Integration of Machine Learning and Remote Sensing to Evaluate the Effects of Soil Salinity, Nitrate, and Moisture on Crop Yields and Economic Returns in the Semi-Arid Region of Ethiopia

Abstract

Soil salinity, soil moisture, and nutrient loss significantly reduce agricultural productivity and economic benefits in the semi-arid regions of Ethiopia. However, knowledge of how to mitigate these risks remains limited. This study examined the combined effects of salinity (EC), soil moisture (Sm), and nitrate (N) on the yield and profitability of banana, cotton, and maize using field-based and satellite data with seven machine learning algorithms. Our results showed that a higher EC level reduced crop yields, whereas sufficient Sm and N improved productivity and income. Among the models, Random Forest (RF) performed the best, achieving high accuracy (e.g., R² = 0.998 for cotton, 0.869 for banana, and 0.793 for maize). SHapley Additive exPlanations (SHAP) analysis further identified EC as the most critical determinant, highlighting the priority of salinity mitigation, alongside water and nutrient management. These findings provide farmers and decision-makers with practical insights into how to sustain crop productivity, improve livelihoods, and strengthen food security in semi-arid regions.

1. Introduction

Soil salinity and related environmental stresses increasingly undermine agricultural productivity and smallholder livelihoods, particularly in semi-arid regions [1,2,3]. A global modeling study estimated annual food losses of approximately 12.4 × 10¹⁴ kilocalories due to elevated water salinity, which is enough to feed 170 million people daily [4]. Field-level studies have revealed even more severe localized impacts: salinity can halve irrigated yields, devalue land, and lead to abandonment [5,6]. These findings point to the importance of locally adapted evaluations reflecting smallholder conditions.

Economic loss analyses in California’s Central Valley have shown that irrigation intensity and water salinity jointly determine profitability. At modeled rates, 3 mm/day produced no profit, while 6 mm/day yielded up to USD 1000 ha⁻¹ for alfalfa and tomatoes; tree crops such as almonds and grapes required >8 mm/day to remain viable under salinity levels ranging from 0.5 to 5.5 dS/m [6]. These findings collectively demonstrate how salinity thresholds interact with crop type and water availability, which is critical for designing context-sensitive interventions. Likewise, spatial analyses in Inner Mongolia revealed that soil salt content (SSC) had a stronger influence on productivity than soil water content (SWC), with maize being more affected than sunflowers [7]. Such evidence underscores the need to disentangle the effects of salinity and moisture across crops and regions.

In Brazil and Mexico, poor irrigation and groundwater salinity have led to land abandonment, unemployment, and shrinking profit margins [8,9]. In Iran, annual salinity-related losses exceed USD 1 billion [3], whereas in Morocco, increased salinity demands more leaching water, straining limited resources [10]. Salinity emerges not only as a biophysical constraint but also as an economic stressor that undermines livelihood resilience.

Despite these insights, most global studies have focused on temperate regions or staple crops, often overlooking the compounded effects of salinity, moisture stress, and nutrient deficiency in tropical semi-arid zones. This gap is particularly pronounced in Ethiopia, where research has largely concentrated on the highland and midland areas, emphasizing teff, barley, wheat, and maize [11,12,13]. In contrast, salinity-prone lowlands and cash crops, such as bananas, cotton, and maize, remain underexplored [14,15,16]. Recent applications of machine learning and remote sensing have seldom addressed the economic consequences of soil stress in vulnerable systems, particularly those affecting smallholders in semi-arid sub-Saharan Africa [17,18,19,20].

In arid, semi-arid areas of Ethiopia, including the present study area (Figure 1a), rising salinity, nitrate depletion, and moisture stress have caused widespread land abandonment and yield losses of major crops [14,21,22,23]. Such multiple stresses directly threaten farmers’ income and regional food security [14,23,24]. Cotton, one of the most economically important crops, has experienced declines in yield, quality, and marketability due to salinity, moisture stress, and nutrient deficiency [25,26,27,28,29,30,31,32]. Additionally, productivity and income have declined owing to insufficient rainfall, land degradation, and poor irrigation practices [22,33,34,35,36]. Severe crop water stress further exacerbates yield loss [14,23,36]. Previous research indicates that the combined effects of salinity, low moisture, and nitrogen imbalance can worsen cotton losses by over 170%, although optimized management can mitigate these impacts [37,38,39].

Many previous studies still rely on classical models that inadequately capture the nonlinear interactions among salinity, moisture, and nutrients [40,41,42]. ML algorithms have improved yield prediction under different management conditions [17,18,19,43,44,45,46,47,48], and RS has enabled large-scale assessments of land cover, soil moisture and salinity [19,33]. Recent studies combining ML and RS have successfully estimated yields by integrating soil nitrate, salinity, and moisture content [20]; however, few have focused on semi-arid regions. In Ethiopia, some studies have applied ML to cash crops in semi-arid zones [14,15,16,49,50,51]; however, most still emphasize rainfall, temperature, and cultivated area, with limited attention to soil moisture stress, nutrient dynamics and salinity [12,13].

From the literature, three significant gaps persist in global and regional studies: (i) a limited focus on salinity-affected semi-arid regions, particularly in Ethiopia, where data to inform farmers are scarce; (ii) inadequately quantified productivity gaps in banana, maize, and cotton under the combined stresses of salinity, moisture deficiency, and nitrate scarcity; and (iii) a lack of economic evaluations for smallholders contending with these challenges in irrigated semi-arid systems.

