Tiaty is a region in Baringo County, which is part of the Rift Valley Province of Kenya (refer to Figure 1 ). It is situated on the floor of the rift valley, with an average altitude of 1000 m above sea level. It lies between latitudes 00˚13' S and 1˚40' N and longitudes 35˚36' E and 36˚30' E.

The climate is classified as semiarid, with temperate to warm temperatures, and is prone to recurrence. The region experiences a semiarid climate characterized by moderate to warm temperatures and a high likelihood of periodic droughts ( Greiner et al., 2021a ). This climate classification is reflected in the dominant vegetation of the acacia bush savannah, which is well adapted to thrive in arid conditions ( Greiner et al., 2021b ; Vehrs, 2016 ).

The rugged topography, consisting of lowland plains, rolling hills, and mountain ranges, further influences local climate patterns ( Greiner et al., 2021b ). These physical features can create variations in temperature, precipitation, and wind patterns across a region. The average temperature ranges from 25˚C to 30˚C, with occasional highs exceeding 35˚C and a mean annual rainfall ranging from 450 - 900 mm ( Kapoi & Charles, 2015 ). The average temperature ranges from 25˚C to 30˚C, with occasional highs exceeding 35˚C, and a mean annual rainfall ranging from 450 - 900 mm ( Kapoi & Charles, 2015 ). The dominant vegetation is open Acacia/Combretum wooded grassland, supporting primarily livestock production through specialized pastoralism for the past two centuries ( Bollig & Österle, 2013 ).

The research utilized Landsat-9 Operational Land Imager (OLI) and Thermal Infrared (TIRS) data with a 30 m resolution ( Earth Resources Observation and Science (EROS) Center, 2020 ). The data, encompassing all spectral bands, were acquired during the agronomic period of January to April 2023. The potential impact of seasonal variations on spectral signatures will be considered in the analysis. Additional auxiliary boundary GIS data for Tiaty in Baringo County were acquired from RCMRD, Kenya, to ensure accurate spatial referencing and classification within the study area boundaries.

The Landsat-9 images were processed via Google Earth Engine applications, where the cloud mask method removed pixels affected by clouds. To ensure that classifications were performed within the appropriate limits, the images were clipped to the boundaries of the study area. Spectral indices, as outlined in the workflow in Figure 2 , with detailed formulas provided in Table 1 , were employed for classification purposes.

Cloud masking in the Google Earth Engine (GEE) was employed to identify and eliminate pixels contaminated by clouds in satellite imagery, thereby enhancing the quality and reliability of the analyses. GEE’s built-in cloud masking function was used, employing bitwise operations to identify and mask cloudy pixels. The clip function (.clip ()) in GEE was implemented to confine the satellite imagery dataset to the area of interest (AOI) corresponding to the user-defined geometry

of Tiaty, the study area. This focused the analysis on the boundaries of the study area, resulting in reduced computational requirements. The effective integration of cloud masking and clipping is crucial for generating high-quality data for analysis.

The normalized difference vegetation index (NDVI) was used to differentiate between dense shrubland and dryland forests. The normalized difference water index (NDWI) was used to distinguish between water bodies and wetlands, whereas the bare soil index (BSI) was used to identify bare surfaces. The normalized difference built-up index (NDBI) was also applied to identify urban built-up areas. These indices were selected based on their established effectiveness in differentiating between various land cover types in arid and semi-arid environments.

The overall accuracy is the number of correctly classified pixels divided by the total number of validation pixels, as shown in Equation (1) ( Elmahdy & Mohamed, 2023 ; Pham et al., 2023 ).

(1)

where:

$n_{ii}$ is the number of correctly classified pixels for class i.

$N$ is the total number of validation pixels.

The kappa coefficient is a measure of the agreement between the classification map and the reference data, considering the agreement that might be expected by chance, as shown in Equation (2):

$\text{Kappa}=\left(p_{\text{o}}-p_{\text{e}}\right)/\left(1-p_{\text{e}}\right)$ (2)

where $p_{\text{o}}$ is the probability of observed agreement and where $p_{\text{e}}$ is the probability of chance agreement.

i) F-1 score

The harmonic means of precision and recall, the F-1 score, are used to evaluate machine learning and information retrieval. An imbalance between positive and negative categories makes model performance evaluation more balanced ( Pham et al., 2023 ). The F-1 score is the harmonic mean of precision and recall, or the ratio of genuine positives to true and false positives, which can be evaluated via Equations (3)-(5):

