Curve (AUC), which ranges from 0 to 1, with higher values indicating better accuracy. In this study, regions with AUC > 0.5 were classified as optimal, consistent with prior recommendations [66,67,68,69,70,71]. The complete process is summarized in Figure 2.

2.2. Phase 2: Soil Sampling, Spatial Interpolation, and Deterministic Modeling

The soil sampling and testing phase consists of three main steps (Figure 3). The first step involves selecting sampling sites within the optimal region and defining the required physical and chemical soil properties. Once sampling points are established, a validation step ensures that selected sites correspond to rural and non-infrastructure areas, using the Bare Soil Index (BSI) proposed by [72], expressed as Equation (1):

$$BSI={\displaystyle \frac{\left(SWIR2+R\right)-(NIR+B)}{\left(SWIR2+R\right)+(NIR+B)}}$$

(1)

where SWIR2 = shortwave infrared (Band 12), NIR = near-infrared (Band 8), R = red band, and B = blue band—all from Sentinel-2 imagery [73,74,75,76,77,78,79]. Before analysis, Sentinel-2 Level 1C data were atmospherically corrected using the Sen2Cor processor to obtain surface reflectance (Level 2A). Cloud and shadow masking were performed using the Scene Classification Layer (SCL), and all spectral bands were resampled to 10 m for consistency. Spectral subsets focused on the red-edge, NIR, and SWIR regions, and dry-season imagery was selected to minimize soil moisture interference, thereby enhancing soil spectral sensitivity.

The soil tests quantified physical and chemical properties, including macro- and micronutrients, with emphasis on those most relevant to Arabica coffee. The second step involved conducting laboratory analyses, which provided reference data for comparison with remote sensing values. At this stage, Sentinel-2 Band 12 was prioritized due to its proven correlation with soil attributes [68,79,80,81,82,83,84,85,86].

To model spatial variability, a deterministic interpolation approach—Inverse Distance Weighting (IDW)—was applied instead of a geostatistical method. IDW was chosen for its simplicity, computational efficiency, and suitability for small datasets, as it does not require variogram modeling or assumptions of spatial autocorrelation. These characteristics make it appropriate for heterogeneous agricultural landscapes with sparse sampling density.

The final step establishes the correlation between laboratory soil data and Sentinel-2 spectral information, serving as a validation of remote sensing reliability. When correlation strength fell below the defined acceptance threshold, the IDW interpolation was used to derive soil attributes. Conversely, when significant correlations were identified, soil properties were estimated using fitted regression models derived from spectral relationships.

It is important to note that the correlation analysis in this study was based on bivariate relationships between the SWIR2 band and individual soil attributes. While this approach provides valuable initial insights, it does not fully capture multivariate or nonlinear dynamics. Future work should incorporate advanced modeling techniques, such as Partial Least Squares Regression (PLSR) or Random Forests, to leverage the full spectral dataset and improve predictive accuracy across soil variables [80,87,88,89,90].

2.3. Phase 3: Soil Fertility Index for Arabica Coffee Crops

The third phase identifies soil fertility levels within the optimal region (Figure 4) using results from the previous analyses. The first step is to identify the soil properties most strongly associated with fertility using the Geographically Weighted Regression (GWR) model. Before applying GWR, a resampling of pixel resolution was performed to reduce computational load. In this study, a 500 m hexagonal grid was used, with each cell value representing the average of 100 random points extracted from the raster data within the hexagon. This process was applied across all analyzed soil properties to ensure statistical robustness.

The second step focuses on calculating the Soil Fertility Index (SFI) by defining optimal value ranges for each parameter and applying the following equation, Equation (2):

$$SFI={\sum }_{i=1}^n{w}_i·P_i$$

(2)

where $P_i$ represents the binary state (1 = within optimal range; 0 = otherwise) and $w_i$ the corresponding weight.

In this study, the SFI was computed using five macronutrients—nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), and magnesium (Mg)—each assigned an equal weight ($w_i=0.20$) to ensure balanced contribution. Sulfur (S) was excluded due to missing data. Although 23 soil parameters were analyzed overall, only those with complete datasets and agronomic relevance were used in the index to ensure methodological rigor and avoid bias.

The final product is a spatially explicit raster map depicting soil fertility distribution, offering a practical tool for precision agriculture and sustainable management in Arabica coffee systems.

3. Results

3.1. Phase 1: Optimal Region Delimitation

Lonya Grande district was selected as the study area, located in the Utcubamba province of the Peruvian Amazon. This region stands out as the main coffee producer in the department (SEIA [91]), covering approximately 365 square kilometers (≈36,500 ha) and a perimeter of about 87 km. Figure 5 displays the location of the study area.

The optimal region was determined using the MaxEnt model, which accounted for 19 environmental and 3 physical variables. The former are related to temperature and precipitation, while the latter are associated with terrain elevation, land cover, and vegetation index.

Table 1. List of the Environmental aspects utilized for estimating the Optimal region.