This study addresses these limitations by integrating machine learning and remote sensing to (a) estimate crop yields under varying salinity, moisture, and nitrate conditions; (b) assess the economic impacts of banana, maize, and cotton cultivation among smallholders in southern Ethiopia; and (c) identify key predictors and algorithms for sustainable land management. By bridging empirical field data with advanced analytics, this study offers actionable insights for mitigating land degradation, enhancing productivity, and strengthening food security in semi-arid regions.

2. Materials and Methods

2.1. Study Site Description and Research Processes

This study was conducted in the Sille irrigation scheme, with geographical coordinates of 37°28′30″ to 37°30′50″ E and 5°50′0″ to 5°54′10″ N, at an altitude of 1118 m (Figure 1a) in the semi-arid region of southern Ethiopia. The scheme serves three villages (Eligo, Cahfe, and Mage) (Figure 1a). Furrow and flood irrigation systems are standard methods of crop production in these areas. The study area has low rainfall intensity, high evapotranspiration, and high soil moisture stress during the dry season. The temperature ranges from 17 to 33 °C, with annual precipitation of less than 850 mm, according to data from the Arba Minch Meteorological Station (AMS) under the National Meteorological Institute (NMI). The major irrigated crops in the area are cotton, maize, and banana, and their distribution across the irrigated lands was determined using satellite-derived indices (Figure 1b). The overall research sequence is shown in Figure 2 to provide a comprehensive overview of the study process.

2.2. In Situ Data Collection, Sampling Procedures, and Measurements

In situ data were collected from 10 October 2023 to 20 October 2023 to assess the spatial variations in soil salinity, nitrate, and moisture across the irrigated area. Soil samples for these parameters were collected across the field, and other agronomic parameters measurement procedures are briefly presented in Figure 3a–f. For sample collection, we applied a stratified random sampling approach, which enabled comprehensive coverage of both high- and low-salinity zones in the study area. Stratified random sampling is an effective method for use in heterogeneous environments characterized by variable soil features and crop cultivation practices. We applied standardized methods for soil sampling, crop yield measurements, and soil nitrate and soil moisture measurements. The crop yield measurements taken from various farmers’ field was used as replication (crop yield data taken from the field of 9 farmers in the vicinity of each sampling point). This method was particularly effective for the heterogeneous environment of the study area, which is characterized by variations in soil properties and agricultural practices. However, the soil sampling and crop yield measurements from the field were limited due to (i) the shortage of time to collect data across all the sampling points; (ii) the topographical and climatic features of the study area; (iii) limitations related to field data collectors; and (iv) financial shortage during field data collection. Therefore, the field data could not be used directly for model development. To overcome data scarcity and improve the prediction reliability, remote sensing data were incorporated to estimate crop yields and economic losses. Detailed sampling and measurement steps are summarized in Figure 3a–f.

(a) Soil moisture determination: To quantify soil-moisture variation within the irrigated fields, soil samples were collected from the top layer (0–15 cm), as shown in Figure 3. The gravimetric method [52] was used to determine the soil moisture content (SM) using Equation (1):

$$\mathrm{SM}={\displaystyle \frac{\left(FW-DW\right)}{DW}}\times 100\mathrm{\%}$$

(1)

where SM represents the soil moisture content (%), FW represents the fresh weight (gm), and DW represents the dry weight (gm).

(b) Soil salinity (EC): To assess salinity levels, 45 soil samples were collected across the irrigation scheme from a depth of 0–15 cm using a soil sampler (auger). Each sample was prepared by mixing distilled water and soil at a ratio of 10:1 (100 mL of water to 10 gm of soil) for extraction. The mixture was stirred, allowed to stand for 30 s, and filtered through filter paper to remove soil particles from the extract. Salinity parameters, including pH, EC, and TDS, were measured using a combo meter (from www.horiba.com, Horiba, Ltd., Okayama, Japan (accessed on 10 June 2023)). The EC probe (ds/m) was immersed in the filtered extract, and a stabilized reading was recorded. The TDS (meq/L) mode was subsequently activated, and the probe was placed in the same solution to obtain the TDS values in mg/L. The pH was measured by rinsing the pH probe, immersing it in the extract, and recording the stabilized value. All readings were repeated to ensure for accuracy, and all data—including EC, TDS, pH, sampling location, and date—were documented for subsequent analysis.

(c) Crop yield: Crop yield (kg/m²) was measured for three major crops across the irrigated areas, with plot sizes of 5 m × 5 m for cotton and maize and 9 m × 9 m for banana. Data obtained from the farmers’ field were used as replicates. The net harvested yield data per plot were converted to a hectare basis for each crop type. Purposive sampling was applied to ensure representative coverage of all irrigated zones. Banana data were collected from 35 farmers’ fields, maize from 19, and cotton from 25. For each crop, measurements from different portions of the irrigated area were averaged to obtain representative values.