The F1 score is shown in Equation (3):

$F1=2\cdot \frac{\text{Precision}\cdot \text{Recall}}{\text{Precision}+\text{Recall}}$ (3)

where:

$\text{Precision}=\frac{TP}{TP+FP}$ (4)

$\text{Recall}=\frac{TP}{TP+FN}$ (5)

where TP = true positives, FP = false positives and FN = false negatives.

ii) Weighted F-1 score

Although F-1 ratings are significant, the weighted F1 score serves to assess the model’s performance across all categories ( Hinojosa Lee et al., 2024 ). Weights tailored to problem specifications provide a more focused and refined assessment of model performance ( Pham et al., 2023 ). The weights can be modified according to the specific demands of the task, facilitating a more precise and nuanced assessment of the model’s performance ( Lipton et al., 2014 ; Mayer et al., 2021 ; Pham et al., 2023 ). In these instances, the weighted F1 score more effectively evaluates the model’s performance by incorporating the priority of precision and recall ( Lipton et al., 2014 ). The weighted F-1 can be evaluated via Equation (6).

$\begin{array}{c}\text{Weighted F1-score}=\underset{i=1}{\overset{N}{{\displaystyle \sum }}}\left(\frac{{\text{Support}}_i}{\text{Total Support}}\right)\times {\text{F1-score}}_i\end{array}$ (6)

where:

N = Total number of classes, Support i = Number of true instances for the i-th class. Total Support = Sum of all support values. F1-score_i = F1 score for the i-th class.

iii) Execution run-time

The execution duration of machine learning algorithms in the Google Earth Engine (GEE) is directly affected by several pivotal factors, including dataset size, algorithm complexity, available bandwidth for data transfer, and memory utilization, all of which are crucial in ascertaining the computational time necessary for performing machine learning tasks on GEE ( Zhao et al., 2024 ). Understanding and improving these components will result in more effective and successful outputs and execution times ( Farda, 2017 ).

Pixel-based supervised classification in the Google Earth Engine utilizes several classifiers within the Google Earth Engine. The classifiers include random forest (RF), support vector machine (SVM), classification and regression trees (CARTs), tree boosting (GBT), and naïve Bayes. Based on the knowledge of the study area, predominantly pastoral and agro-pastoral livelihood, a set of defined land use and land cover classes was determined, and a description of each class provided, as shown in Table 2 , was established. Training sites were established for the classification process.

Figure 3(a) illustrates a land use and land cover classification employing the support vector machine technique. The product inadequately classified the category labelled “Others”, which includes shadows, clouds, bare soil, bare rock sections, and any other ambiguous classifications. Furthermore, there is inadequate classification of cropland, built-up areas, and grassland.

Figure 3(b) illustrates a land use and land cover classification using the Naïve Bayes machine technique. This method misclassifies and generalizes all the classes.

Figure 3(c) illustrates a land use and land cover classification using the Gradient Boost machine technique. The product has overclassified grassland in the southern part of the region.

Figure 3(d) illustrates a land use and land cover classification using the Classification and Regression Tree technique. The product excessively classified water bodies and built-up areas, resulting in speckle noise with circular patterns shape.

Figure 3(e) shows a classification of land use and land cover using the Random Forest methodology. The product provides the enhanced classification of land use/cover categories; however, an imbalanced classification in cropland cover has been noted in the northwestern part of the study area.

i) Evaluation of Accuracy and Runtime Performance Metrics

Table 3 presents the detailed accuracies and run times of the algorithms used, where the random forest method demonstrated an overall accuracy of 84.4% and a kappa of 81.7%, with a runtime of 2.757 seconds. The gradient tree boost (BGT) achieved an accuracy of 76.9% and a kappa of 73.1%; however, it had a slightly longer runtime of 2.91 seconds. The classification and regression tree (CART) technique exhibited moderate efficacy, achieving 68.8% accuracy and a kappa of 63.5%, while requiring a duration of 2.999 s, implying a trade-off between precision and computational expense.

In contrast, the SVM achieved an accuracy of only 36.25%, a kappa statistic of 25.33%, and recorded the most extended runtime of 3.868 seconds, making it the least efficient option. In contrast, naïve Bayes displayed the lowest accuracy at 28.75%, a kappa of 17.12, and a duration of 2.758 seconds, indicating reduced predictive capability and limited efficiency.

