A Robust Crop Recommendation System Leveraging Soil and Climate Parameters

Introduction

Agriculture is a cornerstone of the global economy, essential for supplying food, raw materials and employment. The global population is projected to reach 9.7 billion by 2050 with demand for agricultural products expected to increase by 60% from 2010 levels [3,4]. This growth is putting immense pressure on existing agricultural practices to enhance efficiency and sustainability. Traditional farming methods, which often rely on the farmer’s intuition and experience, may not be sufficient to meet these demands sustainably. In particular, Africa faces significant challenges with its population expected to double to approximately 2.5 billion by 2050 and nearly 20% of its population already affected by food insecurity [3,5]. Innovative approaches, such as machine learning, must therefore be explored to optimize agricultural productivity and ensure food security in the face of climatic and demographic changes.

In recent years, machine learning (ML) have revolutionized various industries, including agriculture. These technologies offer powerful tools for analyzing vast datasets and deriving actionable insights that can significantly enhance decision- making processes. Specifically, ML can be used to develop advanced crop recommendation systems that consider factors such as soil characteristics, climatic conditions and historical crop performance to provide tailored advice to farmers. Additionally, recent advancements in IoT-based data collection enable real-time environmental monitoring, which can greatly improve the accuracy of these ML-based recommendations by integrating live data on weather, soil moisture and crop health [6–8]. These developments in IoT technology facilitate dynamic, timely decision-making, making ML-based crop recommendation systems more adaptable to changing conditions on the field.

This study aims to investigate the potential of machine learning techniques in optimizing agricultural practices by developing and evaluating a crop recommendation system. The primary objective is to compare the performance of different ML models, namely Random Forest (RF), Support Vector Machine (SVM), Convolutional Neural Network (CNN) and Long Short- Term Memory Network (LSTM), in providing accurate crop recommendations [2,9,10]. By assessing precision, recall and F1 scores of these models, we seek to identify the most effective approach for enhancing agricultural productivity. The remainder of this paper is structured as follows: Section 2 reviews existing approaches in crop recommendation and agricultural optimization using machine learning and presents the taxonomy of crop recommendation models, including KNN, LSTM, ensemble methods and soil/yield-based systems. Section 4 analyzes the experimental results and discusses their implications in the context of precision agriculture. Finally, Section ?? concludes the study by summarizing key findings, identifying limitations and outlining future research directions.

Related Works

This section reviews the existing literature on the application of ML techniques in agricultural optimization, focusing on crop recommendation systems, crop yield prediction systems, soil analysis, spectral analysis and climate prediction.

Crop Recommendation Methods

Precision agriculture increasingly relies on machine learning (ML) algorithms to analyze soil and climate data for recommending the most suitable crops. Table 1 categorizes crop recommendation approaches based on the methods and techniques employed. One of the foundational approaches involves KNN-based systems, which leverage similarity metrics to compare current soil and climate conditions with historical data. Studies such as have used KNN to recommend crops adapted to specific soil properties by analyzing feature proximity between agricultural zones [11]. Building upon temporal aspects of agricultural data, RNN-based systems, particularly those using Long Short-Term Memory (LSTM) networks such as Sreemathy et al. (2023) and Rajkumar et al. (2023), aim to capture seasonality and time-dependent pat- terns [2,12]. For instance, Sreemathy and Prasath proposed a BiLSTM- MERNN model to improve long-term prediction reliability by modeling temporal dependencies in crop performance [2].

To further enhance robustness and accuracy, ensemble methods have been adopted. These approaches combine multiple models— such as Random Forest (RF) [13–15]. Support Vector Machine (SVM) and Naive Bayes (NB)—to improve predictive performance across diverse agricultural scenarios, as demonstrated in studies like [13,14,16,17]. Complementing these techniques, general ML- based approaches apply standard algorithms independently to capture complex, non-linear interactions among variables such as soil nutrients, pH levels and climate data. Algorithms like RF, SVM, Na¨i±ve Bayes, Bagging and Boosting have been widely explored in this regard. Focusing more narrowly on soil characteristics, soil condition-based methods prioritize nutrient profiling (e.g., NPK) and soil pH as primary decision variables [13,15,18,19]. Research such as highlights the effectiveness of these models for guiding regionally adapted crop selections [6,13,14]. In contrast, yield prediction-based systems emphasize forecasting future productivity by analyzing historical crop yield patterns. By incorporating past outcomes into current recommendations, these models offer a decision-theoretic approach to optimizing land use [7,8,12,20,21]. Finally, spectral analysis-based approaches make use of hyperspectral remote sensing to assess soil texture and composition over large areas. This technique has proven particularly valuable for large-scale land monitoring, as evidenced by the work of Vibhute et al. and Datta et al. [6,22].