(d) Soil nitrate (NO₃⁻): To quantify the nitrate status of the soil, 45 soil samples were taken from the field from a depth of 0–15 cm, cleaned, and then ground prior to analysis. The soil nitrate content was analyzed using a LAQUAtwin NO₃⁻ meter (www.horiba.com, Horiba, Okayama, Japan (accessed on 10 June 2023)). The soil extract was prepared following standard laboratory procedures. The LAQUAtwin meter was calibrated, and a solution was prepared at a ratio of 10:1 (water to soil), consistent with the procedures used for other parameters. To avoid contamination of the solution, the electrode was rinsed between measurements, and each measurement was repeated three times. To minimize variability and ensure field data accuracy, repeated measurements of the soil samples for soil moisture, soil nitrate, and yield were taken. The soil nitrate concentration (NC, mg/L) of the soil samples was determined using Equation (2) [53]:

$${{\mathrm{NO}}_3}^{-}={\displaystyle \frac{Nitrateconcetration\left(\frac{\mathrm{m}\mathrm{g}}{\mathrm{L}}\right)\times VolumeofExtraction(\mathrm{L})}{Weightofsoilsample\left(\mathrm{k}\mathrm{g}\right)}}$$

(2)

2.3. Statistical Data Analysis

After completing field data collection, sample analysis, and parameter determination, the variability and deviations of the field data and remote sensing data across the sampling points were assessed using analysis of variance (ANOVA). Descriptive statistics, including standard deviation and maximum and minimum ranges of each parameter, are presented in

Table 1. Descriptive statistics of soil parameters and crop yields.

Parameters	Mean ± SD	Minimum	Median	Maximum
YC	0.92 ± 0.44	0.14	0.86	2.05
YM	4.50 ± 0.79	2.53	4.75	6.17
YB	6.81 ± 0.84	4.85	7.10	8.71
SM	0.25 ± 0.09	0.08	0.25	0.51
N	3.69 ± 2.60	1.05	2.56	21.67
EC	18.31 ± 4.17	7.63	18.87	25.61

Note: YC, yield of cotton (ton/ha); YM, yield of maize (ton/ha); YB, yield of banana (ton/ha); SM, soil moisture (%); N, soil nitrate (mg/L); EC, soil salinity (dS/cm).

under Section 3.1. Spatial distribution maps of the three crops were generated using ArcGIS Pro (version 3.5, ESRI, Redlands, CA, USA). Satellite images were accessed and preprocessed through Google Earth Engine integrated with Google Colab. Machine learning model development, training, testing, and evaluation were conducted in Python 3.12 within the Google Colab environment.

2.4. Satellite Data Acquisition, Integration, and Index Determination

To derive satellite-based indices for salinity, moisture, and nitrate in Equations (3)–(5), three satellite image sources—landsat8 (https://landsat.gsfc.nasa.gov/ accessed on 15 February 2024), MODIS (https://www.earthdata.nasa.gov/ accessed on 15 February 2024), and Sentinel-2 (https://dataspace.copernicus.eu/ accessed on 15 February 2024) were utilized. These satellite sources have better resolutions for the selected indices and are freely accessible [53]. Landsat 8 has a resolution of 30 m with a 16-day frequency of data availability and provides data for salinity indices.

MODIS, with a resolution of 250 m to 1 km, provides data on soil moisture and nitrate indices for vegetation and soils based on daily satellite data. Sentinel-2 images with resolutions ranging from 10 to 60 m for the visible and near-infrared bands were employed to evaluate soil-moisture dynamics within the study area.

Satellite images were collected between 10 October 2023 and 30 January 2024, corresponding main cropping season in the study area when cloud free data are most available. Image preprocessing and index computation were performed in Google Earth Engine (GEE) integrated with the Google Colab, a cloud-based platform widely used for large-scale remote-sensing analysis [53,54]. A RF classifier implemented in GEE [55] was trained and tested using point-based field observations to ensure index accuracy.

• The normalized difference vegetation index (NDVI) was calculated to identify crop health and its areal coverage in the study area using Equation (3):

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

  (3)

  where NIR and RED represent the near-infrared band and red band, respectively.
• The normalized differential salinity index (NDSI) was used to determine the salinity distribution in the irrigated lands, calculated according to Equation (4) [56,57].

  $$\mathrm{NDSI}={\displaystyle \frac{R-NIR}{R+NIR}}$$

  (4)

  where R denotes the red band and NIR is the near-infrared band.
• The normalized difference moisture index (NDMI) was determined to assess the vegetation soil moisture conditions across the study area [58] and is calculated using Equation (5):

  $$\mathrm{NDMI}={\displaystyle \frac{NIR-MIR}{NIR+MIR}}$$

  (5)

  where NIR, RED, and MIR are the near-infrared, red, and mid-infrared bands, respectively. The NDMI value ranges from −1 to 1, and a value close to 0 indicates soil moisture stress.
• The soil nitrate levels of the irrigated lands were determined by developing the linear regression model of the vegetation index (NDVI) using the field-observed data according to previously described methods [53,59].

