Deep Learning IoT Enabled Driven Data Centric Crop Yield Forecasting Using Optimized Temporal-CRN Networks with BIANO Algorithm

Agriculture plays a pivotal role in ensuring food security and maintaining the economic resilience of many parts of the world, especially in developing countries, where agricultural production is a direct contributor to livelihoods. High population rate, rising climatic insecurity, and scarcity of natural resources have heightened the strain on the traditional farming systems, which have made the traditional ways of farming inadequate in providing food to the rising demands. This is increasingly turning modern agriculture into a data-driven and technology-enabled approach that empowers farmers to make evidence-based decisions on time. Real-time environmental monitoring combined with smart analytical systems has great potential in streamlining crop management, reducing risks in production, and improving the overall performance of yields. The foremost importance is caused by the necessity to offer accurate and early predictions of crop yield to make well-informed choices in terms of irrigation, nutrients, post-harvest storing, and logistics to the market to reduce losses and increase financial benefits. It is on this paradigm that a combination of sensing technologies and new computational intelligence has emerged as a fount of sustainable and efficient agricultural development. The emergence of the Internet of Things (IoT) technologies in the recent past, as well as the growth of artificial intelligence, has revolutionized the sphere of agricultural data acquisition and analysis fundamentally. Related sensors in the IoT yield ongoing data on the nature of the soil, microclimatic zones, and the physiological state of crops during the growth cycle, and generate gigantic volumes of heterogeneous data rich in time. At the same time, deep learning techniques are exceptionally efficient in both capturing the nonlinear correlations, which are complex in nature, and in identifying meaningful patterns in the huge volume of agricultural data. Nevertheless, agricultural systems are dynamic, nonlinear, and extremely noisy, and hence a massive challenge in terms of a precise forecast of the output. Conventional machine learning and deep learning architectures are likely to be low power to capture long-term temporal correlations, and to put more emphasis on the phenological stage, which has a disproportionate effect on end yield. These shortcomings have spawned an increasing interest in hybrid modeling schemes that integrate more advanced temporal learning schemes, attention-based feature weighting, and optimizing feature selection schemes in order to more effectively utilize multi-source agricultural data.

3.1 Proposed methodology

The suggested methodology in Figure 1 offers both integrated IoT-based, data-centric deep learning for the correct prediction of crop products through real-time agricultural monitoring, sophisticated data preprocessing, optimal feature learning, and a TA-CRN architecture. The constant data gathering is carried out by heterogeneous IoT sensor nodes that are installed on agricultural fields to measure soil physicochemical characteristics, microclimatic changes, and crop growth indices at small temporal scales. The data obtained due to environmental disturbances, sensor drift, and communication are high-dimensional, noisy, asynchronous, and nonlinear in nature. To overcome such difficulties, the multi-stage preprocessing pipeline is organized, consisting of noise filtering, missing-value imputation, normalization, standardization, and temporal synchronization. Then, a two-step dimensionality reduction plan that incorporates the AFC and ICA is adopted to obtain small nonlinear latent representations and statistically independent components, in effect, minimizing redundancy with salient agronomic patterns.

Figure 1. Proposed architecture

After dimensionality reduction, multi-level feature extraction is done on traditional statistical descriptors, higher-order statistical measures, and two adaptive features, such as adaptive weighted skewness and adaptive weighted kurtosis, to be more sensitive to distributional asymmetry, tail behavior, and yield-critical extreme events. To achieve higher predictive performance and learning efficiency, a BIANO algorithm is adopted to select optimal feature subsets and dynamically tune hyperparameters, yielding compact, discriminative feature sets. The optimized features are then input into the TA-CRN model, which incorporates the use of convolutional layers to extract local patterns, recurrent units to model the temporal dependencies, and an attention mechanism to focus on the yield-critical growth stages and time steps. The framework is also trained and validated using standard performance metrics, is scalable and generalizable, and is relevant to real-time precision agriculture applications.