                                Table 1: Categories of Precision Agriculture Models with Descriptions and References

This article adopts ML-based crop recommendation methods due to their capacity to model complex, nonlinear relationships in agricultural data while remaining computationally feasible for deployment in low-resource environments. Their empirical performance and simplicity of implementation make them a compelling choice for precision agriculture applications.

ML-based Crop Recommendation Methods

Machine Learning (ML) has emerged as a cornerstone of modern crop recommendation systems, enabling data-driven decision- making based on complex interactions among soil characteristics, climate conditions and crop suitability. These systems employ a range of supervised learning algorithms to model non-linear relationships and enhance precision in agricultural planning. Figure 1 presents a typical ML-based crop recommendation pipeline, which often integrates ensemble learning techniques. The process typically begins with the acquisition and preprocessing of environmental and agronomic data, followed by feature engineering to extract meaningful patterns. Subsequently, models are trained using algorithms such as K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machine (SVM), Na¨i±ve Bayes (NB) and Bagging (BG), along with ensemble strategies like the VotingClassifier to aggregate predictions for improved robustness and accuracy.

                      Figure 1: ML-based Crop Recommendation Workflow Integrating Ensemble Methods

To provide a comprehensive overview of ML models used in this domain, Table 2 presents a comparative summary derived from an in-depth literature review. The mod- els are evaluated based on five critical dimensions: accuracy/performance, processing time, scalability, robustness to outliers and the quality of crop recommendations.

                           Table 2: Comparative Table of ML-Based Crop Recommendation Methods (from literature review)

These models were evaluated according to five relevant criteria. Among them, performance indicates the prediction accuracy of the model. Crop recommendation quality reflects the relevance of the recommendations made, while processing time evaluates time efficiency. Scalability measures the ability of the algorithm to handle increasing data volumes and outlier robustness assesses the algorithm’s resistance to noise and anomalies in the dataset. From the table, it is evident that ensemble methods stand out with top scores in terms of both accuracy and recommendation quality. RF and SVM also achieve excellent results, while KNN performs well in accuracy but is limited in scalability. Despite longer computation times, all models—except for NB and CHAID—demonstrate good robustness to outliers with RF and BG particularly strong in this regard. In light of these comparative findings, the selected models for implementation in this study include KNN, RF, NB, SVM, BG and the ensemble Voting Classifier method. These approaches have consistently shown the ability to deliver accurate and meaningful crop recommendations with satisfactory computational performance.

Datasets

To evaluate machine learning models for crop recommendation, we considered three publicly available datasets summarized in Table 3. These datasets differ in size, attribute variety, structure and data quality.

                 Table 3: Comparison of Datasets used in Crop Recommendation and Yield Prediction Studies

The datasets differ significantly in terms of scale, dimensionality and quality. While Datasets A and B offer larger sample sizes and mixed-type features, both suffer from missing values, which can impact model training and introduce biases if not addressed carefully. Dataset C, though smaller in size, is fully cleaned and contains no missing values. It includes essential agricultural features such as nitrogen, phosphorus, potassium (NPK), temperature, humidity, pH and rainfall—all critical for soil and climate-based crop prediction. Due to its completeness and relevance, we selected Dataset C for our study to ensure methodological rigor and reproducibility. Moreover, the dataset’s compact size allows for faster iteration and model tuning without the complexities of imputation or variable encoding.

Positioning

In line with ensemble-centric studies and soil/spectral analyses, our focus is methodological rigor and interpretability on a curated soil–climate dataset with a deployment lens for West African smallholders [6,16,17, 22,24].

Methods

Dataset and Task

We use the curated crop recommendation dataset (“Dataset C”) of N =2,200 samples and seven numeric features (N, P, K, pH, rainfall, humidity, temperature) with crop labels [1]. No missing values are reported. The task is multi-class crop classification from soil and climate measurements.

