We illustrate the working steps of the proposed crop production and recommendation paradigm in Figure 1 . The first stage is dataset preparation, which delivers a suitable data format for training and testing using the proposed ensemble learning and the existing investigated ML approaches after the necessary preprocessing and feature extraction. After that, the evaluation and analysis are performed based on the experimental results. Finally, the recommended system is presented for suggesting suitable crops for cultivating a specific region in the next season.

Top-down crop production prediction is depicted in the proposed methodology shown in Figure 1 . After collecting raw data, we structure it crop-by-crop for the five regular crops in our research, jointly with the environmental variables during each of the crops’ relevant months. To create a dataset suitable for ML training, we handle missing values, mitigate wrong format and wrong data and modify raw data as required. In keeping with the standard ML practice, we split the final dataset into training and testing segments for the purposes of model training and testing with various evaluation methods. The ML models are trained independently using 80% of the training data and evaluated using the remaining 20%. After training and testing several models, we tabulate the evaluation outcome in terms of MSE, RMSE, MAE, and R ². To determine the superiority of the proposed ensemble KRR, we conduct a DM significant test to compare it to the state-of-the-art benchmark ML models and determine its relative performance.

We collect the raw data samples from four different agricultural organizations in Bangladesh, which are BMD, BADC, BRRI, and BBS from 1969 to 2021 of different seasons, as shown in Table 4 . Kharif and Rabi are the two harvest seasons in Bangladesh for the majority of crops. The environment varies depending on the harvest season. The months responsible for crop production are considered when constructing the dataset for each crop. The specific crop’s weather information for the corresponding month is then provided. For example, Aus rice is harvested during the Kharif season, from June through August. It indicates that the monthly environmental data are regarded as weather data for Aus rice. Other crops’ samples are generated using the same manner, and the months corresponding to each season are listed in Table 4 . In the final dataset, the data samples include 7,000 records of five categories of crops (Aus rice, Aman rice, Boro rice, potato, and wheat), each having 1,400 samples of different districts of Bangladesh. In particular, we prepare eight attributes (rainfall, maximum temperature, minimum temperature, humidity, wind speed, cloud coverage, bright sunshine, production area, and production amount) from a total of 17 original attributes to predict the production of a particular crop, as shown in Table 3 . As the weather of the different crops is different for the month, we take the average of maximum temperature, minimum temperature, rainfall, humidity, wind speed, cloud coverage, and bright sunshine for each crop according to the month.

After generating the machine-learnable dataset, we split the dataset into training and testing sets. Then, we build the proposed ensemble ML (KRR) and investigated ML models (SVR, NB, CB, RF, and RR) using the training dataset and evaluate the trained models on the testing dataset to assemble the results. We consider four state-of-the-art performance indicators, i.e., mean absolute error (MAE), mean square error (MSE), root MSE (RMSE), and R ² score to evaluate the proposed and investigated ML models. All the experiments (dataset preprocessing, model training, model testing, and result processing) are accomplished using the Python programming language.

We preprocess the collected raw data to make them machine-learnable datasets. As all of the values of our dataset are numeric, we need not label any data during data preprocessing. Besides this, we normalize the data lastly to make it more trainable.

The preprocessing steps are as follows.

In this section, we recapitulate the working principles of the investigated ML and ensemble ML approaches (SVR, NB, RR, RF, and CB) and present our proposed ensemble ML paradigm (KRR). We select five benchmark ML models instead of all ML algorithms in a strategy. ML algorithms can be classified based on architecture and working procedure. We choose SVM as the representative of the distance-based ML algorithm, RR from the group of regularization ML algorithms, RF as the representative of the bagging ensemble algorithm, NB as the member of the Bayes theorem means probability-based ML algorithm, and CB as the representative of boosting ensemble algorithm. We select those five algorithms to represent all ML algorithms in our analysis, train them individually using our dataset, and measure their performances to compare with the proposed ensemble KRR.