3.2 Data collection

The data collection phase aimed to have high-resolution and season-long monitoring of crop development in a wide-ranging environment of actual agricultural fields, as given in Table 1. The experiment was done on rice (Oryza sativa) as it is a staple cereal crop that is highly sensitive to soil moisture, variation in temperature, and changes in climate, and therefore a perfect subject to explore the potential of advanced IoT-based predictive models. The experimental agricultural field in the state of Tamil Nadu, India, was used to gather data, which was a tropical area and had heavy practices of rice growing. The entire crop growth was monitored, which was the Kharif season, June to November, which contains such important phenological stages as the development of the seedlings, tillering, the development of the panicle, flowering, and maturity. The heterogeneous dataset was formed by combining real-time measurements of the IoT sensors, weather station measurements, and agronomic records at the field level. Parameters of the soil were constantly measured at different depths to measure the root-zone dynamics, whereas parameters on the atmosphere and crop-level were measured to explain the external environmental stressors and physiological responses. To have ground-truth points of reference to model training and validation, besides sensor-obtained data, historical yield data, and agronomic logs kept by farm authorities were also included. The gathered dataset therefore forms a multi-source, multi-temporal, and multi-dimensional agricultural data warehouse, which can strongly learn nonlinear interactions between soil, weather, crop development, and yield performance.

Table 1. Detailed description of the crop yield dataset used in the study

It is important to note that the present dataset was intentionally constructed as a high-resolution single-site longitudinal field dataset rather than a broad multi-region benchmark. The primary objective of this study was to evaluate the methodological effectiveness of the proposed TA-CRN–BIANO framework under tightly monitored real-time agronomic conditions, where sensor synchronization, temporal continuity, and label reliability could be maintained across the full crop lifecycle. Accordingly, the dataset covers one complete Kharif-season rice cultivation cycle from a single irrigated lowland paddy environment in Tamil Nadu, India, and should therefore be interpreted as a controlled proof-of-concept deployment dataset rather than a universally representative agroecological benchmark.

3.3 IoT-based agricultural data acquisition

An IoT sensor network cell was distributed over the paddy field, and its placement was strategic in order to record the soil, atmospheric, and canopy conditions of crops. Soil sensors were also placed in root-zone depths to detect moisture availability, temperature change, pH balance, and electrical conductivity, all of which directly affect nutrient uptake and plant well-being. At the same time, weather stations planted around the field took atmospheric parameters that included the air temperature, relative humidity, rainfall, wind speed, and solar radiation, which guaranteed proper depiction of microclimatic conditions that influence crop growth (Table 2).

Table 2. IoT sensors, parameters, and data acquisition specifications

All the sensors were linked by the low-power wireless communication standards, which allowed the possibility of easy data transfer to an IoT gateway and a cloud-based storage system. The sensor readings were time-stamped in order to allow the time correspondence between non-homogeneous data streams. The system of acquisition was designed to be in effect during the whole season of cropping and detect both slow tendencies and sudden changes in the environment, like heat stress or high rainfall. This high-resolution sensing infrastructure guaranteed the presence of high-quality temporal sequences needed for deep learning-based temporal modellings based on TA-CRN structures.

3.4 Data preprocessing

 The methodological components used in the proposed framework include several well-established learning primitives, namely autoencoder-based latent compression, independent component decomposition, convolutional recurrent temporal modeling, and attention-based sequence weighting. Since these components are already well known in the deep learning literature, the present section intentionally emphasizes their task-specific integration rather than reiterating generic theoretical foundations. The principal methodological contribution of this work lies not in the isolated use of these standard modules, but in their ordered co-optimization for agricultural IoT data, where nonlinear compression, redundancy disentanglement, adaptive higher-order agronomic descriptors, and nutrient-inspired feature–hyperparameter search are jointly coupled before temporal classification. Accordingly, the following subsections focus primarily on the operational role, inter-module dependency, and contribution of each block within the proposed TA-CRN–BIANO pipeline.

Preprocessing is thus necessary to improve the quality of data and generate a consistent learning behavior among deep neural models like TA-CRN. Within the framework proposed, preprocessing of data will be done in a multi-staged pipeline of data noise removal, integration of missing values, data normalization, data standardization, and temporal alignment. This pipeline provides that raw sensor signals can be converted to clean, consistent, and temporally coherent sequences that can be used in temporal-based and recurrent neural learning. Predictive accuracy is not only enhanced by the preprocessing stage but also converges faster and reduces overfitting when training a model. Mathematically, let the raw IoT dataset be represented as (Eq. (1))

$\mathbf{X}=\left\{ {{x}_{i,j}}\left( t \right) \right\},i=1,2,\ldots ,N,j=1,2,\ldots ,M$    (1)

where, ${{x}_{i,j}}\left( t \right)$ denotes the reading of the $j$- The sensor parameter at time $t$ for the $i$- the observation, $N$ is the number of samples, and $M$ is the number of monitored parameters.

