Cap-DiBiL: an automated model for crop water requirement prediction and suitable crop recommendation in agriculture

In this technological era, several approaches used to provide the information about suitable crop recommendation means which is crop is suitable for soil. Some of approaches depends on the IoT smart agricultural-devices to gather information from surrounding area. However, several collection of data are used to predict the crops details but it not efficient to provide better performance. Therefore, the proposed model uses various techniques to improve the performance efficiently. Some steps involved in the proposed model as data pre-processing, feature extraction, feature selection, water requirement prediction and recommendation. Initially, the collected IoT data from dataset are pre-processed using data normalization, missing value imputation and one-hot encoding. Then, extract the features from pre-processed data using Gated Residual autoencoder (GRA) model, whereas optimal features are selected using Chaotic Northern Goshawk Optimization (ChaNgo) algorithm. Based on the farmland details, the crop water requirement prediction and suitable crop recommendation due to the market price are carried out using a novel hybrid deep learning model called Channel capsule-assisted stacked dilated Bi-LSTM (Cap-DiBiL). The channel capsule network predicts the crop water requirement and stacked dilated Bi-LSTM is used for suitable crop recommendations such as millets, rice and other crops. Then the proposed model analyses the performance and compares it with several existing techniques to prove the proposed model’s enhancement. The proposed model improved the accuracy as 98.18% for predicting the crop water requirement and crop recommendation. The performance of proposed model for Precision, Recall and F1 score also enhanced as 98.31%, 98.18% and 98.20%.