Aspect		Units	Range Value		Scale	Description
Name	Type		Min.	Max.
bio01	Environmental	°C	−29	32	0.1	Annual Mean Temperature
bio02		°C	0.9	21.4	0.1	Mean Diurnal Range (Maximum–Minimum Temperature)
bio03		%	7	96	1	Isothermality (bio02/bio07 × 100)
bio04		°C	0.62	227.21	0.01	Temperature Seasonality (Standard Deviation × 100)
bio05		°C	−9.6	49	0.1	Maximum Temperature of the Warmest Month
bio06		°C	−57.3	25.8	0.1	Minimum Temperature of the Coldest Month
bio07		°C	5.3	72.5	0.1	Annual Temperature Range (bio05–bio06)
bio08		°C	−28.5	37.8	0.1	Mean Temperature of the Wettest Quarter
bio09		°C	−52.1	36.6	0.1	Mean Temperature of the Driest Quarter
bio10		°C	−14.3	38.3	0.1	Mean Temperature of the Warmest Quarter
bio11		°C	−52.1	28.9	0.1	Mean Temperature of the Coldest Quarter
bio12		mm	0	11,401	1	Annual Precipitation
bio13		mm	0	2949	1	Precipitation of the Wettest Month
bio14		mm	0	752	1	Precipitation of the Driest Month
bio15		Coef. of Var.	0	265	1	Precipitation Seasonality
bio16		mm	0	8019	1	Precipitation of the Wettest Quarter
bio17		mm	0	2495	1	Precipitation of the Driest Quarter
bio18		mm	0	6090	1	Precipitation of the Warmest Quarter
bio19		mm	0	5162	1	Precipitation of the Coldest Quarter
Elevation	Physical	m	−444	8806	1	Terrain elevation
NDVI		-	−2000	10,000	0.0001	Normalized difference vegetation index
LC_Type4		-	-	-	-	Land cover type 4: Annual BGC (Biogeochemical-Argo) classification

lists the variables used, derived from the WORLDCLIM/V1/BIO [93], CGIAR/SRTM90_V4 [94], MODIS/061/MOD13A1 [95], and MODIS/061/MCD12Q [96] datasets. This information was obtained via Google Earth Engine [97], which served as the data source for ecological niche modeling, providing key insights into temperature and precipitation patterns that influence species distribution. Finally, once the information was reported in raster format, pixel resolution standardization was also performed.

Temperature-related variables (i.e., bio01 to bio11) capture annual, seasonal, and extreme thermal conditions, including diurnal range, seasonality, and specific temperature values for the warmest, coldest, wettest, and driest quarters. Meanwhile, precipitation variables (i.e., bio12 to bio19) quantify total, seasonal, and extreme rainfall conditions, which are essential for understanding moisture availability and hydrological stability across different ecosystems. Physical variables capture the elevation above sea level suitable for crop growth (i.e., elevation variable). Likewise, vegetation concentration in the region is represented by the Normalized Difference Vegetation Index (NDVI), which can be combined with land-use data through the annual biogeochemical classification (LC_Type4 variable). It is important to note that bioclimatic and physical predictors are crucial for MaxEnt modeling, allowing for precise habitat suitability assessments based on climatic constraints. Their integration helps to identify the environmental factors driving species distribution in climate-sensitive regions.

Although sunshine hours and accumulated temperature are also known to influence crop phenology, these variables were not included because the model focused on estimating spatial variation in soil fertility rather than modeling plant growth processes. Future studies could incorporate such variables in broader agroecological suitability assessments.

Previous studies have identified the optimal agroclimatic ranges for Arabica coffee. For example, Ballesteros [98] states that Arabica coffee grows best at elevations between 800 and 2100 m above sea level and temperatures around 17–23 °C. Likewise, Parada et al. [99] provide ranges for total annual precipitation and multiple temperature statistics, including maximum, mean, minimum, and extreme values (

Table 2. Recommended agroclimatic variables ranges for Arabica coffee crops retrieved from Parada et al. [99].

Variable	Unit	Optimal Range
Total annual precipitation	mm	1200–1800
Maximum temperature	°C	21–30.6
Mean temperature	°C	17–23
Minimum temperature	°C	10–19.9
Max. extreme temperature	°C	>33
Min. extreme temperature	°C	7–10

). In addition, Figure 6 illustrates the Digital Elevation Model (DEM), annual average temperature, and total annual rainfall for the region of interest.

Afterward, the distribution model for Arabica coffee crops was obtained using the MaxEnt algorithm, considering 22 environmental and physical variables (Table 1), as shown in Figure 7. Subsequently, the optimal regions for Arabica coffee growth were defined as those exhibiting an AUC value greater than 0.5. Values below this threshold indicate that the model does not surpass random prediction, implying a lack of discriminative capacity [66]. Overall, the optimal region covers approximately 52% of the total study area (~191 km²).

3.2. Phase 2

A total of twelve sampling sites were selected for soil collection, representing the four main urban centers in the district: Lonya Grande, Nueva York, Ortiz Arrieta, and Roblepampa. Three rural sampling points were selected for each urban area and validated using the BSI map derived from Sentinel-2 remote sensing data and Equation (1). The Sentinel-2 imagery used in this study was acquired between May and August 2022, corresponding to the dry season in the study area. This period was chosen to minimize cloud interference and soil moisture variability. The data underwent atmospheric correction using the Sen2Cor processor to obtain Level-2A surface reflectance products. Additionally, cloud and shadow pixels were masked using the Scene Classification Layer (SCL), and all bands were resampled to 10 m resolution to ensure consistency in the analysis.

At each site, three soil subsamples were collected within a 10 m radius and subsequently homogenized to form a composite sample, ensuring spatial representativeness and minimizing local variability. In total, 36 individual samples were taken, providing a robust basis for statistical analysis. While administrative boundaries ensured broad spatial coverage, the specific locations within each region were selected based on spectral variability captured by the BSI, enhancing the representativeness and calibration strength of the remote sensing analysis. Figure 8 shows the results for Step 1 of this phase.

Table 3. Soil parameters assessed.

