A Robust Lemuria Framework for efficient crop prediction

Abstract

Agriculture remains a critical pillar of the Indian economy, yet yield forecasting continues to be affected by climatic uncertainty and diverse environmental conditions. To address these challenges, this study introduces the Robust Lemuria Framework (RLF), a deep ensemble hybrid model that integrates a Deep Belief Network (DBN) for hierarchical and non-linear feature abstraction with a diversified ensemble of Random Forest (RF), J48 Decision Tree (DT), and Naïve Bayes (NB) classifiers for stable prediction consensus. The novelty of RLF lies in its two-stage optimized preprocessing pipeline, which applies DBN-based pre-training to eliminate noise, reconstruct missing values, and refine complex agricultural features through non-linear dimensionality reduction. The framework is trained and evaluated using a decade of multi-regional Indian agricultural data (2010–2020), capturing variations across climate zones, crops, and seasonal patterns. Experimental results show that RLF significantly outperforms existing machine learning and deep learning approaches, achieving 98.99% accuracy, 98.54% sensitivity, 99.35% specificity, and an R² score of 0.9994 for yield prediction. These outcomes demonstrate the robustness, scalability, and real-world applicability of the model for agricultural forecasting. Overall, the proposed framework provides a reliable decision-support tool for precision agriculture, contributing to improved crop planning, resource allocation, and policy formulation.

Introduction

The agriculture sector is considered to be an important part of the international economy. Over the past thousand years, the number of individuals working in this field has increased significantly¹. Sustainable development and food security are some of the factors that have to be considered when it comes to developing this sector. According to experts, by 2050, there will be around ten billion people². The steady growth of the global population has placed substantial pressure on the food supply system, raising concerns about long-term food security. As demand continues to rise, essential resources such as water, arable land, and other ecological inputs may become insufficient to support the level of agricultural production required³. An essential consideration in agriculture today is the implementation of more intelligent and efficient methods for managing natural resources.

Researchers across the globe are actively exploring sustainable agricultural practices aimed at improving the productivity of key staple crops, including wheat, rice, and maize. These efforts focus on enhancing yield potential while ensuring long-term environmental balance⁴. These crops are known to be important sources of nutrition for everyone. Due to the high food prices and the natural resource depletion caused by climate change, the research community has been focusing on increasing the productivity of these crops based on climate change and resource requirements⁵. The use of data science tools and sensors in agriculture has raised hopes of farmers for improved and more effective ways of increasing their production. Among the emerging tools, crop-yield forecasting and machine learning techniques have gained significant importance. These approaches enable researchers to examine how different environmental conditions influence crop performance and overall productivity⁶.

The process of developing a yield prediction for crops involves many factors that can affect the production of the crop⁷. These include the weather conditions, soil characteristics, and the type of fertilizers that are used. The uncontrolled factors, on the other hand, include the frequency of irrigations, temperature, and wind speed⁸. In addition to these, other factors like the selection of the appropriate fertilizers and seeds are also taken into consideration to enhance the farm’s efficiency⁹. Because crop productivity is shaped by intricate interactions between climatic variations and human agricultural practices, reliable techniques for forecasting yield have become essential. Accurate prediction models are therefore crucial for understanding and managing these complex dependencies¹⁰.

Artificial neural networks have become a widely adopted computational technique in modern research due to their ability to model complex patterns within data¹¹. This model has the inherent capability to deal with the multiple complex behaviors, however it has some more drawbacks¹². The complexity of implementing artificial neural networks requires the use of parallel processing power¹³. This is why the equipment’s realization is dependent on its structure. One of the most critical issues that ANN faces is the lack of a clear explanation as to how the network works¹⁴. This lack of trust in the network leads to various issues. The structure of a network can be achieved through trial and error. However, showing the problem to the system can be challenging since ANNs can rely on numerical information¹⁵. Before an artificial neural network can be implemented, it has to translate complex problems into numerical values. This process will influence its performance. The duration of a network’s existence is also dependent on the user’s ability. For instance, if the user completes training, the network’s duration will be reduced to a certain value¹⁶.

Recent advancements in deep learning have led to noticeable improvements in crop-yield forecasting, as these models are capable of capturing intricate and non-linear relationships present in agricultural datasets more effectively than conventional machine learning methods¹⁷. This is due to how these models are built upon a framework that includes a variety of components, such as memory pools, gates and activation functions. This framework allows deep learning models to automatically extract various features from the collected data. They can also generate higher-level features by taking into consideration the lower-level models’ configuration¹⁸. Its robust learning capabilities and hierarchal framework make it ideal for both prediction and classification. The agricultural sector of India is struggling with technology integration to achieve the desired outputs¹⁹. Unpredictable climatic conditions and the effects of global warming have made it increasingly difficult for farmers to estimate future crop yields with confidence. This issue is considered a major challenge faced by the industry. Through the use of machine learning techniques, farmers can now receive timely and accurate information about the weather patterns, allowing them to make more informed decisions. Researchers in the past have been developing methods that can help improve India’s agricultural development²⁰.

This study aims to design an integrated predictive framework that combines machine learning methods with deep learning architectures to estimate crop yields more accurately. By generating reliable forecasts, the model supports farmers in making timely and informed decisions, ultimately enhancing overall agricultural output. The central goal of this research is to improve crop production levels, especially in situations where adverse weather conditions may lead to considerable losses. Deep neural networks, consisting of multiple hierarchical non-linear layers, progressively extract complex feature representations, thereby increasing predictive precision. The Robust Lemuria Framework utilises a Deep Belief Network (DBN) for learning representative features and assessing model performance, allowing it to handle heterogeneous agricultural data in an efficient manner.

The RLFs ensemble classification features the J48 Decision Tree, Nave Bayes, and RF. These three algorithms have complementary performance and algorithmic characteristics. Random Forest was chosen because of its capacity to minimize overfitting, manage large and complex feature spaces, and maintain strong generalisation across diverse and variable agricultural datasets. Naïve Bayes(NBs), while based on probabilistic independence assumptions, provides rapid and scalable inference and performs well in cases with limited data or noisy attributes—conditions often present in real-world agricultural datasets. The J48 Decision Tree was included for its interpretability and rule-based structure, which enables intuitive insight into the feature-space relationships influencing crop prediction. The combination of these classifiers creates a balanced ensemble that merges probabilistic reasoning, decision boundaries, and bagging-based robustness. This strategic hybridization improves overall model stability and predictive accuracy when compared to single-model or homogenous ensemble alternatives.

Crop-yield forecasting has become integral to modern precision agriculture, as it directly supports critical activities such as food security planning, distribution of agricultural inputs, and strategic resource management. Although yield prediction is a critical task, many existing forecasting methods rely on traditional machine learning techniques that have limited capability to model the complex and non-linear relationships inherent in agricultural data. In addition, several of these approaches do not incorporate effective data preprocessing mechanisms, making them sensitive to noise and variability. As a result, their performance often declines when applied across different agro-climatic regions. These limitations result in suboptimal predictive accuracy and reduced real-world applicability.