Introduction
Agriculture is the most important field and the backbone of developing countries like India.Agriculture aims to produce food for both human and animal consumption.Various crops like maize, rice, millet, cotton, and paddy are cultivated on the land based on the environmental seasons [1].In many developing countries, most people depend on agriculture; three-quarters of the population depends on agriculture.Nowadays, environmental conditions are unstable due to frequent changes in climate conditions that do not follow a regular pattern [2].Climate change affects the cultivation of crops and production yield.More drought, unpredictable winters, and stronger storms significantly affect crop cultivation in many countries [3].Irrigation is vital in agriculture because a proper water supply only produces good and healthy crops.Irrigation is not only to pour water into the plants but also to know how much water is poured into various crops [4].
Developing irrigation systems in agriculture reduces farmers' work to irrigate their crops and plants.Hence, some of the evapotranspiration techniques are used to provide an effective irrigation strategy for farmers [5].Some of the real time data collection are used to provide efficient result about the need and quality of crop based on time difference of a day.The crop recommendation is selection the crop which is suitable for agricultural land, climate and required amount of water need for crop [6].Fewer technology are developed based on nutrients for crop from the soil, climate, requirement of water and fertilizer.Then, analyse the quality and quantity of fertilizer or water are need for crop growth and production [7].To overcome improper irrigation and insufficient water supply, various irrigation techniques are used based on the crops and farmland [8].Thus, water requirements and crop recommendations are effectively predicted by machine learning (ML), deep learning (DL) algorithms, data mining, and the Internet of Things (IoT).Various types of machine learning algorithms, such as regression, support vector machine (SVM), artificial neural networks (ANN), and decision trees, are used to predict water requirements and crop recommendations [9].ML algorithms recommend crops based on farm land, such as tea crops, maize, rice, beans, and sugarcane.
A hybrid ML algorithm predicts tea crop yield using multiple inputs such as weather, crop, irrigation, fertilizer, and soil [10].These hybrid approaches are used to forecast the water and energy demands of large-scale water distribution irrigation.Sample water forecast and crop prediction techniques are developed with the combination of ANN and other genetic algorithms for automatic crop recommendation [11,12].Based on the amount of water and crop for a particular crop, some IoT frameworks are used to predict crop production and improve precision agriculture using ML.Precision agriculture improves crop production costs while boosting yields by lowering chemical usage for plant and crop cultivation, which is stored in a cloud database [13].
ML algorithms are used to train the system to provide efficient performance for crop and crop water requirement prediction with collected datasets.In this model, the IOT framework using the dataset helps to analyse the attributes and control of remote sensing devices like humidity sensors, temperature sensors, and moisturizer sensors [14].Decision trees and ANN combine to predict crops using meteorological attributes like air temperature, wind speed, and solar radiation [15,16].Various approaches are used to solve the problem using DL algorithms, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long-short-term memory (LSTM).DL algorithms and IOT devices perform crop prediction through wireless sensor networks (WSN) [17,18].Numerous problems occur in traditional systems; thus, efficient technologies are developed to avoid such limitations.
The availability of sufficient water on agricultural lands is vital for achieving optimal crop growth.Consequently, it becomes necessary to determine the specific water requirements of individual crops, commonly called crop water requirements.By understanding and adhering to these measures, farmers can ensure that their crops receive the necessary water for healthy and productive growth.Accurate predictions of water distribution on agricultural lands are crucial to determining the precise amount of water available to compensate for evapotranspiration losses.These predictions are key to generating more accurate estimates of crop water requirements.However, capturing and identifying water distribution in specific lands with precision can often be costly and time-consuming, making it challenging to obtain accurate predictions.To address this issue, several models based on empirical relationships and approaches rather than real-time field conditions have been developed.However, they often lead to inaccurate predictions of water distribution on the land.In addition, the recommendation of suitable crops for the farmers highly supports a better yield.But in the existing recommendation algorithms, the recommendation accuracy is declined, and the feature learning capability becomes less, resulting in high over fitting and convergence issues.To overcome these problems, the proposed model was developed for crop recommendation and crop water prediction.Some of the essential contributions are described as follows: • To introduce an automated methodology in the proposed research to predict crop water requirements and recommend suitable crops effectively.
• To extract the effective features using the Gated Residual Autoencoder (GRA) model and select the optimal features using the Chaotic Northern Goshawk Optimization (ChaNgo) algorithm • To introduce a novel hybrid model called Channel capsule-assisted stacked dilated Bi-LSTM (Cap-DiBiL) with enhanced accuracy and fewer error rates.Here, a channel capsule network is used to predict the crop water requirement, and stacked dilated Bi-LSTM is used for suitable crop recommendations.
• To compare the performances of a proposed work with the existing state-of-the-art methods in terms of different performance metrics to prove the performance superiority of a proposed method.
To overcome the loss of features and provide efficient performance during extraction and selection, the proposed model is used.The pre-processing technique involves three techniques used to improve the quality of the data.Then, the hybrid method of the GRA technique is used to extract the feature from pre-processed data efficiently and improve the quality of the feature.The ChaNgo model is used to improve the selection process efficiently, and it is also used for loss function improvement.Finally, the Cap-DiBiL is used to predict crop recommendations and crop water requirements.These hybrid techniques were used to improve the performance and obtain efficient predictions based on the farmland.
The organization of the paper is structured as follows: section 2 represents a survey of the related work, and the techniques are noticed with its limitation.Section 3 details the proposal model description and function with the flow of techniques.Results and discussion are analysed and described in section 4, and the paper's conclusion is described in section 5.