2.5. Relative Yield and Deviation Determination

The relative yield performance of each crop under salinity, moisture, and nitrate stress conditions was quantified following the approach described in [60]. The combined effects of these stressors were evaluated using Equation (6):

$${Yr}_i=100-\left(B_i\times (EC-a)-C_i\times (N-N_{threshold})-D_i\times (SM-{SM}_{threshold})\right)$$

(6)

where Yr_i is the crop yield relative to the potential (under no salinity); a is the crop salinity threshold, in dS/m; b_i is the slope expressed in percentage per dS/m; and EC is the predicted (or measured) salinity level (dS/m) of the soil. The salinity threshold value (ds/m) and loss reduction slope (%) of each crop were obtained from different studies, with the salinity tolerance threshold (a) being 7.7, 1.7, and 3.5 ds/m, respectively, and the slope (B) being 5.2%, 12%, and 25%, respectively, for cotton, maize, and banana [60]. C_i and D_i represent the slopes of yield reduction associated with nitrate (% per kg/ha) and soil moisture deficit (% per mm/m), respectively. N is the soil nitrate observed (mg/kg), N_threshold is the optimal N rate for each crop (the amount of N fertilizer in kg/ha), SM is the soil moisture obtained (mm/m), and SM_threshold is the optimal soil moisture for each crop (mm/m).

The nitrate and moisture thresholds were adopted from literature values for each crop in the region, while the corresponding yield performance slopes were empirically derived from the field observations.

Following the estimation of the relative yield reduction, we determined the farmer-based deviation ($\mathsf{\Delta }{y}_i)$ of crop yield and compared it with the potential yield ($ai)$. This is the deviation between the relative yield (${Yr}_i)$ from Equation (5) and the potential yield ($a_i)$ determined using Equation (7) for each crop.

$$\mathsf{\Delta }{y}_i={Yr}_i-a_i$$

(7)

After estimating the farmer-based yield deviations, the economic loss of each crop across the sampled farmland (FLEL_i) was evaluated using Equation (8).

$$FLE{L}_i=\mathsf{\Delta }{y}_i\times P_i$$

(8)

where Δy_i denotes the land-based economic deviation for each crop (in USD), and P_i represents the corresponding farm-gate price (USD/kg), which incorporates fertilizer, labor, and irrigation costs. The FLEL was computed for banana, cotton, and maize, with average unit prices (P_i) of 0.8, 0.66, and 0.6 USD/kg, respectively. These prices reflect the farmers’ gate values, which vary seasonally and annually.

To obtain a model with better prediction and ensure its reproducibility, the consistency of the field and satellite-accessed data was first evaluated as follows:

(a)

Model input and parameter set: Key explanatory variables included soil moisture, soil nitrate, and soil salinity, which were used to assess their combined effects on crop yields and farmers’ economic returns. The key soil properties included soil salinity (ds/cm), soil nitrate (mg/kg), and soil moisture (%). Remote sensing indices, including the normalized difference vegetation index (NDVI), normalized difference salinity index (NDSI), and normalized difference soil moisture index (NDSMI), were incorporated into the field-observed data. The field-based crop yields of the three major crops (banana, cotton, and maize, in kg/ha) were also included in the model. Production cost and economic gain were estimated from the current market prices of each crop.

Default hyperparameter settings were applied to each machine-learning model to enhance reproducibility and ensure an unbiased comparison of model performance. The models included a Ridge Regression model (RR) (α = 1.0), a Decision Tree Regressor (DT) (random state = 42), and a RF Regressor (n_estimator = 100 and random state = 42). We also implemented a Gradient Boosting Regressor (GB) (n_estimator = 100 and random state = 42), a Support Vector Regressor (SV) using a radial basis function (RBF) kernel, and a K-Nearest Neighbors Regressor (kNN) based on the average nearest neighbors. Finally, we utilized a Multi-Layer Perceptron (MLP) Regressor with maximum of iteration = 1000 and random state = 42. This selection of models was designed to provide a comprehensive evaluation of the regression performance across various methodologies.

The model assessment was based on a comparative analysis of the selected algorithms thereby avoiding bias from the optimized hyperparameters. Hyperparameter optimization can enhance accuracy, and the current study provides a straightforward comparison and highlights the strengths and weaknesses of the models under consistent conditions. The selected machine learning models were trained using 80% of the data and tested using 20% from all the crop datasets.

(b)

Data integration: The field-observed soil moisture, nitrate, and soil salinity data were normalized using the min-max scaling approach to ensure consistency with the satellite-derived indices. The soil moisture and nitrate data were scaled according to their respective magnitudes. The relationship between the NDIVI of each crop and the soil nitrate concentration was quantified using linear regression across the sampling points [4,59]. Similarly, soil moisture content across the sampling points was estimated based on the correlation between NDSMI and field-observed soil moisture. By correlating and scaling the field-observed and satellite datasets, the crop yields, soil nitrate, soil moisture, and soil salinity of banana, maize, and cotton (N = 189, N = 128, and N = 176, respectively) were determined.

(c)

Machine learning models: The algorithms were selected based on their capacity to capture nonlinear relationships among factors such as yield and economic gains, computational efficiency, and robustness [61]:

(i)

The Ridge regressor (RR) is a suitable approach for addressing multicollinearity in regression models. Under multicollinearity, the least squares estimates are unbiased but exhibit large variance, which can lead to unstable prediction [62]. RR eliminates standard errors by introducing a level of bias into the regression estimates [62].

(ii)

The Random Forest Regressor(RF) is a decision tree-based ensemble method used for classification and regression [63]. Each tree is trained on a random subset of both features and samples. Compared with linear regressors, it may still face challenges such as overfitting and relatively longer prediction times.