This work aims to design an innovative hybrid modelling approach capable of effectively overcoming the limitations observed in existing crop-yield prediction techniques. The framework is called the Robust Lemuria that combines DBNs for deep feature learning with an ensemble of classifiers-RF, Naive Bayes, and J48 Decision Tree—for enhanced prediction robustness. The framework is further supported by a two-stage preprocessing pipeline designed to handle missing values, reduce dimensionality, and eliminate data noise. This comprehensive approach aims to improve both classification accuracy and regression performance in predicting crop yields.

Although machine learning has advanced agricultural forecasting considerably, many current models still face difficulties in representing the highly intricate and non-linear interactions among climatic variables, soil properties, and crop responses—interactions that vary widely across India’s diverse agro-ecological regions. Many conventional models either rely on limited feature sets, use shallow learning architectures, or lack robust preprocessing mechanisms, resulting in reduced accuracy and weak generalisation when exposed to heterogeneous agricultural data. Moreover, the frequent climatic fluctuations experienced in India—such as irregular monsoons, rising temperatures, and erratic rainfall—further expose the limitations of existing techniques. This creates a clear research gap: the need for an integrated, deep-learning-driven framework that can learn rich representations from diverse datasets while ensuring stable and scalable predictions across varying regions and seasons.

In response to these limitations, this research seeks to construct a deep-learning–driven hybrid ensemble framework that integrates the hierarchical representation capabilities of Deep Belief Networks (DBNs) with the decision robustness offered by ensemble classifiers, including Random Forest, J48 Decision Tree, and Naïve Bayes. The intention is to build a predictive system that enhances accuracy while ensuring consistent performance across India’s varied agricultural environments. Through the adoption of a refined two-stage preprocessing strategy and the use of deep feature extraction, the proposed model aims to surpass the shortcomings of existing yield-prediction techniques and deliver a dependable, scalable, and data-centric solution tailored for Indian farming conditions.

Based on the gaps found in earlier crop-yield prediction research, this study puts forward several important contributions, which are summarised below:

1. Hybrid deep ensemble architecture: The Robust Lemuria Framework combines a Deep Belief Network (DBN) for layered feature extraction with an ensemble of classifiers, namely Random Forest, J48 Decision Tree, and Naïve Bayes. This integrated structure improves prediction reliability by uniting deep representation learning with complementary classification strategies.

2. Two-stage optimized preprocessing pipeline: This work proposes a structured preprocessing approach that systematically addresses data noise, missing values, and feature normalization, followed by non-linear dimensionality reduction using a DBN. These steps improve data consistency and ensure that the model is trained on informative and well-conditioned features.

3. Comprehensive evaluation using a decade-long, multi-regional dataset: The model is validated on diverse agricultural data from 2010 to 2020, covering multiple crops, seasons, and climatic zones across India. This establishes the scalability and robustness of the framework.

4. Performance superiority over existing models: Extensive comparisons with classical machine learning algorithms and Comparative results indicate that the proposed approach delivers notably improved accuracy, sensitivity, specificity, and regression performance when evaluated against recent hybrid models.

5. Practical relevance for data-driven agriculture: The framework is designed to support policy planning, resource allocation, and real-time advisory platforms, supporting more informed farming decisions and promoting sustainable agricultural practices.

The motivation behind this work lies in the increasing need for intelligent, data-driven tools to support sustainable agricultural practices in India, a country where farming contributes significantly to GDP and livelihoods. Accurate yield forecasting not only aids farmers in crop planning and input management but also supports governmental policy formulation in areas such as subsidy allocation and food distribution. By leveraging historical data from 2010 to 2020 across multiple crops and regions, the proposed framework is designed to be scalable, adaptable, and practical for real-world deployment in agricultural decision support systems.

To clearly establish the novelty of the proposed RLF, it is essential to highlight how it advances beyond existing methodologies and what unique contributions it brings to the domain of crop yield prediction. Unlike earlier works that typically relied on singular or conventional machine learning models such as Decision Trees, KNN, or even the earlier. The adaptive Lemuria framework combines a variety of DBNs with RF, Naive Baynes, and Decision Trees to provide a comprehensive classification system. This combination allows for both deep hierarchical feature extraction and diverse model consensus, yielding superior predictive performance. Furthermore, it leverages pre-training techniques within the DBN to address data noise and dimensional complexity, which is not commonly emphasized in prior approaches. The framework also introduces an optimized two-stage preprocessing pipeline that enhances data quality before model training, ensuring robustness across varied agricultural conditions.

Additionally, comparative analysis in the paper demonstrates significantly improved In the novelty of the RLF lies in its multi-model fusion architecture, deep feature learning capability, optimized preprocessing strategy, and demonstrated scalability and precision in real-world crop prediction scenarios-distinguishing it clearly from existing literature and making a meaningful contribution to precision agriculture.

Related works

Recent research in crop-yield forecasting has explored a wide range of machine learning and deep learning techniques, with each approach addressing specific challenges associated with agricultural data analysis. These works can generally be grouped into three main categories. The first category includes ensemble and boosting techniques, such as those developed by Shahhosseini et al.²¹ and Keerthana et al.²², which improve accuracy by combining several models. However, these methods depend mainly on shallow feature learning and do not capture deeper patterns in the data, which limits their ability to perform well across different climates and soil types.

A second group of studies focuses on ANN- and LSTM-based deep learning models, such as the work by Khosla et al.²³ and Jhajharia et al.²⁸. Although these models capture non-linear patterns effectively, they often suffer from overfitting, insufficient preprocessing, and limited robustness when exposed to noisy or incomplete agricultural datasets. These limitations reduce their ability to perform reliably across multiple regions or farming scenarios.

The third category includes research that utilises regional or limited datasets, including the work by Kamath et al.²⁴. While these studies offer useful insights for specific districts or crop varieties, their narrow geographic and temporal scope restricts scalability. Consequently, such models struggle to generalise when applied to larger and more diverse agricultural environments like India²⁵.

Collectively, these thematic gaps highlight the need for an integrated framework that combines deep feature abstraction, robust ensemble decision-making, and optimized preprocessing. This motivates the development of the proposed Robust Lemuria Framework, which aims to overcome the limitations observed across ensemble, deep learning, and region-specific approaches (Table 1).

Table 1.

Comparative summary of existing crop yield prediction Methods.