Related works
Some of the related techniques are analysed to improve irrigation, crop prediction and management are noticed as follows.
Elbeltagi et al [19] proposed a technique called the deep neural networks (DNN) method, which was used to incorporate historical data and predict future evapotranspiration (ETc).The main objective of this method was to accurately predict crop water use.Water managers may achieve enhanced, satisfied results in terms of water management and sustainability in agriculture.Determining the water levels at a range of T (min) and T (max) levels.These DNN models generated satisfactory results with high accuracy for the three regions.As a result, these models assisted water managers, planners, and decision-makers in ensuring long-term water sustainability.The limitation of this method was the unavailability of structured historical data sets on crop water requirements.
Kashyap et al [20] projected a deep networking system based on the Internet of Things (IoT) called DLiSA (enabled intelligent irrigation system).The major objective of this method was to solve the growing world population's demands for food and water.Agricultural needs like water savings and irrigation time were concentrated to improve and manage the irrigation scheduler's functionality.The real-time data from three parts is used in the simulation environment settings, calibration and testing, and comparative analysis of the independent test fields.Overall, the proposed model saved more water than the FANNN and T-based models by maintaining the soil moisture deficit within the allowed range.The limitation of this method was the low accuracy level of crop irrigation and water.
Niu et al [21] advised an efficient method called the fixed-threshold technique.The main objective of this paper was to reduce the issues of crop growth and water stress in agriculture.Water stress management and accurate crop trait estimation were required to improve crop water efficiency.The fixed thresholds obtained from ground-based images were successfully applied to images with similar image resolution and variations in image sensors.This method used RF, MLR, and ANN algorithms to predict the crop and the requirement for water for irrigation.This study presented a low-cost and straightforward method for estimating maize FVC and its inter-field variability under various water conditions.In this research, overall accuracy was found to be 99.3%.The limitation of this method was the high occurrence of errors at different stages of crop growth or water stress.Vij et al [22] recommended an efficient technique for the Internet of Things.To monitor the irrigation system, the farmers required a cheaper and more precise solution.The major objective was to solve overirrigation, soil erosion, and crop-specific irrigation problems to ease and efficiently manage irrigation problems.The ML algorithm predicted irrigation patterns based on crop and water scenarios.The limitation of this method was the unavailability of structured data on crop water needs, and the accuracy of the model prediction depends on the availability of data.
Bhuyan et al [23] suggested some techniques for ML algorithms used for crop prediction.The main objective of this method was to predict which crop was best suited to the soil and climate.For crop-type prediction, ML algorithms such as the K-Nearest Neighbours Algorithm (k-NN), support vector machines (SVM), random forests (RF), and gradient boosting (GB) were investigated.High accuracy was promoted by the GB approach to developing a crop forecasting system.In this research, the overall accuracy was found to be 99.20%.The limitation of this method was the requirement for high-level data.The technique and performance are analysed with its limitation described in table 1.
Based on the survey, several technique and approaches were analyse for crop prediction and crop water requirements.Some analysed related works obtained the good performance with few limitations and drawbacks which reduce overall performance of crop recommendation and crop water requirement prediction, such as the unavailability of a dataset and provide low accuracy for the prediction of crop plantation and irrigation.Limited features are used to extract the data, require high-level data for the prediction, and provide high accuracy.High errors obtained in several stages of predicting the crop and water level for irrigation are the existing problems.So, the proposed model is designed to solve the existing problems and improve performance.
Table 1.Existing techniques and performance with its demerits.

Author name and year
Techniques used Merits Demerits Performance

Elbeltagi et al 2020
The technique of deep neural networks (DNN) method that was used to incorporate historical data and predict future ETc The ETc was estimated, and an efficient decision can be made; it provides the result with high accuracy for the three regions.
Unavailability of structured historical data set on crop water requirements.

Kashyap et al 2021
DLiSA technique is used, and it is a deep networking system based on the IoT.
It improves and manages the irrigation schedule functionality, such as saving water and irrigation time.
Provides low accuracy level of crop irrigation and water.
The average volume of the water for required has improved from 23% to 43%.

Niu et al 2021
The fixed-threshold technique is used.It improves the efficiency crop water requirement, water stress management and accurate crop trait estimation.
The high error occurs at different stages of crop growth or water stress.

Accuracy, MAE, RMSE Vij et al 2020 Recommended an efficient technique for the Internet of Things
It solves the over-irrigation, soil erosion, and crop-specific irrigation problems to ease and efficiently manage irrigation problems The unavailability of structured data on crop water requirement and lack of dataset.
The reasonable accuracy of this model is 81.6%

Bhuyan et al 2023 Some techniques of ML algorithms used for crop prediction
High accuracy was promoted by the GB approach in developing a crop forecasting system Requirement of high-level data Accuracy, recall, precision, F1score.

Proposed methodology
In this section, the proposed model for crop recommendation and crop water requirement prediction is described with its equation and work flow.Initially, the collected IoT data, which is either data gathered from an IoT smart device or data collected from a dataset, is stored as the input data used in this model.Then, the data collected from the dataset is pre-processed using data normalization, missing value imputation, and one-hot encoding.Extract the feature using GRA from the pre-processed data and select optimal feature using ChaNgo algorithm.Cap-DiBiL is used to predict for crop recommendation and crop water requirement prediction.Then, it compared the performance with other existing model to describe the superiority of proposed model.The flow work of a proposed model is shown in figure 1.

Data pre-processing
Various steps are involved in pre-processing the gathered data [24] from the IoT devices, such as data normalization, missing value imputation and one-hot encoding.