(iii)

The Support vector Regressor(SV) is a robust and rapid machine learning approach [44]. It has high flexibility in capturing nonlinear relationships [61].

(iv)

The Decision Tree Regressor(DT) is a supervised learning algorithm that infers prediction rules from input features to approximate the target variable [62]. It partitions complex datasets into smaller, homogeneous subsets for improved interpretability and accuracy [64].

(v)

The Multilayer perceprton regressor (MLP) can solve difficult nonlinear problems in a feedforward manner. It can handle massive volumes of data inputs. The power of MLP networks lies in their ability to fit numerous smooth, nonlinear functions with excellent precision.

(vi)

The Gradient Boosting (GB) Regressor builds the model iteratively using a boosting framework, where each stage corrects the residuals of the previous one and optimizes a differentiable loss function [65,66]. The final predictive model is an ensemble of weak learners combined to enhance accuracy and robustness [62,64].

(vii)

The K-Nearest Neighbors (KNN) Regressor is a nonparametric approach that estimates the target values by averaging the outcomes of neighboring samples with similar feature characteristics [65,67,68]. The optimal neighborhood size is determined through cross-validation to minimize the mean-squared error (MSE) [65,67,69].

(d)

Feature importance of models: Feature importance was quantified using the mean SHAP values. The parameters with higher mean SHAP values were identified the most important features, whereas those with a lower mean value were considered less important features in the model prediction [70,71].

(e)

Model performance evaluation: The predictive performance of each model was evaluated using metrics: mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), explained variance score (EVS), and mean absolute percentage error (MAPE). MAE represents the average of the absolute differences between observed and projected values across the dataset [72], as defined in Equation (9). RMSE measures the standard deviation of the residuals [73], as given in Equation (10). R² represents how much of the variability in the dependent variable can be accounted for by the independent variables, as indicated in Equation (11). EVS indicates how well the model predictions account for the variability observed in the dependent variable, as shown in Equation (12). MAPE measures the model accuracy as a percentage, as presented in Equation (13).

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{s}}}{\sum }_{\mathrm{i}=1}^{\mathrm{s}}|{\mathrm{Y}}_{\mathrm{a}\mathrm{c}\mathrm{t},\mathrm{i}}-{\mathrm{Y}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d},\mathrm{i}}|$$

(9)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{s}}}{\sum }_{\mathrm{i}=1}^{\mathrm{s}}{\left({\mathrm{Y}}_{\mathrm{a}\mathrm{c}\mathrm{t},\mathrm{i}}-{\mathrm{Y}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d},\mathrm{i}}\right)}^2}$$

(10)

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum {\left(Y_{act,i}-Y_{pred,i}\right)}^2}{\sum {\left(Y_{act,i}-Y_{avr}\right)}^2}}$$

(11)

$$\mathrm{EVS}=1-{\displaystyle \frac{Var\left(Y_{act}-Y_{pred}\right)}{Var\left(Y_{act}\right)}}$$

(12)

$$\mathrm{MAPE}={\displaystyle \frac{1}{\mathrm{s}}}{\sum }_{\mathrm{i}=1}^{\mathrm{s}}|{\displaystyle \frac{{\mathrm{Y}}_{\mathrm{a}\mathrm{c}\mathrm{t},\mathrm{i}}-{\mathrm{Y}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d},\mathrm{i}}}{{\mathrm{Y}}_{\mathrm{a}\mathrm{c}\mathrm{t},\mathrm{i}}}}|\times 100\%$$

(13)

where Y_act is the actual value of the variables, Y_pred is the predicted value of the variables, and Y_avr is the grand average of the variables.

3. Results

3.1. Summary of Statistical Analysis of Crop Yield and Other Variables

The descriptive statistics in Table 1 indicate substantial variability in crop yields and soil conditions across the study area. Cotton yields ranged from 0.14 to 2.05 tons per hectare, maize from 2.53 to 6.17 tons per hectare, and banana from 4.85 to 8.71 tons per hectare, reflecting pronounced environmental heterogeneity. These yield ranges suggest that crop productivity is highly sensitive to field conditions. Soil salinity, nitrate, and moisture levels also exhibited considerable spatial variations. Soil moisture ranged from 0.08 to 0.51, indicating fluctuations in water availability across sites. The mean soil nitrate concentration was 3.69 ± 2.60 mg/L, pointing to an uneven nutrient distribution within the irrigation area. Collectively, these variations underscore the spatial complexity of soil conditions that affect crop productivity.

The 3D plots (Figure 4a–c) illustrate the relationship between crop yield and major soil variables. A clear negative trend existed between salinity and yield across all crops, confirming the suppressive effect of salinity on productivity. Conversely, yield tended to increase with higher soil moisture and nitrate concentrations up to optimal thresholds, indicating that while water and nutrients enhance yield, salinity remains the dominant limiting factor.

3.2. Yield Prediction

Model predictions revealed clear variations in cotton, maize, and banana yields across farmers’ fields (

Table 2. Model performance metrics for cotton yield and economic gain.