Preprocessing and Leakage Control

All experiments are implemented with scikit-learn Pipelines to prevent data leakage [26]. We standardize features for KNN and SVC using Standard Scaler; tree- based models (RF, BG) consume raw features. Class weights for SVC are tuned on a grid (when applicable). Because the dataset is clean [1], no imputation is required; nonetheless, we clip extreme values at the 0.5/99.5 percentiles in the training folds to mitigate rare outliers.

Models and Hyperparameter Search

We evaluate:

• KNN: k ∈ {3, 5, 7, 9, 11}; distance metric ∈ {Euclidean, Minkowski(p ∈ {1, 2})}.

• SVC (RBF): C ∈ {10−1, 1, 10, 102}, γ ∈ {10−3, 10−2, 10−1}; probability calibration via Platt scaling inside CV when used in soft voting.

• RF: nestimators ∈ {200, 400, 600}, max depth ∈ {None, 10, 20}, min samples leaf ∈ {1, 2, 4}, max features ∈ {sqrt, log2} [15].

• NB (Gaussian): var-smoothing ∈ {10−12, 10−10, 10−8}.

• Bagging: base learner Decision Tree with max depth ∈ {None, 10, 20} and nestimators ∈ {50, 100, 200}.

• Voting (soft): RF + SVC + NB; base estimators use their best inner-CV hyperparameters; soft probabilities require inner calibration for SVC.

Validation Protocol and Metrics

We adopt nested stratified k-fold CV (outer k=10; inner k=5). The inner loop selects hyperparameters; the outer loop estimates generalization. We report macro-averaged

<img src="https://www.opastpublishers.com/scholarly-images/10246-69a7abb78fd3f-a-robust-crop-recommendation-system-leveraging-soil-and-clim.png" width="500" height="600">

Accuracy, Precision, Recall and F1 with 95% confidence intervals from B=1,000 boot- strap resamples of outer-fold predictions. For pairwise model comparisons on the same test predictions, we apply McNemar’s test (continuity-corrected). '

Model Interpretability

To address feature importance and interpretability, we compute:

1. Permutation importance (macro-averaged over outer folds) on the best model to quantify global influence.

2. Partial Dependence (PD) and Individual Conditional Expectation (ICE) for top features (e.g., pH, rainfall, N), reporting agronomically plausible monotonic- ities and interaction hotspots. These plots help practitioners understand how changes in pH or rainfall shift class probabilities, complementing black-box metrics.

Temporal Extension Blueprint

Although Dataset C is static, seasonal and inter-annual dynamics are critical [2]. We outline a future data pipeline aggregating daily weather into monthly descriptors (means, anomalies, variability) feeding a BiLSTM/temporal model; see Figure. 3.

Figure 2: Leakage-safe nested CV with Calibrated Soft-Voting and Post-Hoc Interpretability.

Results and Discussion

In this section, we present and analyze the experimental results obtained from the implementation of the proposed crop recommendation system, illustrated in Fig. ??. The objective is twofold: (i) to assess the predictive performance of several machine learning (ML) models applied to the agricultural dataset and (ii) to interpret these results in light of the underlying soil and climatic characteristics that shape crop suitability. By evaluating individual models as well as ensemble strategies, we aim to identify both the methodological strengths and practical limitations of current approaches.

Results

Figure 4 summarizes the performance of the six classifiers considered in this study. Accuracy, precision, recall and F1-score were computed under a cross-validation protocol, ensuring a fair comparison across methods. KNN. The KNeighborsClassifier achieved strong performance with an accuracy of 98.63% and precision of 98.80%. Despite its simplicity, KNN is memory- intensive and its performance depends heavily on the choice of k and distance metric, making it less scalable for larger datasets.

Figure 3: Blueprint for Future Seasonal Modeling and Real-Time Integration (not used in current experiments).

Figure 4: Performance Metrics (accuracy, precision, recall, F1-score) for all Classifiers.