Data normalization
Normalization is a widely used data preparation technique that converts the values of numeric columns in the dataset to a common scale.Min-Max normalization is a method of adjusting the unnormalized data to predefined upper and lower boundaries.The data are rearranged within the range between 0 to 1 and 1 to 1.
The equation of the min-max normalization is represented in below equation.
Here, the maximum and minimum values of ith feature are denoted as min and max.The data are rescaled as upper and lower bounds are denoted as p max and p min.

Missing value imputation
When the dataset misses some value, the mean value of non-missing values is replaced.The mathematical equation for the missing value imputation is given as, Here, the missing value is denoted as V , i the preceding value from the mission value is represented as V i 1 and the subsequent value from the missing value is denoted as V .
i 1 The natural number is denoted as n.

One-hot encoding
One-hot encoding is the widespread method to transform from categorical to numeric form.When it converts the categorical value into the binary vector with level l, the number of distinct attribute levels are denoted as l.
Any domain can apply this transform, and ML tools mostly assume this method as the default encoding.There are two advantages to using this one-hot encoding of Categorical Attribute transformation Environment (CANE).First, one-hot encoding accepts fast parallel encoding, which encodes each data attribute in a distinct core.Second, the one-hot encoding method allows the user to name the one-hot binary as the full name.

Feature extraction using the GRA model
The feature extraction process is required to ignore the curse of dimensionality problems.The feature extraction of proposed model hybrid three techniques.The GRA model is the advanced and enhanced model that includes three modes: gated recurrent unit (GRU) [25], Residual Attention Network and Autoencoder to improve the feature extraction performance.The GRU model is used to extract useful features to detect energy theft, and it has two gates: the reset gate and the update gate.The mathematical equation for these gates are described in the following equation, Where, the sigmoid is denoted as s and the update, reset gate, candidate hidden state and new hidden state are denoted as U , t R , t k t ¢ and k .t Weight is represented as S, the hidden state at the previous time step is denoted as k t 1 -and the bias term is represented as e.Then the term k tan represents the hyperbolic activation function, and the symbol Ä denotes the Hadamard product.
The residual attention network is designed by stacking multiple attention modules, and every attention module is classified into two branches: mask and trunk.The specified trunk branch provides output ( ) J m with the input m, and the bottom-up and top-down structure is used in the mask branch to learn the size of a mask ( ) B m which is softly weight output features ( ) J m .The fast feedforward and feedback attention processing imitations in between the bottom-up and top-down structure.The neurons of a trunk branch use the output mask as control gates, similar to highway networks.The output of an Attention module is denoted as A,

=
Where, the i ranges overall the spatial positions and the index of a Channel is represented as End-to-end training would be provided for the entire structure.
The attention mask also acts as the gradient update filter during back propagation and feature selector during forward inference.The gradient of a mask for input feature by the soft mask branch is, Where, the mask branch parameter is denoted as q and the truck branch parameter is represented as .j It made the attention module robust to noisy labels.When the trunk parameters can be modified by storing wrong gradients in the mask branch.
Autoencoder is an unsupervised learning technique consisting of three layers: input layer, output layer and hidden layer.The process of autoencoder consists of two stages such as encoding and decoding.The mathematical expression of a process is described in below equation, The input data can be encrypted using the below equations.
The encrypted data could be decrypted using the below equations.
Where, the input data vector is denoted as

=
¼ and the input data of reconstruction layer is where n represents the size of the input vector and m represents the number of hidden layer.The weight connection matrix between the input and hidden layer is represented as T Q .
f Î ´are the input and hidden layer of the bias vectors.( ) h 1 q ⋅ and ( ) h 2 q ⋅ represented as the activation function of a hidden layer neuron and output layer neuron to map the network range to [ ] 0, 1 .The sigmoid function is used as an activation function described in the below equation.
The error between the original and output reconstructed data is minimized while adjusting the encoder and decoder parameters.The output data from the hidden layer unit at a time is assumed to be an optimal lowdimensional representation of the original data, and all data exists in the original data.
Where, the number of input samples is denoted as M. The architecture of a GRA is shown in figure 2.