ID	Soil Parameter	Type	Symbol	Units	Testing Method
PG-1	Sand	Physical	-	%	Bouyoucos
PG-2	Silt		-	%
PG-3	Clay		-	%
PG-4	Saturation		Sw	%	Gravimetric
PG-5	Electrical Conductivity	Chemical	CE	dS/m	Electrochemical
PG-6	pH in H2O		pH_H2O	-
PG-7	pH in CaCl2		pH_CaCl2	-
PG-8	Organic Matter		MO	%	Walkley & Black
PG-9	Field Capacity	Physical	CC	%	Richards
PG-10	Permanent Wilting Point		PWP	%
PG-11	Cation Exchange Capacity	Chemical	CEC	meq/100 g	Mathematical Calculation
MN-1	Nitrogen	Macronutrient	N	ppm	Kjeldahl
MN-2	Phosphorus		P	ppm	Bray I
MN-3	Potassium		K	ppm	Ammonium Acetate
MN-4	Sulfur		S	ppm	CH3COONH4
MN-5	Calcium		Ca	meq/100 g	FAAS
MN-6	Magnesium		Mg	meq/100 g
mn-1	Sodium	Micronutrient	Na	Ppm
mn-2	Copper		Cu	Ppm
mn-3	Zinc		Zn	ppm
mn-4	Manganese		Mn	ppm
mn-5	Iron		Fe	ppm
mn-6	Boron		B	ppm	Colorimetric

FAAS: Flame Atomic Absorption Spectroscopy.

presents a comprehensive characterization of the soil, encompassing key physical, chemical, and nutritional parameters essential for sustainable Arabica coffee production. It includes textural properties—sand, silt, and clay —which are crucial for soil porosity and hydrodynamics, along with key indicators such as saturation (Sw) and electrical conductivity (EC), which reflect water availability and salinity levels. Additionally, the pH in water ($H_{2}O$) and calcium chloride ($CaC{l}_2)$ solutions were analyzed, as they play a fundamental role in nutrient availability and colloidal interactions. Moreover, organic matter (OM), field capacity (FC), and permanent wilting point (PWP) were assessed to evaluate water retention capacity and overall soil fertility. Finally, the effective cation exchange capacity (CEC) was incorporated as a critical indicator of the soil’s ability to retain and supply essential nutrients.

From a nutritional perspective, macro- and micronutrients were quantified using advanced analytical techniques, including Flame Atomic Absorption Spectrometry (FAAS), along with high-precision gravimetric and electrochemical methods. These analyses provide a rigorous scientific framework for optimizing agronomic management and support data-driven decision-making on nutrient availability, water retention, and soil ionic dynamics—factors crucial to maintaining the sustainability and productivity of high-value Arabica coffee cultivation systems.

3.2.1. Results from In-Field Surveys

Figure 9a illustrates the box plot evaluating the general soil parameters (11 in total), including physical properties (sand, silt, clay), moisture-related parameters (saturation, field capacity, and permanent wilting point), and chemical properties (electrical conductivity,$pH$ in $H_{2}O$ and $CaC{l}_2,$ organic matter, and cation exchange capacity). The data are presented on a logarithmic scale, revealing significant variability among samples and emphasizing the heterogeneous nature of soil properties across the study area. Likewise, Figure 9b shows the box plot displaying the distribution of macronutrients (parameters MN-1 to MN-6) and micronutrients (parameters mn-1 to mn-6) using the same logarithmic scale.

Figure 9a confirms that the sand and clay contents (parameters PG-1 and PG-3) display the widest interquartile ranges among physical properties, indicating pronounced textural differences across the sampled locations. This directly affects soil permeability and water retention. Conversely, saturation (PG-4) and field capacity (PG-9) exhibit consistent distributions with high median values, reflecting their importance in maintaining adequate moisture levels for Arabica coffee production. In contrast, electrical conductivity (PG-5) shows minimal variability and low median values, confirming that salinity is not a limiting factor for crop growth in this region. Finally, the organic matter (PG-8) parameter exhibits moderate dispersion, underscoring its role in improving soil structure and nutrient cycling.

To quantify the degree of spatial heterogeneity among key soil nutrients, the Coefficient of Variation (CV) was calculated for phosphorus (P), potassium (K), and organic matter (OM), yielding values of 46.8%, 51.2%, and 42.5%, respectively. These moderate to high CV levels confirm substantial variability across the study area, warranting the application of geospatial interpolation methods, such as IDW, to ensure spatially explicit nutrient management. Furthermore, when compared to agronomic thresholds established for Arabica coffee cultivation (e.g., P: 35–70 mg/kg; K: 150–300 mg/kg; OM: > 2.5%), the mean values for both P and K were below recommended levels, while OM was within acceptable limits. These results point to localized nutrient deficiencies that require site-specific fertilization strategies to optimize crop health and productivity.

Among the macronutrients, calcium (Ca) and potassium (K) show wide variability, with potassium exhibiting the most extensive range, suggesting spatial differences in nutrient management practices or variations in soil mineralogy, as evident in Figure 9b. Meanwhile, nitrogen (N) and phosphorus (P)—both critical for plant growth—show narrower interquartile ranges but consistently low median values, highlighting potential nutrient deficiencies in the area. Regarding micronutrients, iron (Fe) and manganese (Mn) exhibit the greatest variability and highest median values, reflecting their dynamic availability across different soil conditions. Zinc (Zn) and boron (B) show relatively stable distributions but low overall concentrations, suggesting limited availability and potential impacts on coffee plant development.

Figure 10 and Figure 11 illustrate the alignment of eleven soil parameters (sand, silt, clay, saturation, electrical conductivity, pH H₂O, pH CaCl₂, organic matter, field capacity, PWP, and CEC) and all macro- and micronutrients with a normal distribution, validating the statistical consistency of the data. Physical parameters such as sand (R² = 0.96), silt (R² = 0.97), and clay (R² = 0.96) exhibit high consistency, confirming uniform distribution across the analyzed samples despite the natural heterogeneity of the study area. This is essential for modeling soil texture, which directly influences water retention and nutrient availability. Moisture-related parameters, including saturation (R² = 0.82) and field capacity (R² = 0.8243), show moderate correlations, reflecting the influence of management practices and soil composition on water dynamics—critical for Arabica coffee cultivation.