The classical and ensemble ML algorithms and our proposed ensemble ML scheme are applied to predict crop production. The training data train these approaches, and the model learns the data sequences and then forms a prediction. The performance of the ML models is calculated using four evaluation metrics, i.e., MAE, MSE, RMSE, and R ². MSE can be defined as the absolute value of the difference between the predicted and actual value. Using MSE in regression will penalize large errors more than small ones if we assume that the target follows a normal distribution. The MSE is calculated as:

MAE indicates how big of an error we may expect on average from the prediction (Morales and Villalobos, 2023). MSE indicates how close it is to a set of points. It accomplishes this by squaring the distances between the points and the regression line (these distances are the errors). Squaring is required to eliminate any negative signs. The MAE is calculated as:

RMSE is the standard deviation of the residuals (prediction errors) (Glennie and Lichti, 2010). Residuals are a measure of how far the data points are from the regression line; RMSE is a measure of how to spread out these residuals. In other words, it reveals how strongly the data are aggregated around the line of best fit. The RMSE is calculated as:

R ² is a statistical measure of how much variation in a dependent variable can be explained by variation in the independent variables. The main objective of this score is to predict future results based on existing data. The extent to which the model can reproduce observed results is quantified by this measure, which is based on the fraction of the total variation in outcomes that can be attributed to the model. R ² is calculated as:

In the above equations, the Y_iindicates the actual value, Y ^ˆ _iindicates the predicted value, Y¯ indicates the means of the Y values, and n is the total number of samples.

We use Python’s scikit-learn tool to construct the proposed ensemble ML scheme (KRR) as well as the investigated classical and ensemble ML models (SVR, NB, RR, RF, and CB). We consider the actual vs. predicted curve and error metrics (MAE, MSE, RMSE, and R ²) as the evaluation parameters of the trained ML models. We use supervised methods to predict crop production, where the dataset contains eight features and the target is the amount of crop production in a certain area. We take the average results of experiments in three phases, such as 80:20, 50:50, and 30:70 training and testing ratio, where each phase has 10 trials.

To get better performance, we tune the hyperparameters of the proposed ensemble ML scheme as well as the investigated classical and ensemble ML models. The same hyperparameters give a better result for almost all experiments. In particular, SVR gives better results with the linear kernel when c = 100 and gamma is auto while we use 10-fold cross-validation to find the value of gamma and c. Gaussian NB achieves a better result in all cases with the hyperparameters i.e., estimator = model, param_grid = params_nb, and cv = cv_methods. RF finds n_estimator = 20, and random_state = 42 in all cases for better performance. For all experiments, RR gives maximum performance when alpha = 0.01. CB model gives a better result when estimator = model_cvr, cv = 2 n_jobs = −1, and learningrate = 0.05. Our proposed model KRR achieves high accuracy with a low error rate with the hyperparameters alpha = 0.01, n_estimator = 10, and random_state = 42 for almost all experiments.

Aus is considered one of the major crops in Bangladesh. This type of rice is closely related to indica-type rice but it has a distinct genetic group (Chakravarthi and Naravaneni, 2006). Still, this variety is cultivated under environmental stress conditions in Bangladesh and India (Berger et al., 2004). The value of Aus production varies according to the environment and the region of cultivation. Figure 3A represents the actual vs. predicted rice production in every fiscal year from 2015 to 2021 in the Dinajpur zone of Bangladesh. The x-axis symbolizes the fiscal year, while the y-axis reflects rice production (both actual and predicted). It clearly indicates that our proposed algorithm outperforms the other traditional ML and ensemble ML algorithms. We also evaluate MAE, MSE, RMSE, and R ² to measure the model’s goodness of fit in predicting rice production in Figure 3B , which demonstrates that the KRR model fits better than the other models we investigate. Table 5 illustrates that KKR has less mistakes, such as 9.11% MAE, approximately 1% MSE, and 9.17% RMSE, than the others.