Author(s), year	Dataset	Method(s)	Advantages	Limitations	Research gap addressed by this work
Shahhosseini et al. (2021)21	US corn yield + APSIM crop simulation data	Ensemble models (LASSO, XGBoost, linear models)	Improved yield prediction with soil moisture and phenology variables	Focused only on APSIM variables; no deep learning or preprocessing	Lack of hybrid deep learning + simulation model integration
Khosla et al. (2020)23	The information above pertains to the crop yield and rainfall data collected from Visakhapatnam	ANN + rainfall prediction	Region-specific crop prediction using limited variables	Used only weather attributes; lacked preprocessing and spatial generalization	No use of hybrid or ensemble DL techniques; limited features
Kamath et al. (2021)24	Regional yield datasets (India)	Multiple ML algorithms (including LR)	ML use for practical agricultural advice	Scale limitations (100-acre implementation not feasible)	Incomplete exploration of ensemble/hybrid DL models
Keerthana et al. (2021)22	Crop yield datasets (unspecified)	Decision Tree + AdaBoost Regressor	Boosted weak learners for better accuracy	Weak base learner; lacks complex feature engineering	No integration of deep feature learning or DBNs
Morales and Villalobos (2023)26	US agricultural data	RF, Lasso, Ridge, ANN	Comparative baseline against ANN	ANN underperformed; limited generalization	No use of hybrid deep ensemble DL architecture
Jhajharia et al. (2023)28	Rajasthan agricultural dataset	RF, SVM, Lasso, LSTM	RF achieved R2 = 0.963	No feature pre-training; weak LSTM performance	Lacks robust preprocessing and deep-ensemble hybridization
Meghraoui et al. (2024)29	Various DL-based crop prediction studies	Systematic Review (CNN, LSTM)	Identifies trends in CNN/LSTM use	Limited crop variety; weak data preprocessing focus	Need for DBN and multi-stage preprocessing integration
Oikonomidis et al. (2023)30	44 DL studies	Systematic Review (CNN, RNN, LSTM, DNN)	Comprehensive meta-review	Overfitting, poor interpretability, narrow crop focus	Lack of generalizable, interpretable, hybrid DL models

Proposed methodology

Climate factors affect the Production (MT) of agriculture. These include the temperature, wind speed, precipitation, soil parameters, and organic carbon. Apart from these factors, crop production is also influenced by the quality of water available and the condition of the land on which the crops are grown. Due to the continuously changing climate conditions, the Production (MT) of crops has become affected. This study focuses on developing a reliable predictive framework capable of analysing soil-series data and recommending suitable crops across different regions of India. Machine learning techniques are employed to examine and interpret key chemical characteristics of various soil types, while geographical attributes are used to associate each soil series with appropriate crop options. By incorporating a deep neural network, the proposed approach achieves improved robustness and computational efficiency. The overall structure of the Robust Lemuria Framework is illustrated in Fig. 1. Experimental results demonstrate that the proposed model delivers more accurate yield forecasts than existing methods.

Fig. 1.

Robust Lemuria Framework.

A single predictive model often struggles to capture the full complexity of agricultural datasets, which typically contain non-linear patterns, seasonal fluctuations, and region-specific variations. By combining Deep Belief Networks (DBNs) with Decision Tree–based classifiers, the proposed framework leverages the complementary strengths of both approaches. DBNs are highly effective in learning deep, latent, non-linear relationships from high-dimensional climate and soil data, enabling the extraction of meaningful hierarchical features. In contrast, Decision Trees and ensemble methods like Random Forest and J48 are effective at forming clear decision rules, managing different types of input variables, and lowering classification errors by using structured, rule-based splits.

When these techniques are integrated, the DBN first transforms raw agricultural variables into a more discriminative feature space, after which tree-based classifiers make robust and interpretable decisions using these refined representations. This hybridization reduces noise sensitivity, mitigates overfitting, improves generalisation across diverse regions, and delivers significantly higher predictive stability compared to using any single model alone. The synergy between deep feature learning and ensemble decision-making ultimately forms the core advantage of the proposed Robust Lemuria Framework.

Dataset description

Various sources of data have been collected for forecasting rainfall. These include the Indian Meteorological Department, the Department of Agriculture, and the Statistical Institute. The data collected from these departments are used to make informed decisions regarding the development of rainfall. However, they may also contain additional information that could lead to inaccurate predictions. Before starting a research project, the data collected by these agencies are preprocessed to extract the necessary features for the study. These include the maximum and minimum temperatures, moisture, and rainfall. For the current research, we have taken into account the data collected from 2008 to 2017. Although the data collected by these agencies contains a large number of parameters, we need to select the right ones for our study.

The dataset comprises agricultural records collected over a 10-year period, including monthly values for rainfall, temperature, yield, Production (MT), and area(ha) under cultivation for major crops such as rice, wheat, and maize. After preprocessing, a total of a total of approximately 12,000 data points were retained for model training and testing. These entries span multiple seasons and geographical regions across India, ensuring that the model was trained on diverse and representative data. The area under cultivation is reported in hectares (ha), and crop production is measured in metric tonnes (MT).

• The optimal season for different crops, such as wheat, rice, and maize, was determined by analyzing the data on India’s monthly rainfall.

• The mean temperature of various crops, such as wheat, maize, and rice, was analyzed to determine their optimal season in India.

• The total area(ha) of cultivated land for these crops in India was considered for the study. For the Kharif and Rabi seasons, the total area(ha) of land was calculated.

• An area(ha) under irrigation for these crops was calculated for the various seasons in India. This detail was used to determine the optimal season for these crops.

• For the study, the total Production (MT) of these crops was estimated to be around 1 million tons for the Kharif and Rabi seasons.

• The objective of this study is to predict the yield outcomes of different agricultural crops. Such as wheat, rice, and maize, using the area under cultivation and production figures during the Rabi and Kharif seasons.

Although the dataset used in this study was generally balanced across primary crop classes, some imbalances were observed in specific regions and less frequently cultivated crop types. To ensure fair model evaluation, we adopted stratified sampling during the training-test split process to maintain class proportions. Furthermore, rather than relying solely on accuracy, We used the F-measure, Recall, and Precision metrics to analyze the model. These measures help us evaluate how well the model performs when dealing with uneven class distributions. These measures also help maintain the model’s accuracy in predicting less common crop types, preventing the majority class from dominating the results. For future work, more sophisticated techniques for handling class imbalance—such as SMOTE, cost-sensitive approaches, or ensemble resampling methods—could be explored to further enhance prediction performance, especially for underrepresented crop categories.

The dataset employed in this study covers a broad spectrum of agricultural conditions across India, ensuring that there is adequate variety in the data to properly train and evaluate the proposed model. It consists of ten years of crop records (2010–2020) collected from multiple states, spanning varied agro-climatic zones such as humid coastal belts, semi-arid plains, irrigated regions, and rain-fed agricultural systems. The dataset includes major staple crops—rice, wheat, maize, sugarcane, coconut, and others—representing both Kharif and Rabi seasons. In addition, the data incorporate multiple environmental parameters, including rainfall patterns, temperature variations, soil characteristics, irrigation levels, and production metrics.

This wide-ranging dataset helps the model understand how climate, soil conditions, and crop responses are connected across different types of regions. While some states and minor crops show natural variations in data density, the overall dataset is sufficiently comprehensive to support robust generalisation. Furthermore, stratified sampling was employed during model splitting to preserve class balance and ensure fair representation of all crop categories. This diversity strengthens the reliability of the experimental outcomes and validates the framework’s ability to perform accurately across India’s varied agricultural landscapes.

