Jul 2026· International journal of computer information systems and industrial management applications· Vol 18, pp. 294-309· 0 citations
TL;DR
The proposed framework offers a robust, scalable, and interpretable solution for proactive soil health management and efficient fertilizer utilization, contributing toward sustainable agricultural practices as well as reliable and actionable decision support for precision agriculture.
Abstract
Soil health management plays a central role in achieving sustainable agricultural productivity; however, conventional soil testing and fertilizer advisory practices are often labor-intensive, costly, and limited in precision. To overcome these limitations, this work presents a Hybrid Machine Learning and Optimization Framework for predictive soil health maintenance. The proposed framework integrates multiple predictive tasks namely soil nutrient estimation using NPK regression, fertilizer recommendation through classification, Soil Health Index (SHI) computation, and crop suitability prediction within a unified and data-driven decision-support architecture. The predictive modeling layer employs an ensemble of Random Forest, XGBoost, and Deep Neural Network (DNN) models to capture complex nonlinear relationships between soil and environmental parameters. To enhance predictive performance and generalization, model hyperparameters are automatically tuned using a Genetic Algorithm (GA), enabling efficient exploration of the hyperparameter search space. Comprehensive experiments conducted on a real-world soil dataset demonstrate that the proposed optimization-enhanced models consistently outperform baseline approaches.
For soil nutrient prediction, the Hybrid DNN + GA model achieved an approximate 20% reduction in RMSE compared to the baseline DNN, attaining a coefficient of determination R2 of 0.94. In the fertilizer recommendation task, the proposed approach achieved 94.2% classification accuracy with an ROC–AUC of 0.96, while crop suitability prediction reached an accuracy of 92.8%, representing a 4.6% improvement over conventional DNN-based models. The derived Soil Health Index exhibited a mean value of 0.78, with 58% of soil samples classified as Healthy, demonstrating the framework’s ability to translate complex predictions into an interpretable soil vitality indicator. The integration of predictive modeling, metaheuristic optimization, and interpretable soil health assessment enables reliable and actionable decision support for precision agriculture. The proposed framework offers a robust, scalable, and interpretable solution for proactive soil health management and efficient fertilizer utilization, contributing toward sustainable agricultural practices.
Agriculture is a central part of human subsistence, with classification of soil texture and nutrition-based crop recommendation being the central aspects of optimal agricultural practice. Nevertheless, traditional methods are subject to limitations of being less precise, computationally less optimal, and less versatile concerning varying soil and environmental conditions. Current deep learning models are frequently unable to compromise between performance and efficiency, whereas traditional optimization methods fail to handle high-dimensional agriculture data efficiently, resulting in suboptimal suggestions and poor real-time usage. To address these issues, this research presents AgriOptNet, a hybrid deep learning and optimization framework for intelligent soil texture classification and crop recommendation based on nutrition. AgriOptNet novelty is founded upon three integral constituents like Crop Recommendation through Entropy-Regularized Dynamic Deep Q-Learning with Adaptation to the Reward Function (MDQL-RA), optimally dynamic recommendations of crop inputs depending upon the health of soil, yield records, and surrounding environmental aspects and utilizing entropy regularization to accelerate exploration; Classification using a newly invented lightweight deep-learning model called SoilCropNet with a compound based on MobileNetV2, EfficientNetV2, and ShuffleNetV2 and provides precise, and computationally favourable classification along with squeeze-and-excitation as well as depth-wise separable convolutional enhanced properties; Feature selection through newly developed hybrid SailDragon Optimizer (SDO), combining Sailfish Optimization (SFOA) and Dragonfly-Based Optimization (DBOA), to obtain best-informing features for predictions without errors. The proposed AgriOptNet framework demonstrates superior performance with an accuracy of 99.87% and an F1-score of 98.75%, significantly outperforming existing techniques and ensuring high precision and efficiency for real-time precision agriculture applications.
Background: Determining the health of soil and predicting the output of agriculture are complicated processes owing to regional variation and complex nutrient interactions among soil properties. Methods: AgroAdvisor acts as a decision support system using predictive models, particle swarm optimization (PSO) and retrieval-augmented generation (RAG). Soil health parameters such as nitrogen (N), phosphorous (P), potassium (K), pH and moisture levels are analyzed through a hybrid regression model incorporating random forest and gradient boosting. PSO is used for hyperparameter optimization, which improves the accuracy of prediction compared to traditional methods. Result: Experimental testing on soil samples from South India demonstrated a decrease in mean absolute error ranging between 18% to 23% due to PSO optimization. The retrieval-augmented generation technique, based on scientific papers, ICAR/FAO guidelines and local soil management techniques, generated contextually relevant and internally consistent recommendations for fertilizer usage, crop rotation and soil management.