Crop	Parameter	Model	RMSE	R2	MAE	EVS	MAPE
Cotton	Yield	RR	0.000	1.000	0.002	0.999	0.306
		DT	0.000	0.999	0.007	0.990	0.865
		RF	0.000	0.998	0.008	0.999	0.767
		GB	0.000	1.000	0.004	0.997	0.386
		SV	0.005	0.976	0.048	0.977	6.252
		KNN	0.005	0.974	0.052	0.977	6.204
		MLP	0.003	0.985	0.043	0.985	5.896
	Economic benefits	RR	1.888	0.999	1.336	0.999	1.074
		DT	8.031	0.990	3.641	0.999	5.201
		RF	10.843	0.998	4.835	0.999	16.316
		GB	4.979	0.997	1.974	1.000	4.788
		SV	249.181	0.259	192.820	0.166	254.090
		KNN	41.736	0.974	29.422	0.977	41.968
		MLP	215.410	0.311	178.980	0.629	103.652

Note: RR, Ridge Regression; DT, Decision Tree; RF, Random Forest; GB, Gradient Boosting; SV, Support Vector Regression; KNN, K-Nearest Neighbors; MLP, Multi-Layer Perceptron; RMSE, root mean square error; R², coefficient of determination; MAE, mean absolute error; EVS, explained variance score; MAPE, mean absolute percentage error.

,

Table 3. Model performance metrics for maize yield and economic gain.

Crop	Parameter	Model	RMSE	R2	MAE	EVS	MAPE
Maize	Yield	RR	0.479	0.582	0.334	0.587	7.395
		DT	0.418	0.679	0.265	0.695	5.435
		RF	0.344	0.793	0.230	0.223	4.556
		GB	0.346	0.781	0.222	0.791	4.356
		SV	0.413	0.689	0.260	0.708	5.426
		KNN	0.335	0.795	0.258	0.798	5.540
		MLP	0.448	0.633	0.313	0.635	6.881
	Economic benefits	RR	285.27	0.582	199.431	0.587	33.813
		DT	783.080	0.685	159.170	0.697	59.377
		RF	208.414	0.777	136.691	0.790	50.804
		GB	206.576	0.780	132.280	0.791	51.350
		SV	412.41	0.125	347.230	0.126	84.764
		KNN	199.705	0.799	153.768	0.798	38.495
		MLP	360.481	0.332	300.177	0.501	69.028

Note: RR, Ridge Regression; DT, Decision Tree; RF, Random Forest; GB, Gradient Boosting; SV, Support Vector Regression; KNN, K-Nearest Neighbors; MLP, Multi-Layer Perceptron; RMSE, root mean square error; R², coefficient of determination; MAE, mean absolute error; EVS, explained variance score; MAPE, mean absolute percentage error.

and

Table 4. Model performance metrics for banana yield and economic gain.

Crop	Parameter	Model	RMSE	R2	MAE	EVS	MAPE
Banana	Yield	RR	0.436	0.687	0.308	0.696	4.502
		DT	0.288	0.863	0.206	0.863	2.977
		RF	0.217	0.922	0.176	0.923	2.586
		GB	0.246	0.901	0.187	0.901	2.736
		SV	0.277	0.875	0.214	0.883	3.107
		KNN	0.268	0.881	0.207	0.884	3.027
		MLP	0.361	0.784	0.257	0.787	3.764
	Economic benefits	RR	275.920	0.687	195.329	0.696	11.278
		DT	178.720	0.868	125.129	0.869	7.385
		RF	138.850	0.921	111.990	0.921	6.426
		GB	155.608	0.901	118.488	0.901	6.842
		SV	457.460	0.140	396.834	0.140	22.720
		KNN	170.164	0.881	130.788	0.884	7.483
		MLP	1207.720	0.099	1132.815	0.082	25.735

Note: RR, Ridge Regression; DT, Decision Tree; RF, Random Forest; GB, Gradient Boosting; SV, Support Vector Regression; KNN, K-Nearest Neighbors; MLP, Multi-Layer Perceptron; RMSE, root mean square error; R², coefficient of determination; MAE, mean absolute error; EVS, explained variance score; MAPE, mean absolute percentage error.

; Figure 5a–c), highlighting the spatial heterogeneity of soil and environmental conditions. For cotton, RR achieved the highest accuracy (R² = 0.999, RMSE ≈ 0), followed by RF (R² = 0.998, RMSE = 0.004). These results indicate that regularized linear models and ensemble learners effectively capture cotton yield dynamics.

The MLP exhibited intermediate performance, whereas the SV and KNN performed relatively poorly. For maize, GB exhibited the best accuracy (R² = 0.781, RMSE = 0.34), followed by RF. DT and RR were less effective, and KNN produced weakest results, suggesting models capturing nonlinear interactions better predict maize yields. For banana, the DT performed best (R² = 0.863, RMSE = 0.206, MAPE = 2.977), indicating a strong alignment with field-level variability. MLP and SV showed the lowest accuracy, suggesting that tree-based models handle banana yield estimation better. The RF and RR were optimal for cotton, GB for maize, and DT for banana, reflecting crop-specific environmental sensitivities. However, yield predictions alone are insufficient; economic evaluation are essential to translate insights into strategies for smallholders.

3.3. Prediction of Farmers’ Economic Losses

Model predictions indicated that salinity, soil moisture, and nitrate influenced farmers’ economic losses (Table 2, Table 3 and Table 4; Figure 6a–c), with varying accuracy across crops and models.

For cotton, RF, RR, and GB models achieved the highest accuracy, demonstrating reliable performance under combined salinity and nutrient stress. In contrast, the SV, KNN, and MLP models exhibited lower accuracy, suggesting that these algorithms are less suited for economic prediction in heterogeneous field conditions.