Preprocessing

The data set’s preprocessing step is carried out before it is analyzed and transformed into a suitable forecasting model. This stage can also help address several other problems present in the dataset. An important part of this process involves using a neural network model to identify and extract useful patterns and features from the dataset. The proposed RL also eliminates missing data due to various reasons. It then applied various accusation methods to the data to obtain the best possible forecasts. The environmental and harvest data sets were also comprehensive. These data sets were also prepared without missing information.

Importance of two-stage data cleaning

Reliable crop-yield forecasting depends on clean and well-prepared input data, as agricultural datasets commonly include missing entries, recording errors, seasonal variations, and differences across regions. The proposed two-stage data cleaning process strengthens the robustness of the model by improving data reliability at two critical levels.

In Stage 1, the focus is on data quality enhancement, where missing values are systematically reconstructed, outliers caused by sensor malfunction or reporting errors are removed, and numerical attributes are normalized to a uniform scale. This ensures that the model is not adversely influenced by noise, extreme values, or inconsistent measurement units.

In Stage 2, the cleaned data undergo DBN-based non-linear feature refinement, where redundant, weakly correlated, or noisy features are suppressed. The DBN learns deep latent structures within the data, enabling the extraction of compact and highly informative feature representations. This reduces dimensional complexity, prevents overfitting, and enhances the discriminative power of the downstream classifiers.

Together, these two stages ensure that the input reaching the hybrid model is both statistically reliable and structurally meaningful, which significantly enhances prediction accuracy, stability, and generalisation across diverse agricultural conditions.

Deep belief network

The Deep Belief Network (DBN) in the proposed framework is employed primarily for deep feature extraction and non-linear dimensionality reduction, rather than for rainfall forecasting as implied in earlier drafts. Agricultural datasets typically contain heterogeneous variables—such as climate indicators, soil characteristics, and crop production parameters—that exhibit complex, non-linear interactions. The DBN serves as a hierarchical feature-learning mechanism that transforms these raw, high-dimensional inputs into a compact and discriminative representation. By learning latent patterns across multiple layers, the DBN effectively reduces noise, identifies meaningful feature dependencies, and enhances the quality of inputs supplied to the ensemble classifiers. This deep feature abstraction is essential for improving model accuracy and generalisation, especially when dealing with diverse agricultural conditions across India.

The weather rainfall forecast is an essential component of the harvest forecast, since it is based on the observed climate conditions. This section explains the key steps involved in preparing the rainfall-related dataset for model training. If the dataset is not properly processed, it may contain excessive noise, which can negatively affect the model’s learning capability. The pre-training stage of the DBN algorithm plays an important role in organizing and refining the rainfall data. It also supports the efficient development of the deep network structure by enabling faster and more stable learning.

Because the dataset contains complex and high-dimensional information, training the DBN model naturally requires more time. To address this, a pre-training strategy was introduced, which helps the network learn the underlying patterns more efficiently by considering factors such as data size, density, and variability. This approach significantly reduces overall training time while improving learning stability.The proposed DBN algorithm is performed in a hierarchical structure. This project involves the use of a pre-trained data set and various hidden elements. The hidden unit is one of the components that’s supposed to be utilized in training. All of the elements in the set are associated with the visible element set m. The weights learned by the hidden units are then used to compute the probability distribution associated with the visible layer’s state vectors. This process helps the model understand how the input features relate to one another. Finally, the two main components of the network are defined based on these learned relationships.

1

The values presented in this table indicate the state of the visible m_i and the hidden units q_j, their respective weights, and the relationship between them. The terms b_i and a_j denote the terms used to describe these elements, and the table shows the possibility that rainfall data could be presented to a visible vector.

2

The data collected by a weather station do not contain visible or hidden associations. The information is also factorial and can be specified by the various distributions.

3

The learning curve represents the process through which the model gains the information required to learn patterns within the rainfall data.

4

5

Furthermore, u(f|m) can be calculated precisely using:

6

The u(f)-m approach is applied to evaluate the behaviour of the hidden units, taking into account their objective contribution to the learning process. This helps the model operate more efficiently. In this structure, the output generated by one layer becomes the input for the next, allowing the learned features to be gradually revealedClustering techniques are applied to estimate crop-yield outcomes by analysing key influencing factors such as climatic conditions and crop performance indicators. By identifying meaningful groupings within the data, this approach improves prediction accuracy. The overall training process of the proposed Lemuria algorithm is presented in Fig. 2.

Fig. 2.

Pre-training and finetuning.

The output data of each hierarchy is shown in Fig. 2 as hidden layers. The pre-trained data was initially used to input the data for this process, and the result of this process is then used for the next layer. A joint distribution is shown as an unseen layer in h1. It is represented with an equation that shows its energy function and probability distributions. It then tries to generate various samples from these distributions by applying multiple equations. The output is a training sample that shows the conditional distributions that the pre-trained data can provide. These distributions are then analyzed using multiple equations. After generating the probability values, the next step is to apply the multiple equations to the joint distribution to train it. This process stops at the third layer, because it seems to be unable to learn properly when the machine is elevated to something higher. The association strengths are able to be changed by the elements being associated with them, which is a result of the so-called discrepancy trap. The noise generated by this process leads to the association’s arbitrary step until the achievements are satisfied. After the third layer finishes its pre-training, it trains the model using the rainfall data collected from the third layer. This method can help reduce the time it takes for accurate predictions to be made.

Training phase

The training phase of this project involved the use of cluster analysis to train the pre-trained rainfall data. In order to produce tighter clusters, we implement a prototype-based method. The data are divided into k number of clusters according to their respective rainfall values. The objective of this approach is to measure how well each data point belongs to its assigned cluster and how it differs from others. This is done by calculating the squared distance between individual rainfall values and the corresponding cluster centre, which helps assess the strength of cluster membership.

7

The number of samples required to be collected when forming clusters is shown. The hidden and visible unit values are also included. The cluster’s velocity is determined by taking into account its inertia weight.

Random forest

The second generation of tree predictors is a generalization that takes into account the random vector’s profit and the distribution of all trees in a forest. It shows that the tree’s characteristics are similar to those of the other trees in the forest. As the tree grows, the generalization error increases as well. The main reason for this is that the individual strengths of a tree are related to the correlation between them. The RF employs sequences of inputs to improve its accuracy. It also randomly selects the inputs at each node. It then constructs the random features at the nodes randomly. The results of this method are then used to decrease the correlation between the various inputs and the tree’s characteristics. The RF uses the same procedure as the one used in the CART approach. It also applies the same strategy to the development of the tree, except that instead of selecting random features, it chooses the ones that are most appropriate for the task. For instance, if the training dataset has a substitution feature, the tree is grown using the random feature selection method.