P. Nithya, G. S. Kumar· Indian Journal of Agricultur...· 0 citations
Precision agriculture has emerged as a critical paradigm for sustainable food production, necessitating intelligent
systems for crop yield prediction and soil nutrient monitoring. This pa-per presents a novel hybrid deep learning framework
integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and self-attention mechanisms
for accurate crop prediction and real-time soil nutrient analysis. The proposed system employs Internet of Things (IoT) sensors
for continuous monitoring of soil parameters including nitrogen (N), phosphorus (P), potassium (K), pH, moisture content,
temperature, and electrical conductivity. A comprehensive dataset comprising 12,500 agricultural samples from diverse agroclimatic zones was utilized for model training and validation. The hybrid CNN-LSTM-Attention architecture achieves a
prediction accuracy of 96.8% for crop classification and a mean absolute percentage error (MAPE) of 4.2% for yield prediction,
significantly outperforming conventional machine learning approaches. Experimental results demonstrate that the proposed
framework reduces prediction error by 23.5% compared to standalone LSTM models and 31.2% compared to traditional
Random Forest classifiers. The sys-tem provides actionable recommendations for nutrient management, contributing to
optimized fertilizer application and enhanced agricultural sustainability. The integration of edge computing with cloud-based
analytics enables real-time decision support for farmers, achieving latency under 500 milliseconds for prediction queries.
R. Bansod, Durvesh Patil, Bramhadev Wadibhasme· International Journal for Re...· 0 citations
The current study describes an advanced hybrid multi-modal approach that will be used in predicting soil health and crop productivity through the application of several types of deep learning models and machine learning approaches to enhance predictions. In the current work, a hybrid model is considered, and three kinds of pre-trained CNN, known as ResNet50, VGG16, and Mobile Net, are used to extract the features of soil. The pre-fetched features are subsequently processed through multiple fully connected layers with SoftMax activation functions for multi-class types of soil classification, namely clay, sandy, and loam. To enhance the generalization and robustness of the models, data augmentation and normalization techniques are applied to the data. The models are optimized with Adam and SGD optimizers using categorical cross-entropy parameters for the loss function. The results of the experimentation show high accuracy and robust performance for the precision-recall and F1-score metrics. The model would highly benefit precision agriculture applications to facilitate autonomous soil type detection, and to enable farmers and growers to better inform crop selection and nutrient management for sustainable yield optimization models in smart farming.
D. Stephen, D. Ferlin, D. Shahila· International Conference on...· 0 citations
Precision agriculture is an important application scenario for chemical engineering to carry out the clean production and sustainable resource use. In this paper, we propose an explainable data-centric ensemble machine learning framework for the optimization of agrochemical (fertilizer and pesticide) inputs based on utilizing their input efficiency and reducing environmental pollution based on chemical engineering aspects. We utilize the multiple source agricultural data (soil N, P, K, pH, temperature, humidity and rainfall) to make it possible for the precise application and site-specific planting adaptation of agrochemicals.
Integrating three models including Logistic Regression, Support Vector Machine, Decision Tree, we utilize the Voting Classifier and Stochastic Gradient Boosting (SGB) to implement the classification. Together with controlled noise addition and cross validation that utilize data-centric methods, the Voting Classifier performs the best accuracy 90.7% with perfect score balance between precision, recall and F1-score.
SHAPley Additive ExPlanations (SHAP) and permutation feature importance methods are adopted for model interpretation and illustrate the dominant features are rainfall, humidity and nitrogen that are consistent with agricultural chemical transport and nutrient conversion process.
The framework can be further used for VRT systems to realize automated and quantitative inputs and reduce input of fertilizer and pesticides; soil and water pollution is limited, resource use efficiency is high. A generalized, interpretable and engineering-practical framework is proposed which is important for the chemical engineering application of clean production in precision agriculture.
Padma Nilesh Mishra, Kinjal Doshi, Rupali Jadhav et al.· Applied Chemistry for Engine...· 0 citations
The aim of this research is to develop soil health prediction model and to compare the performance of XGBoost with the proposed algorithm, CatBoost, in enhancing fertilizer application decisions. The two groups evaluated the dataset. Group 1 used XGBoost tuned for its hyperparameters, while Group 2 used CatBoost that featured ordered boosting to resolve categorical soil attributes. A Kaggle dataset containing more than 10,000 soil samples was refined, prepared, and split into training, validation, and testing sets for evaluating the model. The XGBoost achieved an average accuracy of 85.87%, whereas CatBoost significantly outperformed it, achieving accuracy between 94.80% and 96.70%, with a total average of 95.55%. Statistical evaluation validated that this discrepancy is extremely significant (t = –60.275, p = 0.000), CatBoost’s enhanced predictive performance This means that CatBoost yields more stable and reliable predictions of soil health. CatBoost outperformed and showed greater stability than XGBoost in modeling soil nutrients for optimizing fertilizer use. Its high and consistent performance allows it to be considered a reliable tool in precision agriculture for making rational decisions on soil health assessment.
N.Prakash, M.Kavinandhini, K.Jeevitha et al.· International Conference Com...· 0 citations