For maize, the GB and RF models outperformed the others in estimating losses from salinity-induced water stress and nutrient deficiencies. The weaker performance of SV and MLP reinforces the importance of using models that can effectively capture nonlinear interactions among soil and climatic factors.

For banana, RR showed moderate accuracy (R² = 0.687), although high error values limited reliability. RF and DT models performed better, while GB achieved the highest accuracy (R² = 0.901, RMSE = 155.608), suggesting strength in capturing complex stressor effects.

Across all crops, SV, KNN, and MLP consistently underperformed. These results indicate that robust models tend to perform more reliably for economic predictions under combined environmental and management-related stressors.

3.4. Feature Importance for Yield and Economic Loss Prediction

SHAP analysis identified soil salinity, moisture, and nitrate as key predictors of crop yield and economic loss (Figure 7a–c and Figure 8a–c).

Salinity exerted the strongest influence on cotton yield (SHAP = 0.378), whereas soil moisture (0.259) and nitrate (0.134) contributed less substantially. For maize, salinity ranked highest (0.769), followed by nitrate (0.259) and moisture (0.205), indicating the secondary role of nutrient stress. For banana, salinity (0.575) was the key stressor, whereas nitrate (0.3134) and moisture (0.123) were less influential. Collectively, these patterns demonstrate that salinity suppresses yield across crops, reinforcing the need for targeted management strategies.

In predicting economic losses, salinity remained the most influential, with mean SHAP values of 213.4 (cotton), 291.4 (maize), and 277.8 (banana). Nitrate contributed significantly to cotton (113.5) and maize (81.7), whereas moisture had a greater influence on banana. The results indicate that economic vulnerability varies among crops and stressors, with salinity acting as primary driver of losses and moisture playing a crop-specific role.

SHAP analysis revealed that increased salinity was associated with reduced soil moisture, indicating a compounding stress effect. Soil nitrate influences yield and economic losses across crops, highlighting its role in productivity. These interactions suggest that salinity, moisture, and nutrient management are interrelated factors influencing agronomic and economic outcomes in semi-arid regions.

4. Discussion

4.1. Yield Productivity and Models’ Predictive Capacity

Soil salinity has severely affected crop yields in the region examined in this study by reducing soil water and nutrient availability [74,75]. For the selected crops, the yield performance was well below regional and national averages; for instance, banana yield was about 8.81 tons/ha, cotton about 0.92 tons/ha, and maize about 4.5 tons/ha. In most sites, soil salinity level exceeded 16 dS/cm, which was far above the crop-specific tolerance thresholds [10,76]. Poor agricultural water management has further aggravated salinity, thereby reducing soil moisture and nutrient availability and lowering yields [23,24].

Among the selected machine learning models, RF displayed the best performance for yield and economic loss estimation in areas affected by soil salinity. Consistent with our findings, RF has shown strong predictive capacity for maize yield in the semi-arid regions of sub-Saharan Africa [77] and wheat yield under varying conditions globally [78]. RF also performs well with spatially distributed data in complex and interrelated conditions, particularly in modeling the effects of organic fertilizers and seed rates [79]. The superior performance of RF relative to other models is likely due to its robustness in managing high-dimensional datasets and its capability to model complex nonlinear relationships [80].

In contrast, GB has shown better prediction of banana yield in some studies [81,82], and XGBoost has been found to outperform RF for maize [83]. Another study reported that RF predicted banana yield with a 70% accuracy [80], suggesting room for improvement. These results indicate that while RF is generally robust, model performance can vary depending on the crop type and data characteristics.

A limitation of the present study is that only salinity, soil moisture, and nitrate measurements were considered. In addition, a relatively limited dataset without seasonality or large spatial coverage was analyzed. Recent studies have highlighted the potential of deep learning approaches, including neural networks and convolutional neural networks (CNNs), to further improve yield estimation in regions with complex conditions [84,85,86].

Despite these limitations, our combined field- and satellite-based analysis provides valuable insights into how soil salinity, moisture, and nitrate affect crop yields in semi-arid regions. These results can serve as a useful benchmark for areas lacking field-based estimates and assist decision-makers in guiding farmers toward more effective salinity and risk management practices.

4.2. Insights on Economic Losses Based on Model Estimations

Soil salinity has significantly reduced farmers’ economic returns in the region examined in this study by decreasing soil moisture and nitrate concentrations. These combined stresses have caused substantial yield and economic losses [87,88]. In semi-arid regions, increasing temperature and rainfall variability further exacerbate crop yields and economic benefits [89,90]. In our study area, yield losses reached up to 100%, resulting pronounced economic disparities among farmers. Optimized nutrient application, soil moisture management, and salinity mitigation have been shown to restore crop yield and economic benefits [10,75,76,89,90,91,92].

Similar patterns have been observed in other regions. In one study in California, salinity combined with water scarcity reduced crop yields by up to 90% [77], while another study found that a unit increase in salinity caused yield losses of approximately 5.38% and revenue reductions up to 364.90 [93]. These findings confirm that salinity directly contributes to both yield and income instability.