The advantages of using a bag are its ability to increase the accuracy and its flexibility. It also provides a method to estimate the correlation coefficient and the ensemble tree combination. In addition, it allows the user to perform the assessment in-bag. The main advantage of out-of-bag estimation is that it takes into account the various components of a sequence, such as the number of classifiers. This method can help reduce the error rate whenever the statistics of the sequence change. It is also useful to pass on the merging of the errors to the other regions. One of the biggest disadvantages of carrying a bag is its high likelihood of being subjected to bias. In the cross-validation process, the degree of bias is unknown. Out-of-bag method is free from bias. The RF applies a third of the data and tests a third of the data. Unlike a bag, out-of-bag method does not require the training data to be performed by the RF. This method achieves faster performance and is more accurate than other decision tree methods. Also, having multiple tree constructions helps the RF perform well with additional tree framework.

Naive Bayes

The Naive Bayes classifier is related to the Bayes theory. It has a probabilistic nature.It also has independence between its features. This type of analysis is commonly used in mathematical terms to represent the knowledge.

8

Where R and S are events and P(R) and P(S) are functions.

The posterior probability of an event R is known as P(R). Similarly, the probability of an event S is called P(S/R). The likelihood of an event happening is referred to as P(S/R). This is the implementation of the Naive Bayes version. The nominal feature probabilities of an event are approximated from the data distribution and the Gaussian distribution. The most prominent class for an event is predicted by the Bayes rule or the classifier. The Naive Bayes is very easy to understand, as its log probabilities are applied automatically.

The Naive Bayes classifier can provide high accuracy. When the redundant features are removed, the performance of the system improves. Moreover, when the data dependencies are minimal, the performance of the system is exceptional. So a minimal execution time is required to train the data.

J48(C4.5) decision tree generator

This study makes use of the C4.5 decision tree algorithm, an improved and more advanced version of the earlier ID3 method. It is constructed using the Weka library’s J48 algorithm. The main method used in the construction of the trees is to find the feature sets having distinct characteristics. This method helps in performing the testing of the features on the root node of a tree. The branches and features of a tree are related to its values. To classify a new instance, one needs to look for the features that are related to the values that are observed in an instance. Doing so allows one to perform the testing on the features that are closest to the class’s values.

The C4.5 algorithm uses the greedy approach to construct decision trees. It takes into account the various information-theoretic estimates and then divides the resulting tree root into subsets. For instance, if there is a small amount of entropy among the labels of a class in a subset corresponding to the labels of the class in the training dataset, the information-theoretic estimates are taken into account to gain the information. The C4.5 algorithm then generates a selection tree that takes into account the various attributes of the root attribute. The goal of the gain ratio principle is to ensure that the various attributes have an average or better gain.

The algorithm iteratively constructs subtrees using the C4.5 framework. It then finds the likely subset of these subtrees that contains a distinct class. The key distinction between ID3 and C4.5 lies in the improvements introduced in C4.5. While ID3 builds decision trees using information gain alone, C4.5 enhances this process by using gain ratio, handling continuous attributes, and managing missing values more effectively.

By implementing the C4.5 algorithm, the goal is to reduce the over fitting of training data. The C4.5 algorithm performs a variety of tasks, such as identifying the best leaf and performing a succession of nodes based on the resubstitution errors. It is considered to be an efficient algorithm when the training data is evaluated. Also, it is fast and accurate when the knowledge is brought in. The main advantage of using C4.5 is that it allows the system to perform various tasks with minimal effort. The C4.5 algorithm also performs well when it comes to selecting the appropriate feature subsets from the redundant and relevant data.

Hyperparameter settings

To ensure model reproducibility and transparency, the hyperparameters used for the DBN and ensemble classifiers in the Robust Lemuria Framework are summarized in Table 2.

Table 2.

Season wise crop prediction.

Season	Production (MT)
Autumn	6.441377e+07
Kharif	4.029970e+09
Rabi	2.051688e+09
Summer	1.706579e+08
Whole year	1.344248e+11
Winter	4.345498e+08

Model	Hyperparameter	Value
DBN	Hidden layers	3
	Neurons per layer	[128, 64, 32]
	Learning rate	0.01
	Batch size	64
	Epochs	100
Random Forest	Number of trees	100
	Maximum depth	None (fully grown)
	Split criterion	Gini index
Naive Bayes	Distribution type	Gaussian
	Smoothing parameter (α)	1.0
J48 Decision Tree	Confidence factor for pruning	0.25
	Min instances per leaf	2

The hyperparameter values were chosen through experimental tuning and guidance from earlier studies, allowing the model to achieve a suitable balance between accuracy and computational cost.

Result and discussion

Simulation environment

We have used the Python programming language to implement the data processing codes. We have also used the Anaconda Spyder integrated development environment, This is a powerful tool for data scientists. We are using a Mac with a Core i3 and 4 GB of RAM.

Input description

Continuous training and improvement of machine learning models make it possible to build systems that can produce dependable future predictions. Crop-yield forecasting is especially important for addressing issues related to climate change and food security. Reliable estimates require ongoing advancements in technology and careful analysis of factors such as climate trends, soil quality, and crop condition. Having access to accurate historical yield records supports farmers and policymakers in making informed and effective decisions. The increasing popularity of precision agriculture has led to the development of a vast amount of data that can help the farmers make informed decisions. This dataset was built using the latest weather and climate data collected in India. Figure 3 shows the sample dataset taken for our research work. Figure 4 shows the Pictorial representation of area(ha) harvested and obtained Production (MT) for the year 2010–2020. Figure 5 shows the Production (MT)s for each state.

Fig. 3.

Example dataset showing seasonal patterns and corresponding crop yield values.

Fig. 4.

Pictorial representation of area (ha) harvested and obtained Production (MT) for the year 2010–2020.

Fig. 5.

Agricultural production (MT) for each state.

Prediction results

Crop yields are influenced by several elements, including farming practices and environmental conditions. Even small changes in climate can lead to significant differences in yield from one year to the next. Accurate yield forecasting can help increase global food production by enabling better planning. With the help of prediction tools, farmers and policymakers can make smarter decisions about imports, exports, and market strategies. These tools can also be used to estimate how newly developed hybrid varieties may perform under different conditions.

Here we showed the predicted results for the overall prediction and crops like rice, coconut and sugarcane. Figure 6 shows the season based outcome and Fig. 7 shows the corresponding plot based on the area(ha) harvested and prediction obtained. Figure 6b shows that, rice will be harvested for the whole year and yield high Production (MT) on winter. Figure 6c shows that, coconut Production (MT) is high except in kharif season. Sugarcane yield is shown in Fig. 6d, its Production (MT) is limited to autumn, rabi and summer seasons.

Fig. 6.

Season based crop prediction outcome (a) Overall (b) rice (c) coconut (d) sugarcane.

Fig. 7.

Crop prediction outcome (a) Overall (b) Rice (c) Coconut (d) Sugarcane.

The Bar plot (Fig. 8) shows the estimated effects of various weather components on corn grown for 52 weeks each year. The vertical axes were adjusted to make the effects comparable across all of the weather components.