The random forest is the most suitable algorithm for this study based on accuracy, reliability, and computational efficiency. Gradient boosting trees (GBT) provides a viable alternative, while CART balances accuracy and interpretability. SVM and naïve Bayes exhibited poor performance and are not recommended for this type of analysis.

ii) Performance Metrics for Accuracy Assessment and F-1 scores, including the Weighted F-1

a. Naïve Bayes (NB)

The naïve Bayes model results in Table 4 yield a producer accuracy (PA) of 3.45% and a user accuracy (UA) of 11.11%, with a low F1 score of 0.05, indicating difficulties in accurately classifying cropland. Dense shrubland achieved a PA of 17.39% and a UA of 40.00%, whereas Dryland Forest had a PA of 72.73% but a UA of only 17.39%, indicating confusion with other land types. Sparse shrubland had a PA of 5.56% and a UA of 20.00%, and built-up/settle land had a PA of 22.73% and a UA of 20.00%, resulting in a low F1 score of 0.21, highlighting issues in identifying built-up areas.

The weighted average F1 score of 0.23 highlights the model’s poor classification performance, indicating significant misclassification across all classes. This low score reflects the model’s struggle to balance precision and recall, leading to unreliable outputs. Thus, the naïve Bayes model is inadequate for land use classification in complex landscapes, limiting its use for environmental monitoring and land-use planning.

b. Gradient Tree Boost

The gradient boost model results in Table 5 for cropland yield a producer accuracy (PA) of 65.52% and a user accuracy (UA) of 73.08%, leading to an F1 score of 0.69. This indicates a significant performance improvement over the naïve Bayes model. Dense shrubland achieved a PA of 65.23% and a UA of 83.33%, and Dryland Forest had a PA of 90.91% and a UA of 62.50%, demonstrating better generalizability across land types. Sparse shrubland and built-up/settlement classes also improved markedly. Sparse shrubland had a PA of 88.89% and a UA of 69.57%, whereas built-up/settle land had a PA of 68.18% and a UA of 100.00%, resulting in a strong F1 score of 0.81. These results show that the model effectively captures distinct land cover categories.

The weighted average F1 score of 0.77 indicates strong classification performance, indicating balanced precision and recall across classes. A higher weighted F1 score suggests that gradient boosting effectively reduces misclassification, making it a more reliable model for land use classification and suitable for environmental monitoring and planning.

c. Support Vector Machine (SVM)

The support vector machine (SVM) model results in Table 6 indicate a producer’s accuracy (PA) of 37.93% and a user’s accuracy (UA) of 23.91% for cropland, yielding a low F1 score of 0.29, reflecting misclassification issues. Dense shrubland achieved a PA of 47.83% and a UA of 33.33%, whereas Dryland Forest had a PA of 81.89% but a UA of only 50.00%, indicating confusion with other land types.

Sparse shrubland and built-up/settle land also performed poorly, with sparse shrubland at PA 5.56% and UA 5.89% and built-up/settle land at PA 22.72% and UA 62.50%, resulting in a low F1 score of 0.33. These results suggest that the SVM model struggles to distinguish similar land cover classes, hindering its effectiveness in complex landscapes.

The weighted average F1 score of 0.36 shows challenges in classification accuracy. This low score highlights inconsistencies in precision and recall across land use classes, making SVM unreliable for land use classification. Therefore, SVMs may not be suitable for environmental monitoring and land-use planning, as they lack balanced classification across diverse land cover types.

d. CART

The CART model results in Table 7 for cropland yield a producer accuracy (PA) of 44.83% and a user accuracy (UA) of 65.00%, resulting in an F1 score of 0.53. This suggests moderate performance, indicating challenges in distinguishing cropland from other land covers. Dense shrubland achieved a PA of 60.87% and a UA of 66.67%, whereas Dryland Forest recorded a PA of 90.91% and a UA of 62.50%, reflecting superior classification. Sparse shrubland and built-up/settlement classes presented sparse shrubland with a PA of 66.67% and a UA of 70.59%, whereas built-up/settlement had a PA of 63.64% and a UA of 77.78%, culminating in a robust F1 score of 0.70. These results imply that the CART model facilitates improved classification of land cover types.