Regarding chemical properties, pH (H₂O) ($R^2=0.95$) and pH (CaCl₂) ($R^2=0.93$) confirm moderately acidic soil conditions, optimal for Arabica coffee growth. Organic matter ($R^2=0.96$) also shows strong alignment with normality, emphasizing its role in enhancing soil fertility and structural stability. However, electrical conductivity ($R^2=0.73$) and cation exchange capacity (CEC, $R^2=0.89$) exhibit greater variability, reflecting complexities in soil chemistry linked to mineral composition and localized salinity. Overall, these findings indicate that the evaluated parameters are statistically consistent and suitable for integration into spatial analysis models such as GWR. However, some factors—particularly CEC and electrical conductivity—may require specific adjustments to improve predictive accuracy.

As a general remark, while box plots emphasize variability and dispersion, probability plots focus on validating the statistical normality of the data and detecting potential biases. The strong alignment of parameters such as sand, clay, pH, and organic matter with the normal distribution line ($R^2>0.95$) underscores their consistency and relative homogeneity, highlighting their robustness as predictive variables within the GWR model applied in later stages.

Figure 12 presents the correlation matrix, providing critical insights into the interactions among soil physical, chemical, and nutrient properties and offering a comprehensive understanding of their dynamics in the context of Arabica coffee production.

On the one hand, physical parameters such as sand, silt, and clay exhibit strong inverse correlations ($r\approx -0.9$), reflecting the characteristic textural balance of soils. Moisture-related parameters, including saturation, field capacity, and permanent wilting point, show moderate to strong correlations ($r\approx 0.7–0.8$), underscoring their importance in water retention and soil resilience under drought conditions, which are crucial for sustaining coffee yields. These relationships emphasize the combined influence of texture and moisture content on soil functionality.

On the other hand, chemical properties reveal significant interactions, with pH (measured in H₂O and CaCl₂) displaying strong internal correlations ($r>0.9$), confirming the stability of soil acidity. Organic matter and cation exchange capacity (CEC) exhibit a moderate positive correlation (r ≈ 0.6), highlighting their joint role in nutrient retention and soil fertility. Among micronutrients, zinc and manganese show moderate positive correlations ($r\approx 0.6$), whereas iron is negatively correlated with pH (r ≈ −0.4), reflecting pH influence on Fe availability. The weaker correlations observed for nitrogen and potassium suggest greater mobility and susceptibility to leaching in highly weathered soils, underscoring the need for targeted nutrient management strategies.

3.2.2. Data Performance/Validation

Figure 13 presents the spatial information for the optimal region derived from the SWIR2 band of Sentinel-2 satellite imagery, while Figure 14 illustrates the evaluation of physical and chemical soil properties based on 12 samples collected from different locations. Their relationship was established using the SWIR2 band reflectance from Sentinel-2 imagery. These properties were used for cross-validation and calibration of the Geographically Weighted Regression (GWR) model in the context of Arabica coffee cultivation in Lonya Grande, Amazonas. The results highlight the capability of the SWIR2 band to capture key soil variations, particularly moisture-related properties that are critical for agricultural management.

Figure 13 presents the spatial information for the optimal region derived from the SWIR2 band of Sentinel-2 satellite imagery, while Figure 14 illustrates the evaluation of physical and chemical soil properties based on 12 samples collected from different locations. Their relationship was established using the SWIR2 band reflectance from Sentinel-2 imagery.

The SWIR2 reflectance values were extracted from atmospherically corrected Sentinel-2 Level-2A imagery (using the Sen2Cor algorithm). Each reflectance value corresponds to a georeferenced sampling point, matched through bilinear interpolation. The correlation plots use normalized surface reflectance (X-axis) against lab-measured soil properties (Y-axis). Axis labels and units have been included in the updated figure to ensure clarity and reproducibility.

These properties were used for cross-validation and calibration of the Geographically Weighted Regression (GWR) model in the context of Arabica coffee cultivation in Lonya Grande, Amazonas. The results highlight the capability of the SWIR2 band to capture key soil variations, particularly moisture-related properties that are critical for agricultural management.

Results from Figure 14 indicate that soil texture—represented by sand, silt, and clay fractions—exhibits weak to moderate correlations with the SWIR2 band ($R^2=0.22$ for sand and $R^2=0.21$ for clay). These correlations suggest that, although SWIR2 partially reflects physical soil properties, its predictive capacity for these variables remains limited and may require complementary spectral bands. Conversely, soil saturation (Sw) shows a strong positive correlation ($R^2=0.44$), highlighting the SWIR2 band’s sensitivity to variations in soil moisture content. This property is especially critical for coffee production because it directly influences root-zone water dynamics.

Regarding chemical properties, pH values (in H₂O and CaCl₂) and cation exchange capacity (CEC) display weak correlations ($R^2\approx 0.05–0.06$), reflecting their indirect influence on SWIR2 spectral reflectance. However, organic matter (OM) demonstrates a moderate correlation ($R^2=0.05$), indicating its impact on soil optical behavior. Additionally, field capacity (FC) and permanent wilting point (PWP) exhibit significant correlations ($R^2=0.41$ and $R^2=0.38$, respectively), underscoring their importance in water retention dynamics and their connection to agricultural productivity.

Similarly, Figure 15 presents the analysis of 12 soil samples, revealing key insights into the relationships between macro- and micronutrients and the SWIR2 band of Sentinel-2 imagery. These relationships were assessed using regression models to quantify the spectral responses of each nutrient, serving as the basis for validating the GWR model in the study area. The results are displayed for macronutrients (columns 1 and 2) and micronutrients (columns 3 and 4), emphasizing their relevance to soil fertility and coffee cultivation in Lonya Grande. Finally, Figure 15 highlights the following findings:

• Macronutrients: Nitrogen (N) shows a weak positive correlation with the SWIR2 band ($R^2=0.05$), while phosphorus (P) also exhibits a minor positive trend ($R^2=0.07$). These limited correlations indicate that, although these elements may influence spectral reflectance, their contribution is overshadowed by other soil properties. Potassium (K), sulfur (S), and calcium (Ca) show stronger negative correlations ($R^2=0.15, 0.02$, and $0.07$, respectively), suggesting that these nutrients affect soil reflectance through their influence on soil structure and surface interactions. Magnesium (Mg) displays the weakest correlation ($R^2=0.05$), indicating a minimal impact on SWIR2 spectral variability.