Fig. 8.

Corn prediction result based on 6 parameters.

Figure 9 shows the estimated impact of various weather elements on the soybean crop during the 52 weeks of the season. The vertical axes have been adjusted to make the comparisons more comparable. In addition, the effects of different elements on the soil are shown in Figs. 8 and 9. The normalized effects of various weather elements were shown in each group. The management practices, soil conditions, and weather components were then adjusted to make them more comparable.Tables 2 and 3 shows the season wise crop prediction and crop type based predictions respectively. From Table 3, most of the crops yield can be harvested for the whole year. From Table 2, coconut yield high Production (MT) among all crops. Figure 10 shows the future crop prediction outcomes for the upcoming years ranges from 2023 to 2028.

Fig. 9.

Soybean prediction result based on 6 parameters.

Table 3.

Crop prediction based on crop type.

Crop	Production (MT)
Coconut	1.299816e+11
Sugarcane	5.535682e+09
Rice	1.605470e+09
Wheat	1.332826e+09
Potato	4.248263e+08

Fig. 10.

Future crop predictions for the year 2023–2028.

Figure 10 presents a multi-year comparative analysis of crop production across various crop types from 2023 to 2028. While the figure offers valuable insights into production trends, its readability can be further enhanced through improved labeling. In particular, the X-axis labeled “Crops” would benefit from clearer orientation and font sizing, while the Y-axis should explicitly state the unit of measurement as “Production in Tonnes” for clarity. Additionally, distinguishing actual and predicted values—especially where forecasting is applied—using labels such as “Actual Yield (MT/ha)” and “Predicted Yield (MT/ha)” would improve interpretability. Color-coded bars for different crops or yield types (e.g., rice, maize, soybean) could further assist in visual differentiation. These enhancements would ensure the figure is not only informative but also accessible to a wider academic and policy-oriented audience.

The regression-based performance of the proposed framework for crop-yield forecasting was assessed using widely accepted evaluation metrics. Figure 11 illustrates the results obtained using four measures: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). The proposed framework’s RMSE score was the highest among the evaluated values. This is because it is sensitive to large errors when dealing with the residuals’ squaring. Nonetheless, this value remains relatively low given the scale of agricultural Production (MT) data, indicating the model’s strong predictive capacity. An MAE of 4.69 shows that the predicted values are, on average, closely aligned with the actual crop yields, indicating consistent and reliable model performance. The MAPE value of 3.88% further confirms the high accuracy of the predictions, suggesting that the model’s estimates deviate by less than 4% from the true yield values on average—well within acceptable bounds for agricultural forecasting systems. Most notably, the R² Score was calculated as 0.9994, The model accounts for more than 99% of the variation in actual yield values, indicating an exceptionally strong fit. This high level of accuracy supports the effectiveness of the proposed hybrid approach, which integrates DBNs for deep feature abstraction and an ensemble of classifiers to enhance predictive performance. Overall, this performance analysis affirms that the RLF offers both low error rates and high generalization, rendering it highly suitable for practical deployment in crop Production (MT) prediction across diverse soil and regional conditions in India.

Fig. 11.

Regression performance metrics of the proposed Robust Lemuria Framework.

Performance analysis

The performance of the Robust Lemuria algorithm is evaluated using key classification metrics such as accuracy, sensitivity, and specificity, which together provide a comprehensive assessment of the model’s predictive capability.

Accuracy

Accuracy refers to the proportion of correctly classified instances among all evaluated samples. It provides a general measure of the model’s effectiveness during the testing phase by indicating how reliably it predicts the correct outcomes.

9

Where TP denotes correctly predicted positive cases, TN represents correctly predicted negative cases, FP indicates cases incorrectly classified as positive, and FN refers to cases incorrectly classified as negative.

Sensitivity (recall)

Sensitivity indicates how well the model can correctly identify actual positive cases. It is determined by comparing the number of correctly predicted positive instances with the total number of positive observations.

10

Specificity

Specificity indicates how effectively the model correctly recognises negative cases. It takes into account the fraction of accurate predictions made from all observations.

11

F-measure

The F-measure combines precision and recall into a single harmonic value, making it especially useful for evaluating model performance on imbalanced datasets. This metric relies on precision as a key component for assessing classification quality.

12

Error rate

The error rate is a measure of how frequently the algorithm makes wrong predictions. A lower error rate means that the data is more reliable. In this paper, we present a robust algorithm that aims to reduce the error rate to 0.0251. This method gives a reliable prediction.

13

Regression evaluation metrics

To evaluate the regression performance of the Robust Lemuria Framework, several standard error and fit metrics were employed:

• Mean Absolute Error (MAE): MAE represents the average difference between the predicted and actual values. Lower MAE indicates that the model’s predictions closely match real outcomes.

• Root Mean Squared Error (RMSE): RMSE measures the square root of the average squared difference between predicted and actual values. Because larger errors are penalized more heavily, this metric highlights significant deviations.

• Mean Absolute Percentage Error (MAPE): MAPE expresses prediction accuracy in percentage terms by averaging the relative errors. It provides an easy-to-interpret measure of how close the forecasts are to actual values.

• R² Score (Coefficient of Determination): The R² value shows how much of the variance in the actual data is explained by the model. A value closer to 1 indicates a strong fit and high predictive accuracy.

Figure 12 illustrates the performance metrics achieved by the proposed Lemuria framework.These include sensitivity, specificity, accuracy, and F-measure.The updated graph addresses previous concerns regarding readability by incorporating explicit axis labels, percentage units, and data markers. The Y-axis clearly indicates the metric values in percentage (%), and each point is annotated with its corresponding value for precise interpretation. The graph reveals that the model achieves the highest performance in Specificity (99.35%), followed by Accuracy (98.99%), Sensitivity (98.54%), and F-measure (97.89%). This suggests that the framework is particularly effective at correctly identifying negative classes (non-relevant predictions) while maintaining a high true positive rate and balanced overall performance. The inclusion of grid lines and metric labels enhances visual clarity, ensuring the figure communicates insights in a scientifically rigorous and accessible manner.

Fig. 12.

Performance evaluation based on accuracy, sensitivity, specificity, and F-measure.

Figure 13 illustrates the key performance indicators of the Robust Lemuria Framework. While the earlier version of the figure conveyed useful information about the model’s accuracy and efficiency, the updated version offers much better clarity. The axes are clearly labeled, making it easier to distinguish performance scores from error measures. Numerical values are also displayed above each point, allowing readers to interpret the results quickly without guessing. This clear presentation draws attention to the very low error rate (0.0251) and MSE (0.0214), which demonstrate strong predictive accuracy, as well as the efficient training time (1.256 s) and testing time (2.231 s). Overall, the improved visual design enhances readability and follows good scientific graphing practices.

Fig. 13.

Performance evaluation based on training time, testing time, error rate, and MSE.