Optimal feature selection using ChaNgo
The optimal feature selection reduces the falsely selected features by half while maintaining the same level of a true positive rate.The northern Goshawk Optimization (Ngo) is used with the Tent Chaotic map to improve the process of optimal feature selection.

Initialization process
The ChaNgo is a population-based algorithm used to search the members of this algorithm.It means a solution for the problem that describes the value of a variable.In mathematical form, each population member acts as a vector and the vector's population together acts as a matrix for the algorithm.The members are randomly initialized for the search space at the beginning of an algorithm.The mathematical expression for the Ngo algorithm is given below as an equation.
´Ẃhere, the population of the northern goshawks is denoted as A and the expected solution as ith solution is represented as A .
i a i j ´is the value of a jth variable which is specified by the ith expected solution and number of prey are denoted as M. Hence, the problem is evaluated by objective function on each population member.These values can be described as a vector using the equation below.
Ẃhere, the vector of objective function value is denoted as Y and Y i obtained by ith as expected solution.The criterion is used to find the best solution of an objective function value.For minimization problems, the value of an objective function is smaller, for maximization problems, a larger value is the better solution.New values are obtained for each iteration of an objective function, and during the iteration, the expected solution is updated.

Mathematical model
The Ngo model is designed to update the population members and the simulation strategy while haunting.The strategy of a northern goshawk has two main behaviour such as, (a) Identify and attack the prey.
(b) Chase the prey and escape operation are the two main phases of a northern goshawk.

Identification of prey
This phase improves the power of exploration over the Ngo algorithm due to selecting the prey randomly in search space.Identifying the optimal region is found in this phase, leading to the global search.Identification of prey and attack of the northern goshawk is a schematic behaviour.The concept of this phase is mathematically expressed using the below equation.


Where, the prey's position is denoted as R i for the i th goshawk and its objective function is denoted as Y .

R i The random selection of natural numbers in interval [
] n 1, is denoted as k and the new status H i new R , 1 for i th solution and H i j new R , , 1 is j th dimension.The objective function value of a first phase is denoted as Y i new R , 1 and the random number is denoted as s in the interval [ ] 0, 1 and K denotes the random number of 1 or 2. s and K are the random number which is used to generate the behaviour of an Ngo algorithm for search and update.

Chase the prey and escape the operation
This phase is used for exploitation that increases the exploitation power of Ngo to local search.This haunt phase is also considered to be close in radius D to the attack position.The mathematical equation of this phase is described in below = <


Where, the iteration counter is denoted as t and the maximum number of iterations is denoted as L. The new status of i th expected solution is H , , 1 is its j th dimension and the objective function value based on this second phase is denoted as Y .
The random vector is used for a goshawk to choose the prey, chase and attack with high speed in the Ngo algorithm, and it is replaced with the Tent chaotic (tc) map, which enhances the effective optimization after the tc is used.The mathematical equation of this phase is assumed as the below equation.
Here, s signifies the current number of iterations is occurred and To enhance the performance of global search, the Tent chaotic is used.The ChaNgo algorithm's several stages are described in the flow chart in figure 3.

Channel capsule-assisted stacked dilated Bi-LSTM (Cap-DiBiL)
This proposed Cap-DiBiL is used to collect the details about the required water for crop and crop recommendation based on the land.This proposed model involves the integration of three models: capsule network, stacked dilated Bi-LSTM and convolutional block attention.The LSTM need more memory and obtain high computational time to process.Then, it not well-suitable for prediction or classification task.It is difficult to train the requirement of more training data and it is slow to process large datasets.Hence, the Cap -DiBiL is introduce to improve the performance and efficiency.
A capsule network contains various layers, each consisting of several capsules.The capsule length is measured by the existence probability of an associated instance [26].When each capsule i, which have installation parameters x i and try to guess the output of next layer with a trainable weight matrix Where, the prediction of capsule i for capsule j is denoted as 