Although phosphorus (P) is not spectrally active and does not directly influence reflectance in the visible or infrared regions, its availability in the soil significantly impacts plant physiological processes, which are detectable through vegetation indices such as NDVI. Phosphorus plays a critical role in energy transfer (ATP), root development, and photosynthesis efficiency. When phosphorus levels are sufficient, crops tend to develop more vigorous root systems and denser canopies, enhancing chlorophyll production and biomass accumulation.

This improvement in vegetative vigor results in increased reflectance in the near-infrared band (due to greater cellular structure and leaf area), leading to higher NDVI values. Therefore, the observed correlation between NDVI and P content is indirect: elevated P availability improves plant health and biomass, which in turn enhances the spectral signal captured by NDVI. Similar indirect relationships between vegetation indices and soil nutrient status have been documented in precision agriculture studies, particularly in crops such as maize and wheat, where NDVI has served as a proxy for nutrient-induced changes in canopy development. Hence, although P itself is not optically active, its physiological effects on plant growth create measurable impacts on spectral indices.

• Micronutrients: The spectral response of micronutrients revealed notable variations. Sodium (Na) and zinc (Zn) exhibited weak correlations ($R^2=0.05$ and $R^2=0.02$), while copper (Cu) and boron (B) showed negligible correlations ($R^2=0.04$ and $R^2=0.03$). Conversely, manganese (Mn) demonstrated a stronger inverse relationship ($R^2=0.1699$), indicating its significant contribution to soil reflectance in the SWIR2 band range. Iron (Fe) with a $R^2=0.04$, reflected a weak correlation, possibly due to its indirect influence on soil surface properties.

Table 4. Summary of the correlation between soil samples and the SWIR2 band of Sentinel-2.

ID	Soil Parameter	${\mathit{R}}^2$	Correlation	ID	Soil Parameter	${\mathit{R}}^2$	Correlation
	General Soil Parameters				Macronutrients
PG-1	Sand	0.2232	Weak	MN-1	Nitrogen (N)	0.0451	none
PG-2	Silt	0.1201	none	MN-2	Phosphorus (P)	0.0664	none
PG-3	Clay	0.2077	Weak	MN-3	Potassium (K)	0.1513	Weak
PG-4	Saturation (Sw)	0.4043	Moderate	MN-4	Sulfur (S)	0.0164	None
PG-5	Electrical Conductivity (CE)	0.1814	Weak	MN-5	Calcium (Ca)	0.0746	none
PG-6	pH (H2O)	0.0494	none	MN-6	Magnesium (Mg)	0.0472	none
PG-7	pH (CaCl2)	0.0598	none		Micronutrients
PG-8	Organic Matter (MO)	0.0489	none	mn-1	Sodium (Na)	0.0462	none
PG-9	Field Capacity (CC)	0.0143	None	mn-2	Copper (Cu)	0.0457	none
PG-10	Permanent Wilting Point (PWP)	0.0137	None	mn-3	Zinc (Zn)	0.0232	None
PG-11	Cation Exchange Capacity (CEC)	0.0558	none	mn-4	Manganese (Mn)	0.1899	Weak
				mn-5	Iron (Fe)	0.0375	none
				mn-6	Boron (B)	0.0292	none

Correlation Classification Criteria: $R^2>0.5\to Strong$; $0.3<R^2\le 0.5\to Moderate$; $0.1<R^2\le 0.3\to Weak$; $R^2\le 0.1\to None.$

summarizes the calculated correlations for all 23 assessed parameters, followed by a classification process into four discrete levels: strong, moderate, weak, and none based on their values. An R² greater than 0.5 indicates a strong correlation. A moderate level corresponds to values greater than 0.3 and less than or equal to 0.5; weak values range from 0.1 to 0.3. Finally, those lower than or equal to 0.1 are assigned a “non-level” label.

Table 4 highlights that most soil parameters exhibit very weak or no correlation with the SWIR2 band of Sentinel-2 reflectance.

3.2.3. IDW

Then, soil parameters for the entire optimal region were obtained using spatial analysis and a spatial interpolation technique, since there was no strong correlation between the sample results and the SWIR2 band from Sentinel-2 imagery. At this stage, the Inverse Distance Weighted (IDW) method was applied to generate raster maps for the complete set of 23 variables, due to its simplicity and straightforward implementation process [32,100]. A 2 m pixel resolution was defined in the calculation process, using the 12 soil samples as baseline data. Figure 16, Figure 17 and Figure 18 show the results for the macronutrients, general soil parameters, and micronutrients, respectively.

The spatial distribution of Organic Matter (OM) exhibited notably higher concentrations in the northern sector of the study area, particularly in areas with gentle slopes and dense vegetation cover. This pattern aligns with NDVI and DEM-derived data, suggesting that reduced slope and enhanced plant biomass accumulation contribute to lower erosion risks and greater retention of organic residues. These areas also coincide with sites managed under traditional agroforestry systems, as confirmed by field interviews and land-use data, where composted coffee pulp and natural litter are more frequently applied—thereby enhancing OM levels.

Conversely, lower OM and phosphorus (P) concentrations were predominantly found in the southern and southwestern zones, characterized by steep gradients and intensive monoculture practices. These conditions favor nutrient leaching and surface runoff, reducing fertility and long-term sustainability. Potassium (K), while more spatially scattered, exhibited slight elevation-related trends, with relatively lower concentrations in high-rainfall, low-lying zones—likely due to increased leaching.

These spatial patterns reveal that soil nutrient variability is not random but is closely linked to environmental covariates, including slope, elevation, vegetation index, and land management history. Such relationships underscore the importance of integrating environmental layers in spatial modeling, enabling a more accurate understanding of nutrient dynamics and their implications for sustainable Arabica coffee cultivation.