From Figure 3A , it can also be observed that there is no linear relationship between the area and environmental data, and Aus rice production. The weather conditions in Bangladesh vary from year to year, and natural disasters may occur. In August 2017, during the Kharif harvest season, an uncertain flood occurred in Dinajpur zone (Das and Rahman, 2018). It damaged the crops and interrupted the production cycle. Our proposed KRR performs effectively during this period, which demonstrates its adaptability to the uncertainty in the environmental data. Owing to a lack of soil fertility and improper management of soil carbon, the post-flood effects on Aus production continue throughout the subsequent growing seasons (Siddique et al., 2022). In this uncertain situation, the proposed KRR performs better than other models. This type of prediction is critical for farmers as well as individuals who depend on harvesting for a living. Such future production prediction aids in the care of alternative solutions to ensure food and industrial raw materials.

Aman rice is grown in Bangladesh during the winter (rabi) season. The cultivation of Aman rice is strongly linked to the environment. Figure 4A illustrates the actual and predicted values using the investigated and proposed ML algorithms. The performance of our proposed algorithm KRR reaches maximum accuracy for each fiscal year. In terms of error measurement parameters, both our proposed KRR and the RF models have the same R ² score. However, our KRR obtains better MAE, MSE, and RMSE than RF and other models, as shown in Table 6 . Figure 4B indicates that our KRR is the best-fit model compared to others. Changes in maximum temperatures have had a significant impact on crop yield in Bangladesh. However, temperature changes confirm that maximum temperature raises the risk for Aman rice while minimum temperature reduces yield variability. Rainfall has increased the risk of Aman rice (Sarker et al., 2019). The great news is that environmental factors in Bangladesh are now changing within a range that allows Aman rice to adapt to the environment. In recent years, Aman rice has been consistently produced because of its versatility (Chakrobarty et al., 2021). Figure 4A shows the consistency of Aman rice production, and in most of the cases, the proposed KRR performs better. The authority can benefit from this method in their long and short plan for food supply in the future.

Boro rice is cultivated in the Kharif season, which has a vital impact on the total rice production in Bangladesh. Figure 5A illustrates the actual vs. predicted bar chart for the Boro production from 2015 to 2021, while Figure 5B indicates that KRR is the best-fit model compared to others. Table 7 indicates that the performance using RF is better than KRR in respect of MAE and MSE. However, RMSE and R ² are good in KRR. In summary, the average performance of KRR is better than the other models.

To this end, we investigate three major rice variations in this part to predict their production concerning the environmental conditions. Given the analyses, we can deduce that our proposed model KRR performs better in predicting different rice production in Bangladesh. Boro rice needs extra irrigation for cultivation. The average production of Boro rice is expected to decrease by over 20% in 2050 and by 50% in 2070 as a result of climate change (Basak et al., 2010). It has been determined that an increase in both the maximum and minimum temperatures is the primary cause of a reduction in yield. Rainfall pattern changes during the growing season have also been observed to impact rice production and irrigation needs. Using the proposed KRR, researchers can track environmental factor changes and then take the necessary steps to select an alternative rice variety or predict the production of the new variety.

In Bangladesh, potato farming takes place throughout the winter season. Sandy loam soils can produce more potatoes than other types of soil (Faraji et al., 2017). In terms of productivity and internal demand, potatoes are a popular crop in Bangladesh. As a result, predicting potato production has a significant influence on the economy. Figure 6A demonstrates a comparison among the actual and predicted values using the investigated and proposed ML approaches, which shows that our proposed KRR approach offers superior prediction in almost all cases. According to Table 8 , our proposed approach KRR outperforms the other investigated ML algorithms in all error measures. In particular, KRR predicts potato production with a minimum MSE of 6.3%. Figure 6B illustrates that KRR is the best-fit model to predict potato production. Potato production in Bangladesh is still at a satisfactory level but it swings to change of environment (Hossain and Abdulla, 2016). In some consecutive years, the production goes down due to heavy cold and attack of unexpected diseases on potato (Islam et al., 2022). To predict production in this kind of uncertain situation, the proposed KRR can be a good solution for the agriculture domain people.