Random Forest. The Random Forest Classifier performed even better with accuracy and precision of 99.54% and 99.56% respectively. It proved highly effective in capturing complex non- linear relationships between soil nutrients and climate variables. How- ever, RF can be computationally demanding and may introduce bias toward majority classes if data are imbalanced. GaussianNB. Surprisingly, GaussianNB also yielded very high accuracy (99.54%) and precision (99.58%). Its efficiency and speed make it attractive for low-resource contexts. Yet, its assumption of feature independence is unrealistic in agricultural data, where correlations between soil nutrients (e.g., N, P, K) often exist. SVC. The Support Vector Classifier reached 99.09% accuracy. It is well-suited for linearly separable problems but faces scalability challenges in higher dimensions or with larger datasets. Kernelized variants mitigate this but at a higher computational cost. Bagging. The Bagging Classifier obtained 99.31% accuracy, reflecting the strength of bootstrap aggregation in reducing variance. Its advantage is most evident when combined with unstable learners, though its gains are less pronounced with stable models such as RF. Voting Ensemble. Finally, the soft-voting ensemble combining RF, SVC and NB delivered the best performance overall with accuracy and precision of 99.77% and 99.78% respectively. This highlights the benefit of integrating complementary classifiers to balance bias–variance trade-offs and enhance robustness across heterogeneous input distributions.

Discussion

The results confirm that supervised ML approaches can deliver high-quality agricultural recommendations when applied to well- structured soil–climate datasets. In particular, ensemble learning (VotingClassifier) consistently outperformed individual models, supporting the view that aggregating diverse decision boundaries reduces overfitting and improves generalization [13,14,18].

Nevertheless, two limitations must be emphasized:

• Dataset Size. The dataset used (2,200 samples) is relatively small, limiting the statistical power of the reported results. High accuracies on small datasets may reflect overfitting and the absence of external validation raises concerns about generalizability to other agroecological zones.

• Static Data. The dataset is historical and static. It does not account for seasonal variations or integrate real-time data streams (e.g., rainfall fluctuations, pest out- breaks). This limits the system’s applicability in rapidly changing environmental conditions. Despite these constraints, the study demonstrates a strong proof- of-concept: crop recommendation systems powered by ML can help farmers optimize decision-making, improve yields and support sustainable resource use.

Future Enhancements should focus on (i) hyperparameter optimization using auto- mated search techniques (e.g., Bayesian optimization), (ii) enlarging and diversifying the dataset to include multiple regions and temporal dynamics, (iii) incorporating IoT-based real-time monitoring and remote sensing to adapt recommendations under changing conditions and (iv) improving interpretability through feature importance, partial dependence and sensitivity analysis. Such developments will enable more reliable, transparent and farmer-centric decision-support systems, bridging the gap between experimental validation and practical deployment.

Conclusion

We provide a transparent, deployment-oriented benchmark of classical ML methods for crop recommendation on a curated soil–climate dataset. Using leakage-safe nested CV, calibrated soft voting and post-hoc interpretability, we obtain top-line accuracy (best single-run: 99.77%) while candidly acknowledging overfitting risks at small N . The value of this work lies in the rigorous protocol, interpretability layer and practical guidance for ethically aligned, accessible systems in low-resource contexts. Future work will prioritize temporal data integration, external validation across agro-ecologies and field trials.