3.3. Phase 3

3.3.1. GWR

Once this stage was reached (before any regression analysis), the data were resampled. The raster maps for the 23 variables were converted into a grid format with hexagon-shaped cells. As a result, the optimal region was represented by 716 hexagons, each measuring 500 m in width and height (area: 21.6 ha). A hexagonal grid structure was selected for its geometric advantages in spatial analysis and interpolation, including uniform neighbor distances and minimized directional bias. The cell size was chosen to match the spatial resolution of Sentinel-2 data (10–20 m) and the spatial support of soil sampling points, as recommended by [101,102]. Subsequently, the assigned parameter value for each hexagon was calculated as the mean of 100 random points selected from the IDW raster map contained within that hexagon. The base condition required that each selected point be at least 35 m apart from the others to avoid bias, and this procedure was applied to all 23 variables.

In the next step, the Geographically Weighted Regression (GWR) model was executed, using a cross-validation (CV) process between the SWIR2 band from Sentinel-2 and the 23 parameters for each sampling site, followed by an adjustment based on the Akaike Information Criterion (AIC). Seven models were evaluated and are summarized in

Table 5. GWR model analyzed.

ID	Case	CV (Max)	AIC (Min)	R2	SWIR2 Function f (Variables)
GWR-01	Macronutrients—Case 0	572.12	602.80	0.9393	f (MN-1, MN-2, MN-3, MN-4, MN-5, MN-6)
GWR-02	Macronutrients—Case 1	435.42	466.33	0.9338	f (MN-1, MN-2, MN-3)
GWR-03	Macronutrients—Case 2	434.04	480.89	0.9342	f (MN-4, MN-5, MN-6)
GWR-04	Micronutrients	804.72	958.62	0.8991	f (mn-1, mn-2, mn-3, mn-4, mn-5, mn-6)
GWR-05	Physical	318.44	439.27	0.9616	f (PG-4, PG-5)
GWR-06	Organic Matter	280.48	403.33	0.9720	f (PG-8)
GWR-07	Acidity	324.39	438.01	0.9608	f (PG-6, PG-7)

. The first three models (GWR-01, GWR-02, GWR-03) used macronutrients as the best estimators of the SWIR2 band, whereas GWR-04 used micronutrients. Likewise, GWR-05 incorporated only physical parameters, and GWR-06 and GWR-07 used organic matter and acidity, respectively. The variables selected for each model were based on expert judgment. GWR statistical analyses were conducted using R version 4.4.2 [103], and the maps were processed and edited in QGIS 3.6.2 [104].

According to Table 5, the model with the highest performance is the GWR-06, with the highest R² (i.e., 0.9616) and the minimum AIC value (i.e., 403), and relying on a single variable, making it easy to apply. However, it is considered that the CV of the model should be the highest. Considering this, the models that adapt best are those of Macronutrients and Micronutrients (GWR-01 to GWR-04). In addition to being the most complex considering the number of variables used to validate it, but according to performance, only the Macronutrients model reaches R² above 90% and has the lowest AIC compared to that of Micronutrients. Therefore, the GWR-01 predictive model generated by the relationship of Macronutrients and SWIR2 was chosen. The output results are summarized in Figure 19 and Figure 20. The former depicts the model performance through statistical variables, and the second one shows the maps with the normalized factor assigned to each parameter.

3.3.2. Soil Fertility Index Map

The macronutrients variability from the maps obtained at phase 2 and adjusted through grid resampling in the first step of this phase is illustrated in Figure 21. It shows the log-scaled boxplot for each parameter identified from the GWR-01 model, MN-1 to MN-6. In addition, the optimal reference values for macronutrients identified from GWR-01 model are summarized in

Table 6. Summary of optimal ranges for the parameters from the GWR-01 model.

Code	Parameter	Units	Reference Values	Category
MN-1	Total Nitrogen (N)	%	0.1–0.2	-
MN-3	Assimilable Potassium (K)	mg/kg	150–300	-
MN-2	Olsen Phosphorus (P)	mg/kg	35–70	-
-	Burriel Phosphorus	mg/kg	43.6–109	-
MN-3	Exchangeable Potassium (K)	meq/100 g	0.5–1.2	Exchangeable cations
MN-5	Calcium (Ca)	meq/L	11.0–25.0	Saturated extract
MN-6	Magnesium (Mg)	meq/L	6.0–14.0	Saturated extract
-	Sodium	meq/L	4.0–17.0	Saturated extract
MN-3	Potassium (K)	meq/L	1.0–5.0	Saturated extract
-	Exchangeable Ca/Mg Ratio	-	4.0–6.0	Exchangeable cations
-	Exchangeable K/Mg Ratio	-	0.3–0.8	Exchangeable cations
-	Ca/Mg Ratio	-	1.5–3.0	Saturated extract
-	K/Ca Ratio	-	0.15–0.25	Saturated extract
-	K/Mg Ratio	-	0.3–0.8	Saturated extract

, highlighting the importance of soil chemical balance and nutrient availability in agricultural fertility related to Arabica coffee crops. That information was retrieved from [105]. For this case, only 5 of the 6 parameters have reliable information, excluding the sulfur (MN-4). These results provide a framework for evaluating the suitability of the GWR-01 predictive model and for optimizing soil management practices to improve agricultural productivity and maintain edaphic ecosystem sustainability.