The weighted average F1 score of 0.68 suggests that CART offers enhanced precision and recall across various classes, rendering it a more dependable option for environmental monitoring and land-use planning. However, some misclassification issues remain, particularly in differentiating cropland from shrubland.

e. Rand Forest (RF)

The results for the random forest model in Table 8 regarding Cropland indicate a producer’s accuracy (PA) of 86.21% and a user’s accuracy (UA) of 78.13%, culminating in a high F1 score of 0.82. This reflects strong classification performance, with minimal misclassification errors. Dense shrubland achieved a PA of 86.61% and a UA of 79.17%. In contrast, Dryland Forest had a PA of 90.91% and a UA of 71.43%, underscoring the model’s capacity to distinguish between different land cover types accurately.

The sparse shrubland and built-up/settlement classes also demonstrated commendable performance. Sparse shrubland had a PA of 94.44% and a UA of 89.47%, whereas built-up/settle land had a PA of 77.27% and a UA of 94.44%, resulting in a high F1 score of 0.85. Tables 4-8 provide a detailed breakdown of the performance metrics for each algorithm, including producer’s accuracy, user’s accuracy, F1-scores for individual land cover classes, and the weighted average F1-score. These results further support the conclusions drawn from the overall accuracy and kappa statistics.

Land use and land cover (LULC) classification is essential in remote sensing and influences urban planning, agriculture, and ecosystem management. This discussion evaluates five machine learning techniques: SVM, naïve Bayes, gradient tree boost, CART, and random forest for LULC classification, as shown in Figures 3(a)-(e) . While SVM manages high-dimensional data and establishes solid classification boundaries, it struggles with complex land cover types, such as shadows and clouds and classes with high variability, such as cropland and grassland ( Figure 3(a) ). These limitations are likely due to the algorithm’s sensitivity to parameter settings and the need for careful feature selection and optimization.

Naïve Bayes struggles due to its feature independence assumption, often violated in remote sensing data with correlated spectral bands and spatial features ( Figure 3(b) ). This inherent limitation renders it unsuitable for complex LULC tasks, particularly those involving high-dimensional datasets. The gradient tree boost can handle nonlinear relationships and improve accuracy through ensemble learning; this has been observed in previous studies ( Pham et al., 2023 ). In this study, these methods tend to overclassify dominant classes such as grasslands, likely due to class imbalance or overfitting ( Figure 3(c) ). This highlights the importance of addressing class imbalance and employing regularization techniques to prevent model bias. Similarly, CART shows overclassification of water bodies and built-up areas, causing speckle noise and circular patterns, possibly due to overly complex decision boundaries ( Figure 3(d) ). These results suggest that parameter optimization, such as pruning, is crucial for improving CART’s performance in heterogeneous landscapes.

The random forest model is the most effective of the five methods; it uses ensemble learning and feature randomness to manage LULC data complexity and improve classification accuracy ( Figure 3(e) ). However, it shows imbalanced cropland classification in the northwestern area, likely due to insufficient training samples or challenges in capturing spectral variability; previous research identified a similar scenario while classifying in the Sahel region of West Africa [48]. Despite this limitation, the random forest model demonstrates robust performance for LULC classification, provided that sufficient and representative training data are used. This study underscores the necessity of suitable machine-learning techniques for LULC classification and highlights the significance of hybrid approaches and enhanced feature extraction across diverse landscapes.

The performance metrics of the five machine learning algorithms for land cover classification emphasize their accuracy, kappa statistics, and efficiency. The random forest method stands out as the top performer, achieving 84.4% accuracy and 81.7% kappa, while running for 2.757 seconds. This highlights its effectiveness in managing complex data efficiently. However, the accuracy achieved in this study could be further improved through the exploration of more advanced techniques. Studies in Ethiopia ( Fentaw & Abegaz, 2024 ) have reported accuracies approaching 90% using hybrid random forest-CNN models with Sentinel-2 data, suggesting that hybrid approaches could enhance the accuracy of LULC mapping. Furthermore, research in Saudi Arabia ( Sun & Ongsomwang, 2023 ) has shown the advantages of deep learning (U-Net) over random forests for edge detection in arid regions, indicating the potential for further improvements.