References

1. Rao, M. S., Singh, A., Reddy, N. S., & Acharya, D. U. (2022). Crop prediction using machine learning. In Journal of Physics: Conference Series (Vol. 2161, No. 1, p. 012033). IOP Publishing.
2. Sreemathy, J., & Prasath, N. (2023, April). Crop recommendation with BiLSTM-MERNN algorithm for precision agriculture. In 2023 International Conference on Networking and Communications (ICNWC) (pp. 1-5). IEEE.
3. Food and Agriculture Organization of the United Nations. The Global Popula- tion Projected to Reach 9.7 Billion by 2050. Accessed: 2024-07-28. n.d.
4. World Bank. The demand for agricultural products is set to increase by 60% compared to 2010 levels. [Online; accessed 2024-07-28]. 2010.
5. United Nations. Africa faces significant challenges, with its population expected to double to approximately 2.5 billion by 2050. Accessed: 2024-07-28. n.d.
6. Vibhute, A. D., Kale, K. V., & Gaikwad, S. V. (2024). Machine learning-enabled soil classification for precision agriculture: a study on spectral analysis and soil property determination. Applied Geomatics, 16(1), 181-190.
7. Lee, H., & Moon, A. (2014, February). Development of yield prediction system based on real-time agricultural meteorological information. In 16th international conference on advanced communication technology (pp. 1292-1295). IEEE.
8. González Sánchez, A., Frausto Solís, J., & Ojeda Bustamante,W. (2014). Predictive ability of machine learning methods for massive crop yield prediction.
9. Thilakarathne, N. N., Bakar, M. S. A., Abas, P. E., & Yassin,H. (2022). A cloud enabled crop recommendation platform for machine learning-driven precision farming. Sensors, 22(16), 6299.
10. Rybacki, P., Niemann, J., Derouiche, S., Chetehouna, S., Boulaares, I., Seghir, N. M., ... & Osuch, A. (2024).Convolutional neural network (cnn) model for the classification of varieties of date palm fruits (Phoenix dactylifera L.). Sensors, 24(2), 558.
11. Rawat, P., Bajaj, M., Vats, S., & Sharma, V. (2023, March). An analysis of crop recommendation systems employing diverse machine learning methodologies. In 2023 International Conference on Device Intelligence, Computing and Communication Technologies,(DICCT) (pp. 619-624). IEEE.
12. Rajkumar, N., & Mukunthan, M. A. (2023, December). Efficient Crop yield Analysis Prediction in Modern Agriculture System using Machine Learning Algorithm. In 2023 International Conference on Data Science, Agents & Artificial Intelligence (ICDSAAI) (pp. 1-4). IEEE.
13. Patil, N. A., & Mane, S. (2024). Crop recommendation in precision agriculture using machine learning techniques.
14. Bhatt, N., & Varma, S. (2023, February). Recommendation System for Crops Integrating with Specific soil parameters by Machine Learning Techniques. In 2023 IEEE International Students’ Conference on Electrical, Electronics and Computer Science (SCEECS) (pp. 1-8). IEEE.
15. Hansch, R. (2018). Handbook of random forests: Theory and applications for remote sensing. World Scientific Publishing Co Pte Ltd.
16. Reddy, A. (2019). Crop recommendation system to maximize crop yield in ramtek region using machine learning. International Journal of Scientific Research in Science and Technology.
17. Akshatha, K. R., & Shreedhara, K. S. (2018). Implementation of machine learning algorithms for crop recommendation using precision agriculture. International Journal of Research in Engineering, Science and Management (IJRESM), 1(6), 58-60.
18. Sani, S. R., Ummadi, S. V. S., Thota, S., Muthineni, N., Swargam, V. S. S., & Ravella, T. S. (2023, May). Crop recommendation system using random forest algorithm in machine learning. In 2023 2nd International Conference on Applied Artificial Intelligence and Computing (ICAAIC) (pp. 501-505). IEEE.
19. Reddy, A. (2019). Crop recommendation system to maximize crop yield in ramtek region using machine learning. International Journal of Scientific Research in Science and Technology.
20. Savla, A., Israni, N., Dhawan, P., Mandholia, A., Bhadada, H., & Bhardwaj, S. (2015, March). Survey of classification algorithms for formulating yield prediction accuracy in precision agriculture. In 2015 International Conference on Innovations in Information, Embedded and Communication Systems (ICIIECS) (pp. 1-7). IEEE.
21. Shinde, K., Andrei, J., & Oke, A. (2015). Web based recommendation system for farmers. International Journal on Recent and Innovation Trends in Computing and Communication, 3(3), 41-52.
22. Datta, D., Paul, M., Murshed, M., Teng, S. W., & Schmidtke,L. (2023). Comparative analysis of machine and deep learningmodels for soil properties prediction from hyperspectral visual band. Environments, 10(5), 77.
23. Kumar, R., Gupta, M., & Singh, U. (2023, June). Precision agriculture crop recommendation system using knn algorithm. In 2023 International Conference on IoT, Communication and Automation Technology (ICICAT) (pp. 1-6). IEEE.
24. Kulkarni, N. H., Srinivasan, G. N., Sagar, B. M., & Cauvery,N. K. (2018, December). Improving crop productivity through a crop recommendation system using ensembling technique. In 2018 3rd international conference on computationalsystems and information technology for sustainable solutions (CSITSS) (pp. 114-119). IEEE.
25. Nishant, P. S., Venkat, P. S., Avinash, B. L., & Jabber, B. (2020, June). Crop yield prediction based on Indian agriculture using machine learning. In 2020 international conference for emerging technology (INCET) (pp. 1-4). IEEE.
26. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ... & Duchesnay, É. (2011). Scikit- learn: Machine learning in Python. The Journal of machine Learning research, 12, 2825-2830.