From Table 6, it can be observed that total nitrogen (MN-1) falls within a narrow range (0.1–0.2%), indicating the need for continuous monitoring to prevent production limitations. Phosphorus (MN-2), in its Olsen and Burriel forms, shows a similar trend ensuring efficient nutrient uptake, while assimilable potassium (MN-3) varies between 150 and 300 mg/kg, guaranteeing its availability in the soil. Exchangeable cations show critical relationships between Ca/Mg (4.0–6.0) and K/Mg (0.3–0.8), reflecting the importance of ionic balance in nutrient retention and availability. In the saturated extract, calcium (MN-5) and magnesium (MN-6) exhibit appropriate concentrations to ensure soil structural stability, while potassium (MN-3) in solution ranges from 1.0 to 5.0 $\mathrm{m}\mathrm{e}\mathrm{q}/\mathrm{L}$, which is crucial for plant nutrition. The relationships among Ca/Mg, K/Ca, and K/Mg emphasize the need to maintain balanced proportions to avoid antagonistic effects on nutrient absorption.

When performing the SFI calculation, the weight factor ($wi$) for the five parameters with available information was set equally to ensure uniform probability across all, i.e., $wi = 0.20$. For sulfur, a value of zero was assigned due to the lack of reliable data. Additionally, a sensitivity analysis was conducted to account for variability in results when defining the optimal parameter ranges, as shown in Table 6; the considered variation levels were 0%, 20%, and 40%. Specifically, the lower bound was obtained by multiplying the minimum optimal range value by (1 − variation), while the upper bound was calculated as the maximum optimal range value multiplied by (1 + variation). The resulting outputs are presented in Figure 22, which depicts the SFI maps and their corresponding sensitivity analysis for the optimal study region after applying Equation (2).

Results from Figure 22 confirm that the maximum SFI is 0.60 regardless of the sensitivity analysis cases, meaning that at least 3 out of the 5 parameters are within the optimal ranges recommended by previous studies. For the 0% variation case, this condition represented less than 1% of the optimal area (i.e., ~22 Ha). Likewise, the SFI values of 0.40, 0.20, and 0.00 cover an area of 1750 Ha, 13,694 Ha, and 0 Ha, respectively. Such a condition highlights the need to increase soil fertility by incorporating fertilizers that compensate for the lack of micronutrients. Finally, the variability among the sensitivity analysis cases for the SFI vs. the percentage of area is depicted in Figure 23.

The sensitivity cases do not alter the overall trend in the area represented by SFI values of 0.2 and 0.4, with maximum differences of 9% and 2%, respectively (see Figure 23). On the contrary, the total area with a 0.6 result depicts a high sensitivity to the optimal range definition, increasing by up to 3 times between levels. Finally, it is observed that there are no cases where an SFI value is equal to 0, nor greater than or equal to 0.8. This means that at least one parameter is within the optimal range for the most critical condition, or in the other scenario, three parameters present a nutrient content within the desired intervals.

A comparative assessment between the IDW interpolation maps and the spectral correlation models (GWR–SWIR2) reveals both convergence and complementarity. For example, in the case of organic matter (OM), both methods identified higher values in the northern areas, indicating consistent spatial patterns. However, the IDW maps captured finer local variations around sampling sites, while the spectral models provided smoother gradients reflecting broader environmental influences. Overall, IDW offers higher reliability for localized estimates based on measured points, whereas the spectral–GWR approach is better suited for large-scale prediction of soil fertility trends. Thus, integrating both methods enhances spatial accuracy and supports a more comprehensive understanding of soil variability.

4. Discussion

The heterogeneous nature of soils in the study area is evident from the variability observed in both physical and chemical parameters during the Phase 2 field surveys. This spatial heterogeneity directly influences water retention, nutrient availability, and overall soil fertility. Parameters related to soil moisture (saturation and field capacity) align with the water requirements of Arabica coffee, confirming the region’s suitability for its cultivation. However, the low concentrations and high spatial variability of essential macronutrients (nitrogen, phosphorus, and potassium) underscore the need for targeted and site-specific fertilization strategies to optimize crop productivity. Similarly, the uneven distribution of micronutrients such as zinc and boron emphasizes the importance of localized interventions to improve specific deficiencies. These findings support integrating remote sensing and spatial modeling techniques (e.g., GWR) into precision agriculture frameworks, enabling the detection of soil variability and promoting efficient, data-driven management practices tailored to regional conditions.

The validation process, which compared field measurements with SWIR2 band data from Sentinel-2 imagery, revealed varying degrees of correlation among soil nutrients, emphasizing the selective sensitivity of the SWIR2 band to specific soil properties. Stronger associations with potassium, sulfur, and manganese indicate their greater influence on spectral reflectance. In contrast, nutrients such as copper and boron exhibited negligible correlations, suggesting they are less responsive to SWIR2 reflectance. The main findings can be summarized as follows:

• For general soil parameters, saturation (Sw) exhibited the highest correlation (moderate; see Table 4). This indicates that the SWIR2 band effectively captures variations in soil moisture, consistent with previous studies that have highlighted its sensitivity to soil water content. However, other physical and chemical variables—such as texture (sand, silt, clay) and pH—showed weak or negligible relationships, suggesting that additional spectral bands or data sources are needed to improve predictive power.
• For macronutrients, potassium (K) presented the strongest relationship (weak correlation; Table 4), indicating some potential for SWIR2-based prediction of potassium spatial distribution. In contrast, nitrogen (N), phosphorus (P), sulfur (S), calcium (Ca), and magnesium (Mg) exhibited very weak or no correlation, suggesting that their spatial variability is not adequately captured by SWIR2 data alone.
• For micronutrients, manganese (Mn) demonstrated the highest correlation (weak; Table 4), indicating a modest relationship between the SWIR2 band and Mn concentrations. Other micronutrients—including sodium (Na), copper (Cu), zinc (Zn), iron (Fe), and boron (B)—showed minimal or no correlation, reinforcing the need for multi-band spectral analysis or advanced machine learning methods to enhance soil nutrient prediction.