In Bangladesh, the production of wheat is decreasing on average by 0.44% each year. People are cultivating different crops instead of wheat for more benefits and a higher production rate. The prediction of the production of wheat can improve the production rate of wheat. We use the same ML and ensemble ML algorithms to predict wheat production in the Dinajpur zone of Bangladesh.

Figure 7A indicates a comparison among the actual and predicted values of wheat rice production using the investigated (SVR, NB, RF, RR, and CB) and proposed (KRR) ML models in each fiscal year from 2015 to 2021. It demonstrates that the performance of our proposed KRR is better than other ML models. In terms of other error metrics, KRR achieves the best result than the other investigated ML approaches, as illustrated in Table 9 and Figure 7B . Wheat production in South Asia climbed from 15 million tons in the 1960s to 95.5 million tons in 2004–2005. It still needs to increase at a rate of 2%–2.5% every year till the middle of the 21st century (Chatrath et al., 2007). Because there is little scope for growing wheat field areas, the main task will be to crack the yield barrier utilizing practical genetic and morphological techniques. Other issues are unique to the highly productive rice–wheat farming system prevalent in the Indo-Gangetic plains. Though the production is at a low level, the 2017–2078 and 2018–2019 time periods have broken records. Previously, we discussed that the damage of the Aus rice due to flood plays a vital role in this segment (Das and Rahman, 2018). People engaged more in cultivating wheat to recover the damage to the economy in the period. It indicates that cultivating more wheat can be a solution to increase the amount of wheat production, which leads the agricultural economy in another direction.

To summarize, we can state that, on average, our proposed model KRR outperforms the others in predicting crop production for all crops considered in this work. Regarding MSE, MAE, RMSE, and R ², the proposed KRR performs better. From Tables 5 to 9 , we can clearly differentiate the performance of each model for prediction. In almost all cases, the MSE, MAE, and RMSE values of KRR are smaller than those of the other models, which indicates that KRR shows minimum error in the case of prediction compared to other ML models. However, the R ² value of the KRR is larger than the other models in the above-mentioned tables. It also creates a comparison among the models that KRR is a better fit to the dataset than existing benchmark ML models. To find the superiority of KNN, the DM significant test is also performed below.

Dataset is one of the significant components of the recommendation system (Zhang et al., 2018). A dataset should train the model used for prediction and recommendation to understand the nature of the system. Then, it can predict and recommend the outcome. In this recommendation system, environmental, area, and production data of the crops are taken into consideration. Five main crops of Bangladesh are considered here as the source of data. The data collection procedure and preprocessing are discussed in detail in the Methods and measurements section.

This stage is one of the crucial parts of the crop recommendation system. The preprocessed dataset is used to train the ML models, and the model predicts the production of different crops. The model makes specific crop predictions based on the area, previous production, and environmental data related to the specific crop production period. Our proposed high-performance model KRR described in the Proposed paradigm section is considered as the potential ML model in the recommendation system.

We consider the standard environmental data as a threshold for the recommendation. In a specific area, crop production depends on environmental properties like temperature, rainfall, wind speed, and sunshine. The environmental variables can be predicted by our proposed model besides the crop production prediction. Threshold values are taken from the local agriculture office and set to our model for the recommendation.

It is the final step of the recommender system. The model predicts the production for an area, and we set the threshold for this system. The threshold is compared with the predicted environmental data, and the crops’ production is considered the main element of recommendation. The recommendation is reached by comparing the threshold with the predicted data for each season’s crop. The top match is the recommended crop for the season in the area. The input of this system is the season and the list of crops, and the output is the recommended crop among the selected crops. As an extension of this work, in the future, a mobile application will be developed that will help farmers gain easy access to this system. However, in this current state, they can use it with the help of experts with a technical knowledge of putting the inputs and synchronizing with this system.