The computational complexity of the Robust Lemuria Framework arises from its hybrid architecture, which combines DBNs for feature learning with an ensemble of classifiers (RF, NB, and J48). DBNs involve several stages of unsupervised pre-training and later fine-tuning, which naturally increases their training time compared to simpler, shallow models. However, once trained, inference is relatively fast. The ensemble classification component balances computational load by combining low-complexity algorithms (NB and DT) with the more computationally intensive RF. According to the evaluation, the framework can run smoothly with a training duration of 1.26 s and a test duration of 2.231. While the model is more demanding compared to simpler architectures, these results show that it can run smoothly, its performance gain-particularly in predictive accuracy and robustness-justifies the added cost. Moreover, the framework remains scalable and suitable for deployment in agricultural decision support systems with moderate computing resources.

Figure 14 compares the crop-yield values predicted by the Robust Lemuria Framework with the actual observed yields. The scatter plot shows that most points are closely aligned with the diagonal line, indicating strong agreement between predicted and real values. This close alignment highlights the effectiveness of the multi-stage preprocessing strategy and the DBN-based feature extraction, which together help the model learn complex and non-linear patterns present in agricultural data.

Fig. 14.

Predicted vs. actual yield and error distribution for the Robust Lemuria Framework.

The axes have been clearly labeled as Actual (X-axis) and Predicted (Y-axis); however, specifying the units—metric tonnes (MT)—would further improve clarity and allow easier replication of results. Including a reference line (y = x) could also enhance visual interpretation by showing the ideal prediction pattern.

The consistent clustering of points near the diagonal confirms that the model is performing optimally, without signs of overfitting or underfitting. These results demonstrate that the Robust Lemuria Framework can reliably generalize across different crops and conditions, making it a strong candidate for practical crop-yield prediction applications in India.

Comparison analysis

We compare the proposed robust version of the Lemuria algorithm with the existing methodologies. These include the SVM, CLARA, KNN, PAM, and Adaptive Lemuria.

Table 5 compares the performance of the proposed Lemuria framework with several commonly used machine learning methods, including SVM, CLARA, KNN, DBSCAN, and Adaptive Lemuria. The results indicate that the proposed approach delivers more reliable and accurate predictions for crop-yield estimation than the existing techniques. This comparison is based on important evaluation metrics such as precision, recall, and F-measure, which collectively demonstrate the improved effectiveness of the Lemuria framework. All models compared in Table 5 were implemented and evaluated under identical experimental conditions. The same dataset (2010–2020 agricultural data), stratified 70/15/15 split (training/validation/testing), preprocessing steps (missing value imputation, noise removal, normalization), and feature set were applied across all models. This ensures that the reported differences in precision, recall, and F-measure reflect genuine algorithmic performance and not discrepancies in data handling.

Table 5.

Performance comparison of Robust Lemuria with existing classifier.

Method	Precision	Recall	F-measure
CLARA28	60.6	64.3	61.1
SVM29	45.7	66.6	54.2
Decision tree (DT)29	81.4	91.6	86.2
DBSCAN28	84	81.5	82.2
KNN29	91.6	91.6	91.6
K-means28	94	92.3	91.5
PAM28	96.1	95.7	95
Adaptive Lemuria30	97	96.2	96.49
Proposed Robust Lemuria	98.21	97.23	97.89

The Table 5; Fig. 15 shows the performance comparison between different approaches when it comes to handling large datasets. The data collected by these systems are compared with the same number of values. The goal of this analysis was to find out if the proposed Lemuria framework could perform well against various machine learning classifiers, including CLARA, SVM, Decision Tree, DBSCAN, KNN, K-Means, PAM, and Adaptive Lemuria. As illustrated in Fig. 15 and detailed in Table 4, the proposed framework achieves the highest values across all key metrics: Precision (98.21%), Recall (97.23%), and F-measure (97.89%). This significantly outperforms conventional models like SVM (Precision: 45.7%, Recall: 66.6%, F-measure: 54.2%) and even advanced clustering-based approaches such as PAM and K-means. While Adaptive Lemuria performs competitively (F-measure: 96.49%), Even so, its performance remains below that of the proposed model, particularly in terms of accuracy and recall. This reinforces the strength of the hybrid design introduced in this study, which combines DBN-based hierarchical feature extraction with an ensemble of classifiers—Random Forest, Naïve Bayes, and J48—supported by a structured two-stage preprocessing process. Together, these components enable the RLF to deliver higher prediction accuracy while maintaining robustness and scalability when used on complex agricultural datasets encountered in real-world conditions.

Fig. 15.

Performance comparison between the Robust Lemuria Framework and existing classification models.

Table 4.

Regression performance metrics.

Metric	Value
Mean absolute error (MAE)	4.69
Root mean squared error (RMSE)	22.20
Mean absolute percentage error (MAPE)	3.88%
R2 score	0.9994

The comparative results indicate that the Robust Lemuria Framework achieves better performance than many commonly applied machine learning and deep learning approaches for crop-yield prediction. Traditional models such as SVM, KNN, DBSCAN, and basic Decision Trees generally show lower accuracy, as they struggle to model the complex and non-linear relationships present in agricultural data. Even hybrid approaches such as Adaptive Lemuria and boosted decision-tree methods tend to lose effectiveness when dealing with diverse climate patterns and varying soil characteristics.

In contrast, the proposed model consistently records 98.99% accuracy, 98.54% sensitivity, 99.35% specificity, and an R² of 0.9994, outperforming all baseline algorithms evaluated in the study. The improvement is attributed to the DBN’s ability to learn deep, latent feature patterns and the ensemble classifiers’ strength in constructing robust decision boundaries. The comparative results presented in Table 5; Fig. 15 confirm that the hybrid deep–ensemble architecture not only yields higher predictive accuracy but also maintains stability across diverse crop types and seasons. This establishes the proposed framework as a significantly more reliable and scalable solution than existing models for large-scale agricultural prediction tasks.

The proposed lemuria algorithm is more reliable and accurate than the previous ones when it comes to reliability and accuracy. It takes into account different measures such as precision and recall, and it performs better than the current generation of approaches when it comes to these aspects. In addition, it has lower training time and error rate.

The implementation of the RLF can be done in the farming industry to enhance the precision and efficiency of operations. This is due to how the data collected by the various machines and sensors utilized in farming can contribute to the overall market price.The paper aims to help solve various farming problems and improve the efficiency of the country’s agriculture sector. By enabling more accurate crop yield forecasting, the RLF supports strategic planning in agriculture, including optimal resource allocation, timely harvesting, and reduction in input waste. These efficiencies not only enhance farmer profitability but also contribute to stabilizing food prices, improving food security, and guiding evidence-based policymaking-ultimately promoting sustainable economic growth in India’s agriculture-dependent regions.