|
x .j i The prediction processing is considered based on a coefficient using the routing by agreement process.It determines the actual output of the capsule j, which is denoted as e j in below equation ( Where, the agreement between the prediction and output is denoted as p ij and the given score for prediction is denoted as r .
ij It determines the contribution of prediction to output.Routing by agreement differs from the CNN to CapsNet function and identifies the spatial relation.
In CapsNet loss function, the capsule k final loss of summation is denoted as f k and it is measured as below, Where, the class k is present, otherwise zero is denoted as J .
k The hyper-parameters of a model are denoted as the terms u , + u -and .b The final loss of summation overall the f , k e.The deep convolutional neural network area is used to extract the hierarchical feature from the images, such as high-level network outputs abstract semantic features and low-level network outputs simple geometric features.The feedforward convolutional neural network uses the convolutional block attention model (CBAM), a simple and effective module which is an intermediate feature map sequence.Then adaptive feature optimization are provided from the multiply attention map by input feature map.The weighted result is the output of convolution layer which first pass through the channel attention module, and the final weighted result gets from the output of a convolution layer passed through the spatial attention module.
The mathematical equation of channel attention and spatial attention modules is given below. ( ) )) ( ) the input of a feature map is denoted as E, the channel attention module is represented as ( ) A E c and the sigmoid function is denoted as .
s The fully connected multi-layer perceptron is denoted as mlp.The average pooling and maximum pooling function is denoted as avgpool and pool max .The convolution operation is denoted as the function g is a 7 7.
´The mathematical equation for the output formula after introducing the residual network based on the Channel and spatial attention module is represented as, Where, the output feature map of a channel attention module is denoted as E , c the channel attention module operation is denoted as A .
c The input feature map is represented as E, and element-by-element multiplication is denoted as .
Ä The spatial attention output feature map is represented as E .s Acquiring the idea of feature map space and the channel information of the previous layer produces unique weight, and the channel attention process is utilized by the channel attention.After combining the attention mechanism, combine the weight with the initial attribute map to address the goal of integration.The existence probability of an associated instance measures the capsule length.It enhances the detection network's ability to detect the target's feature expression which improve the target accuracy detection.The architecture of the Cap-DiBiL model is shown in figure 4.
The stable architecture of long-short-term memory (LSTM) is used to overcome the vanishing gradient problem and convey essential information in the LSTM network.The traditional Recurrent Neural Network (RNN) may fail to capture the long-term dependencies between the feature vectors due to the vanishing gradient problem during model training [27].There are three gates used in the LSTM model as input gate n , t the forget gate g t and the output gate u .
t These gates control the memory cell activation vector.

Forget gate
The forget gate is used to know the amount of information that is forgotten from the previous state s t

Performance constraints evaluating the proposed model
Performance constraints, such as accuracy, recall, precision, error, loss, kappa and F1 score, are used to measure the proposed model's performance.These performance constraints are described with their formulation.

Accuracy
The accuracy rate is a performance and correct prediction of the proposed model.The mathematical equation is represented below.
( ) accuracy ts tn ts wn tn ws Where, ts denotes the number of correctly selected samples and ws represents the number of incorrectly classified samples.Then, tn denotes the number of incorrectly selected samples.

Recall
The ratio of the number of positive samples detected is known as the recall rate, and the mathematical equation of a recall rate is described below, ( ) r ts ts ws 100% 41 = + Ẃhere, the recall is denoted as r.

Precision
The ratio of accuracy rate over several predicted positive samples with all positive samples.The mathematical equation for the precision rate is represented below, ( ) p ts ts ws 100% 42 = + Ẃhere, the precision is denoted as p.

Error
During testing, the input as entire data is well-trained to predict the density map.Two types of error are obtained: mean absolute error (MAE) and the root mean squared error (RMSE).MAE used to measure the average magnitude of errors in set and RMSE used to measure the quality of predictions.The mathematical equation of these metrics is denoted as, Where, the number of data in a proposed data set is denoted as n and the estimated count of error is denoted as P. i The corresponding actual count is denoted as P .represents the expected accuracy rate and accuracy denotes the occurred accuracy value of a proposed model.