Overall, these findings indicate that SWIR2 reflectance alone is insufficient for accurately modeling most soil properties, except for soil moisture (Sw) and, to a lesser extent, potassium (K) and manganese (Mn). The results suggest that incorporating other spectral bands—particularly NIR and SWIR1—and applying advanced spatial analysis models could substantially enhance predictive capability. Additionally, local soil heterogeneity and environmental variability likely contribute to the low correlation values, highlighting the importance of future validation with multispectral or hyperspectral data for precision agriculture. Ultimately, integrating spectral and spatial data provides a robust and scalable framework for identifying soil fertility gradients, optimizing resource use, and supporting sustainable Arabica coffee production.

The IDW interpolation method was applied to generate maps for 19 soil parameters within the optimal region, based on 12 field sampling sites. This approach was adopted because there were no strong correlations with SWIR2 data. However, several interpolated maps—particularly for clay, organic matter, nitrogen, phosphorus, potassium, sodium, and iron (Figure 16, Figure 17 and Figure 18)—showed spatial patterns strongly influenced by the limited number of input samples. This limitation suggests that future studies should increase sampling density to better capture fine-scale variability, consistent with the high heterogeneity observed in the field data.

Conversely, other parameters displayed more uniform spatial distributions across the study area. Despite the limitations of sampling density, the generated maps provide valuable spatial insights for soil management, offering farmers practical alternatives to traditional field measurements for identifying suitable zones for coffee production. Therefore, future research should aim to expand soil sampling coverage and incorporate additional explanatory variables—such as topography or proximity to urban areas—to refine crop suitability assessments and optimize resource allocation.

The selected GWR-01 model demonstrated the strongest overall performance, with results showing high dependence across all six macronutrients (MN-1 to MN-6). This model achieved the highest cross-validation (CV = 572.12), coefficient of determination (R² = 93.93%), and acceptable Akaike Information Criterion (AIC = 602.80). Although some models using general soil parameters or micronutrients yielded slightly higher R² values, they were excluded due to lower cross-validation performance. One model (GWR-06) outperformed GWR-01 in terms of statistical fit but relied solely on organic matter, which would bias results and fail to represent the multi-nutrient complexity essential for Arabica coffee cultivation.

Subsequently, a cross-validation procedure was conducted using the macronutrient variables from GWR-01 to generate the Soil Fertility Index (SFI) map. The classification was based on recommended nutrient thresholds for Arabica coffee, though reference ranges were available for only five of the six nutrients. The sensitivity analysis showed a maximum SFI value of 0.60, indicating that at least three of the five parameters met optimal conditions in roughly 10% (~19 km²) of the study area. Furthermore, the results revealed that at least one macronutrient consistently fell within the optimal range across most of the region. These findings reinforce the role of macronutrients as key indicators of soil fertility and agronomic performance, complementing traditional environmental factors such as temperature, elevation, and precipitation.

5. Conclusions

This study proposes a well-structured and practical methodological framework for large-scale soil fertility assessment in coffee cultivation. By integrating satellite imagery with field survey campaigns, the proposed methodology combines Inverse Distance Weighting (IDW) and Geographically Weighted Regression (GWR) models to predict the suitability of agricultural regions. The approach is organized into three sequential phases, utilizing Sentinel-2 SWIR2 band data as the primary input, which has shown a strong correlation with key soil fertility indicators.

The case study was carried out in the Lonya Grande region of the Peruvian Amazon, covering approximately 365 km². Field surveys and soil analyses were conducted to assess the availability of macro- and micronutrients, providing a comprehensive dataset for model calibration and validation. This integration ensured high accuracy in soil fertility estimation and enabled the identification of optimal zones for coffee cultivation.

Soil property analysis revealed significant heterogeneity within the study area. The results confirmed that the soils possess adequate water retention capacity, with stable values for saturation (Sw), field capacity (FC), and permanent wilting point (PWP), indicating favorable conditions for coffee growth. However, deficiencies in essential macronutrients such as nitrogen (N), phosphorus (P), and potassium (K) were identified, potentially limiting yields and highlighting the need for targeted fertilization strategies. Among the micronutrients, iron (Fe) and manganese (Mn) exhibited the greatest variability, emphasizing the importance of differentiated soil management to optimize nutrient uptake and improve crop productivity.

A key finding of this study is that traditional methods for assessing soil suitability—based on general parameters such as terrain elevation, rainfall, and average temperature—fail to capture the spatial distribution of essential nutrients. This limitation significantly reduces their accuracy in predicting soil fertility for specific crops. In contrast, the proposed methodology successfully generated a high-resolution raster-based nutrient distribution map, offering a valuable tool for precision agriculture and data-driven land management.

Spatial analysis confirmed the reliability of the soil sampling network, showing strong correlations with satellite-derived data. On the one hand, the soil test results for variables such as sand ($R^2 = 0.96$), silt ($R^2 = 0.97$), and clay ($R^2 = 0.96$) exhibited consistent distribution, supporting their integration into predictive soil fertility models. However, parameters such as cation exchange capacity (CEC) and electrical conductivity (EC) displayed higher variability, with values of $R^2 = 0.89$ and $R^2 = 0.73$, respectively, reflecting the complexity of nutrient retention and transport processes in the soil.

On the other hand, when trying to capture the relationships among soil sampling results and the SWIR2 band using traditional statistical methods, there was no clear correlation. Nonetheless, after applying the spatial analysis GWR model, all macronutrients were identified as the most suitable variable that fits with the SWIR2 band, reaching a $R^2 = 0.94$. In addition, the soil fertility index displayed a high dependency on the optimal range for the assessed parameters, focused on a high precision when predicting the physicochemical soil features.

From an applied perspective, the integration of satellite-based remote sensing with spatial analysis modeling represents a significant advancement in agricultural soil monitoring and management. This methodology enables precise soil fertility assessment in remote areas, facilitating data-driven decision-making for crop management and resource allocation. It is recommended that this approach be adopted in regional and national farming programs to optimize and supervise fertilizer application, enhance productivity, and promote sustainable coffee production. Furthermore, the incorporation of advanced machine learning models could further improve predictive accuracy and refine soil management practices across the agricultural sector.