From the field observations, we conclude that increasing soil salinity is directly associated with soil moisture stress and nutrient uptake efficiency, which together lead to quantifiable yield reductions and economic losses. Banana and maize farmlands in the study area face severe salinity problems with reduced soil fertility, leading to substantial yield declines. High salinity also deteriorates soil structure, reduces aeration, and limits nutrient uptake in the root zone.

Similarly, to yield estimation, RF showed strong performance in estimating the economic losses of the three studied crops in the region, although GB Regressor sometimes outperformed RF. By contrast, MLP and SV performed poorly, reflecting their limited capability in capturing complex relationships in heterogeneous environments.

Overall, the economic loss estimation models provide valuable insights into minimizing yield and income reductions in the region. Three key implications can be drawn. First, real-time soil and water monitoring data should be integrated into management strategies to reduce economic losses. Second, RF and other advanced models can guide salinity management to improve crop yield and economic returns. Third, more advanced data-driven modeling, including hybrid approaches, should be applied in the future for more effective management of salinity to improve crop yields and agricultural sustainability in semi-arid regions.

4.3. Prioritization of Determinants for Remedial Strategies

The mean SHAP values showed that soil salinity significantly influenced crop yield and economic losses, followed by soil moisture and soil nitrate, as shown in Figure 5a–c and Figure 6a–c. This implies that increasing salinity primarily drives yield reduction, highlighting the urgent need for salinity management to sustain the productivity of marginalized lands and ensure food security.

Increasing salinity further causes soil moisture stress and nutrient deficiency, thereby accelerating soil infertility. Across the study sites, salinity also caused surface cracking, nutrient depletion, aeration problems, and fertilizer leaching. These interrelated challenges intensified soil degradation and food insecurity, putting farmers at greater risk. In addition, soil fertility declines and nitrogen fixation problems further contributed to yield variability among farmers.

These findings are consistent with other model-based results from different regions. For example, the SALTMED study identified soil salinity as the main determinant of faba bean and sunflower production [94,95]. Other studies have also highlighted that in regions affected by soil moisture stress, developing optimization models for yield and water allocation can enhance crop productivity and improve farmers’ resilience [96,97,98]. Similarly, a study showed that salinity significantly contributed to cotton yield and economic gains [99].

Considering temperature together with soil moisture improved yield prediction, Pandyawith a mean SHAP value of 0.47 [100]. Another study confirmed that atmospheric pressure and temperature are key contributors to maize yield in complex environments, similar to the findings for the study area [101]. These results indicate that addressing salinity, together with moisture and nitrate depletion, is critical for reducing economic losses, while also recognizing that climatic factors amplify stress conditions.

Taken together, the SHAP values imply that prioritizing salinity management could improve yields and reduce economic losses in the region. For sustainable production, integrated strategies that also account for climatic variability should be developed. Such strategies will support farmers, decision-makers, and policymakers in drafting and implementing effective measures to ensure sustainable land use and enhanced crop productivity.

4.4. Limitations and Future Interventions

This study focused on three key environmental stressors—soil salinity, moisture, and nitrate—but did not account for other influential factors, such as soil carbon, phosphate, and organic matter, which also affect crop yield and economic outcomes. The analysis relied on seasonal and spatial data from farmers’ fields, providing valuable insights but limiting the temporal depth and broader generalizability. These gaps constrain the scalability and policy relevance of this study’s findings.

To address these limitations and strengthen future interventions, research should prioritize the following:

• Enhancing salinity, moisture, and nitrate conditions through cover crops (e.g., legumes, salt-tolerant species) and organic amendments (e.g., compost, green manure) is supported by long-term trials and model-based optimization.
• Developing and applying crop-specific salt-leaching fractions to reduce salinity buildup in irrigated lands.
• Adapting nitrate fertilizer application to site-specific nutrient balances can improve productivity and profitability.
• Engaging farmers, water managers, and policymakers in co-developing strategies for integrated soil and water management.
• Incorporating data-driven models (e.g., RF, ensemble learners) that integrate climatic, agronomic, and socioeconomic variables for more robust prediction and decision support. These directions will help build resilient, evidence-based strategies for sustaining crop productivity and farmer livelihoods in salinity-affected semi-arid regions

5. Conclusions

In Ethiopia’s semi-arid regions, where field-based data are scarce, integrating satellite observations with machine learning offers valuable insights for farmers and decision-makers. In this study, seven ML algorithms were applied to predict crop yields and economic losses using soil salinity, moisture, and nitrate data derived from field and remote sensing sources.

Among the models, RF consistently outperformed the others, achieving high yield prediction accuracy for cotton (R² = 0.998), banana (R² = 0.922), and maize (R² = 0.793). RF also proved effective in estimating economic losses, reflecting its strength in capturing complex environmental interactions and quantifying the financial impact of salinity, moisture stress, and nitrate depletion.

Feature importance analysis confirmed that soil salinity was the most influential factor, followed by nitrate and moisture. This underscores the need to prioritize salinity mitigation, alongside water and nutrient management, to sustain productivity and reduce economic vulnerability.

These findings provide actionable guidance for improving the resilience of banana, maize, and cotton systems in semi-arid, irrigated regions. They also support efforts to enhance farmer income and reduce land marginalization.

Future research should integrate climatic variables, long-term monitoring, farmer participation, and broader agroecological and economic dimensions to better address the complexities of sustainable agriculture under salinity stress.