F1 score
The F1 score is an evaluation metrics used to measure the proposed model's accuracy while comparing the performance between two classes.It measured by precision and recall rate is described in the below equation.magnitude of an error and is analysed and compared with the existing model.The computational time of a proposed model is analysed and compared with the existing model, which are shown in figure 9.
The computational time of a proposed model is analysed to verify the running time and compared to the other existing model.Based on the analysis of figure 9, the computational time of a proposed model provides less      used to measure average magnitude of error which is lowest value than other existing model.The MAE used to reduce the large outlier and RMSE is biased estimator which is used to measure the predication rate in both crop recommendation and crop water requirement predication.Finally, the computational time of the proposed method is analysed and it obtained as 16 s to perform the process of crop recommendation and crop water requirement predication.Then it compared with other model performance indicate that the proposed model obtain lowest time to perform.These result are obtained from proposed model described the overall performance of entire model and its efficiency.

Conclusion
The IoT device has made a major revolution in the agricultural field, and the proposed model is also introduced to provide efficient performance for crop recommendation and crop water requirement prediction.The proposed model improves efficiency by using several techniques to perform the tasks of crop recommendation and crop water requirement prediction.The data collected from the dataset is pre-processed to improve the quality of the data.The GRA model is used to extract the feature, and the ChaNgo model is used for optimal feature selection.Finally, the proposed novel Cap-DiBiL technique is used to grow crops suitable for land and require water for crop growth.Then, the performance of the proposed model is analysed for various metrics, such as the enhanced accuracy rate of 96.36% to 98.18% and the precision value of 98.31%.A proposed model's recall and F1 score are enhanced as 98.18% and 98.20%.The MCC value provided by the proposed model is 97.97%, and the Kappa value is 97.97%.Analysing the overall performance of the proposed model to obtain efficient performance and predict the crop recommendation and crop water requirement In the future, hybrid optimization techniques will be explore these techniques to improve performance and use advanced datasets.Link https://.kaggle.com/datasets/atharvaingle/crop-recommendation-dataset(Note: Additionally, the crop water requiement details and market price details will be manually added along with the dataset).https://github.com/Ravi2kinus/Dataset.git.

Figure 1 .
Figure 1.Workflow of the proposed model.
consumption of resources and accurately restore data during model training.T T jk ij t = are usually exist then, Q n 1 1 reconstruction error function between E and the mean squared error function D. The mathematical equation of error function shown in below.

Figure 2
Figure 2 Architecture of GRA.
The kappa value is used to measure the efficiency of a proposed model.The mathematical equation of kappa is evaluated as given below, Where, ex accuracy .

Figure 6 .
Figure 6.comparison of the proposed model with existing (a) training accuracy, (b) training loss, (c) testing accuracy and (d) testing loss.

Figure 8 .
Figure 8.Comparison of MAE and RMSE error.
1 -based on the hidden state d t 1 -and present input a .t The forget gate provides the output between 0 and 1.The mathematical equation for the forget gate is represented below, context of web pages.These techniques are used in the proposed model for crop water requirement prediction and recommendation.
-Where, the bias vector is denoted as m , h v rg and v .kg The weight matrix lies between r t and g , t the weight matrix lies between k t 1 -and g .t and local

16
Environ.Res.Commun.5 (2023) 095016 R K Munaganuri and Y N Rao running time than other existing models.The computational time of a proposed model takes 16 min; it is 6 min lesser than the Bi-LSTM model.4.3.Discussion The performance of proposed model is analyse for each metric separately and compare it with other existing model.The overall performance of proposed model obtain efficient result such as Accuracy rate as 98.18%, Recall value as 98.18%, F1 score as 98.20%, Kappa value as 97.97%, Recall as 98.98% and MCC value as 97.97% describes the efficient performance.These performance rate of the proposed model obtained higher performance than existing model which described the efficiency and improvement of proposed model.The training and testing of proposed model are analysed to measure the performance during the training with subset of data and testing process occurred to measure or verify the overall accuracy and loss during testing.The performance of accuracy and loss for both training and testing phase of proposed model represents the accuracy and loss rate briefly.The accuracy of the proposed model for both training and testing are slightly improved the performance closed to 1.00 Epochs and the loss rate re slightly reduced the loss rate during the training and testing of the proposed model.The performance.The ROC curve describes the performance of proposed which is used to evaluate the performance of classification model at True Positive Rate TPR and False Positive Rate FPR.The curve top-left corner from the ROC Curve indicates the performance and described with other existing model shows the supervisory of proposed model.The ROC Curve is only used to measure the binary classification problems.Then, the MAE and RMSE rate of propose model is analysed and it indicate the MAE