The gradient tree boost is a strong alternative with 76.9% accuracy and 73.1% kappa, though it has a longer runtime of 2.91 seconds. Despite its competitive accuracy, its high computational cost limits its suitability for large-scale use. The classification and regression tree (CART) achieves 68.8% accuracy and 63.5% kappa with a runtime of 2.999 seconds. While less accurate, CART is suitable when interpretability is prioritized over precision. Machine learning models like RF and GBT are often considered “black boxes,” hindering the understanding of their decision-making processes. SHAP analysis can address this limitation by quantifying the contribution of each feature to individual predictions, improving model transparency and informing evidence-based decision-making in land cover mapping ( Avci et al., 2023 ; Farda, 2017 ).

The SVM and naïve Bayes methods perform poorly, making them unsuitable for LULC classification. The SVM has only 36.25% accuracy and 25.33% kappa, with the most extended runtime of 3.868 s, indicating poor predictive ability and high computational cost. Naïve Bayes is even worse, with the lowest accuracy (28.75%) and kappa coefficient (17.12), despite a runtime similar to that of random forest (2.758 s). These results highlight the limitations of these methods in handling the complexities of LULC data, particularly in arid and semi-arid regions. The superior performance of the random forest aligns with studies in China that have reported better results compared to SVM ( Avci et al., 2023 ). Previous research has also shown that increasing the number of training samples can improve the accuracy and efficiency of random forests in land cover classification ( Badapalli et al., 2025 ; Zhao et al., 2024 ). Class imbalance can bias land cover classification results, leading to issues such as grassland overclassification (GBT) and poor cropland detection (Naïve Bayes) ( Almalki et al., 2022 ). Techniques like SMOTE (generating synthetic samples) and Focal Loss (penalizing rare class misclassification) can mitigate these issues ( Hinojosa Lee et al., 2024 ).

The F1-scores and weighted average F1-scores provide a more detailed assessment of the algorithms’ performance across individual land cover classes. Naïve Bayes exhibits the lowest average F1 score (0.23), reflecting its poor performance in classifying all land cover types. Gradient boosting and CART perform reasonably well, with F1 scores of 0.77 and 0.68, respectively. Gradient Boost excels in cropland (0.69) and built-up areas (0.81), effectively managing nonlinear relationships and minimizing misclassification. The CART shows strong results in dryland forests (0.70) and sparse shrublands (0.70) but struggles with croplands (0.53). Both models balance precision and recall, which is ideal for LULC classification. However, gradient boosting’s higher F1 score indicates superior reliability for highly accurate applications, while CART’s interpretability remains valuable.

The random forest outperforms all the models, with a weighted average F1 score of 0.85, the highest of the tested algorithms. It accurately classifies cropland (F1 score of 0.82), sparse shrubland (0.89), and built-up areas (0.85), with few misclassification errors. This superior performance stems from its ensemble learning approach, effectively managing the complexity and variability of LULC data. The high F1 scores show a balanced trade-off between precision and recall, making the random forest model the most reliable model for LULC classification. Its robustness makes it ideal for environmental monitoring and land-use planning applications requiring precise land cover data classification.

The study employed basic parameterization for the machine learning classifiers, limiting their potential performance. Future research should focus on advanced hyperparameter tuning to optimize the performance of each algorithm. Random Forest utilized 70 trees, and Gradient Boosting Trees employed 100, but other critical parameters like “maxDepth” and “learningRate” were absent. SVM, CART, and Naïve Bayes relied on default settings with no hyperparameter tuning. This lack of optimization contributed to suboptimal results, particularly for SVM (36.25% accuracy) and Naïve Bayes (28.75% accuracy). The absence of fine-tuning, especially concerning “maxDepth”, “minLeafPopulation”, “learningRate”, and kernel configurations, significantly limited these models’ robustness and generalization capability.

This study highlights the potential for improved classification accuracy through targeted refinement strategies. Future work should prioritize advanced hyperparameter tuning across all models, including exploring techniques to address class imbalance.

The study’s limitations include using single-date Landsat-9 imagery (January-April 2023), neglecting seasonal dynamics and potentially misclassifying spectrally similar classes. The lack of hyperparameter tuning and the absence of deep learning methods and advanced techniques for handling class imbalance restricted the comparability with state-of-the-art approaches. The absence of post-classification refinement, including noise filtering and texture analysis, diminished map accuracy. Addressing these gaps with multi-temporal data, hybrid modeling, and advanced hyperparameter tuning while managing class imbalances is crucial for enhancing the robustness of future LULC studies in arid and semi-arid regions.