Conclusion

Agriculture remains a cornerstone of India’s economic progress, playing a key role in ensuring national food security. Predicting crop yields accurately is essential, as production levels are influenced by many factors, including climatic conditions, historical trends, and the use of fertilizers and pesticides. Traditionally, yield estimation depended heavily on farmers’ experience and their understanding of local fields and crop behaviour. In contrast, this work presents a contemporary solution through the Robust Lemuria Framework, which uses machine learning methods to estimate the yields of major staple crops with higher precision. The model provides timely and dependable predictions, helping farmers apply fertilizers, nutrients, and other agricultural inputs more efficiently. As machine learning becomes increasingly integrated into agriculture, it supports better decision-making and contributes to improved crop productivity, representing a meaningful advancement in modern farm management practices.

Limitations and future scope

The proposed framework for crop yield prediction, which is known as the Robust Lemuria, exhibited strong performance in both the regression and classification tests. The model primarily relies on historical soil, crop, and climate data, which may not capture real-time environmental dynamics or policy changes. Additionally, the framework was validated using region-specific data from India, and its performance across other agro-ecological zones requires further exploration. The model’s scalability for large-scale deployment, especially in low-resource settings, remains an open challenge. Future research could focus on integrating remote sensing data, real-time weather feeds, and economic variables to enhance predictive robustness. Moreover, extending the model to support adaptive learning and dynamic crop recommendation systems would significantly increase its practical utility for agricultural stakeholders.

The proposed Robust Lemuria Framework is designed with a flexible and data-driven architecture, making it adaptable to agricultural systems beyond India. Since the model primarily relies on climatic variables, soil characteristics, and crop production indicators—parameters that are universally present in global farming ecosystems—the framework can be retrained and calibrated for other countries with minimal structural changes. The DBN component is capable of learning complex, region-specific patterns within any new dataset, while the ensemble classifiers exhibit strong generalisation even when farming practices, crop calendars, and agro-climatic conditions differ significantly.

However, effective deployment in another country requires the availability of reliable local datasets, including historical weather patterns, soil profiles, and crop records. Once such data are integrated, the model can capture region-specific dependencies and provide accurate yield predictions. This flexibility shows that the framework is not confined to Indian agriculture alone and can be scaled for use in other regions, supporting decision-making across different climatic conditions and farming practices worldwide.Agricultural systems are highly dynamic, with continuous changes in farming practices, crop varieties, irrigation strategies, and climate conditions. As a result, model performance can gradually decline if it is not updated to reflect evolving patterns. The proposed Robust Lemuria Framework is designed to be adaptable, but like all data-driven models, it relies on timely and representative input data. Significant shifts in rainfall distribution, temperature extremes, or the adoption of new agronomic techniques may alter historical yield relationships. To maintain accuracy, the model should be periodically retrained with the latest seasonal and annual data.

The DBN component is particularly effective in adjusting to new trends because it can relearn latent representations when exposed to updated datasets. Similarly, the ensemble classifiers adjust decision boundaries according to modified feature distributions. Therefore, while the framework is robust to moderate fluctuations, regular model updates—preferably annually or after major climatic or policy changes—are recommended to sustain high predictive performance. Incorporating real-time sensor data or satellite-based measurements in future versions can further enhance adaptability to rapidly changing agricultural environments.

Although this study concentrates on building a high-accuracy hybrid deep–ensemble model, future work may explore the use of advanced deep learning architectures, such as transformer encoder–decoder models, to better capture long-term relationships in crop growth and climate patterns. These architectures have shown promising results in recent agricultural research, particularly in attention-guided feature learning and multi-source data fusion. Furthermore, while the current model functions as a high-performing predictive system, it can be extended with light-weight explainability components—such as basic feature-importance scores or external model-agnostic tools—to provide transparency without altering the model’s core structure.

Future studies may also focus on creating a user-friendly web or mobile platform that allows farmers, agricultural officers, and policymakers to easily access and interact with the model’s predictions. Such application layers can improve accessibility and operational usefulness without modifying the underlying predictive framework. This direction aligns with emerging smart-farming systems that focus on user-friendly integration rather than architectural changes. Finally, although the DBN inherently performs non-linear feature abstraction, future studies may explore global–local feature selection strategies from recent literature to further enhance model interpretability and reduce computational load when working with larger datasets.

Policy recommendations

The proposed Robust Lemuria Framework can play a significant role in strengthening agricultural policy-making by offering data-driven insights into crop performance, regional productivity, and expected yield fluctuations. Accurate yield forecasts enable government agencies to plan resource allocation more efficiently, such as determining the optimal distribution of fertilizers, seeds, irrigation resources, and crop insurance coverage. By identifying regions likely to experience lower productivity due to climatic or soil constraints, policymakers can design targeted subsidy schemes to support vulnerable farming communities.

Additionally, the model’s predictive outputs can help agricultural departments monitor food supply trends, anticipate market shortages, and regulate procurement policies more effectively. Long-range yield forecasting can also assist strategic planning efforts under national initiatives such as the National Food Security Mission (NFSM) and the Pradhan Mantri Krishi Sinchai Yojana (PMKSY). Overall, the framework provides a scientific basis for designing equitable, region-specific, and climate-responsive policies, ensuring that government support reaches the areas where it is most needed.

Policy recommendations for the Indian agricultural sector

The study explored the use of data in Indian agriculture to improve decision-making processes. It identified several actionable suggestions.

1. Integrate Predictive Analytics into Agricultural Extension Services: The RLF can be adapted by Krishi Vigyan Kendras (KVKs) and state-level agricultural departments to offer timely and location-specific crop yield forecasts to farmers.

2. Support Region-Specific Crop Planning: Policymakers can use insights from soil-crop prediction models to design region-wise crop advisories based on suitability, thereby optimizing productivity and reducing crop failure risks.

3. Inform Subsidy and Resource Allocation Policies: Predictive modeling can help estimate future production needs and guide equitable distribution of fertilizers, irrigation resources, and subsidies based on expected crop outputs.

4. Promote Investment in Smart Agriculture Infrastructure: The study supports the development of digital infrastructure, including real-time weather monitoring and soil sensors, to further improve model accuracy and operational scalability.

5. Encourage Open Agricultural Data Initiatives: Ensuring access to updated and granular agricultural data will enhance the reliability of models like Robust Lemuria and support evidence-based policy formulation.

These recommendations are aligned with the objectives of India’s National Mission for Sustainable Agriculture (NMSA) and the Digital India initiative, contributing to long-term food security and climate resilience.

Author contributions

M. Tamilselvi - Conceptualization, Data curation, Formal analysis, Methodology, Validation, Writing – original draft.S. Vishnupriya - Supervision, Formal analysis, Writing – review & editing.K. Ushanandhini - Supervision, Conceptualization, Writing – review & editing.A. Dhanamathi - Supervision, Visualization, Writing – review & editing.

Data availability

The dataset used in this study is publicly available and can be accessed directly at the ICRISAT Data Repository: (http://data.icrisat.org/dld/src/crops.html). The data were utilised solely for research purposes to support the findings reported in this manuscript.

Declarations

Competing interests

The authors declare no competing interests.