3.4.1 Data cleaning and noise removal

Noise in agricultural sensor data arises from environmental interference, sensor aging, calibration drift, and abrupt weather events. To address this, a hybrid noise filtering strategy is employed, combining statistical thresholding and smoothing-based techniques. Firstly, an interquartile range (IQR) criterion is used to identify outliers, and the criterion is (Eq. (2)):

$\begin{gathered}\mathrm{IQR}=Q_3-Q_1 \\ x \in \text { Outlier if } x<Q_1-1.5(\mathrm{IQR}) \text { or } x>Q_3+ 1.5(\mathrm{IQR})\end{gathered}$    (2)

The outliers detected are fixed via local interpolation or eliminated when it is physically implausible. Then, an average filter is used to eliminate high noise in (Eq. (3)):

$\hat{x}\left( t \right)=\frac{1}{w}\underset{k=t-w+1}{\overset{t}{\mathop \sum }}\,x\left( k \right)$    (3)

where, w denotes the size of the window. This is done to make sure there is a smooth transition across time, and the agronomic trends that are meaningful (e.g., progressive drying of soil, increasing temperature) are maintained.

3.4.2 Handling missing and inconsistent sensor data

Missing values are caused by loss of packets, malfunctioning of sensors, or loss of power. In order to provide continuity of data, missing observations are filled in by a hybrid temporal interpolation method. In short discontinuities, linear interpolation is used in (Eq. (4)):

$x\left( t \right)=x\left( {{t}_{1}} \right)+\frac{t-{{t}_{1}}}{{{t}_{2}}-{{t}_{1}}}\left[ x\left( {{t}_{2}} \right)-x\left( {{t}_{1}} \right) \right]$    (4)

where, ${{t}_{1}}$ and ${{t}_{2}}$ denote the nearest valid timestamps. For longer gaps, a mean-based temporal imputation is employed in (Eq. (5)):

$x\left( t \right)=\frac{1}{K}\underset{k=1}{\overset{K}{\mathop \sum }}\,x\left( t-k \right)$    (5)

Inconsistent values that do not meet physical constraints (e.g., negative rainfall values or extreme pH values) are detected by use of domain-specific rules and rectified. This is the strategy that does not create artificial bias in the dataset completeness.

3.4.3 Data normalization and standardization

Because of the differences in scale of sensor parameters, normalization and standardization are vital in avoiding feature dominance in the case of neural learning. The first step is min-max normalization of bounded values within a given range in (Eq. (6)):

${{x}_{\text{norm}}}=\frac{x-{{x}_{\text{min}}}}{{{x}_{\text{max}}}-{{x}_{\text{min}}}}$    (6)

Then, the standardization of the Z-score is performed to have a zero mean and unit variance in (Eq. (7)):

${{x}_{\text{std}}}=\frac{x-\mu }{\sigma }$    (7)

where, $\mu$ and $\sigma$ represent the mean and standard deviation, respectively. This bilateral change makes gradient propagation stable and increases the consistency of firing of temporal neurons of the TA.

3.4.4 Temporal alignment of multi-source IoT data

Temporal alignment is needed to coordinate heterogeneous data streams since the various sensors have different sampling frequencies. There is a popular time grid Tcis which is defined, and all sensor readings are resampled with interpolation (Eq. (8)):

${{x}_{j}}\left( {{T}_{c}} \right)=\mathcal{I}\left( {{x}_{j}}\left( t \right) \right)$    (8)

where, $\mathcal{I}\left( \cdot  \right)$ denotes interpolation. This step ensures that each feature vector is considered at a time. ${{T}_{c}}$ represents a coherent agronomic state, enabling accurate temporal dependency modeling in the CRN layer.

3.5 Two-fold dimensionality reduction

In order to handle these issues, the two-fold dimensionality reduction approach is taken using the AFC with the CA. This is a hybrid method using the nonlinear representation learning ability of autoencoders and the statistical independence modeling ability of ICA, creating small, informative, and decorrelated feature representations that can be used in deep temporal learning.

Let the preprocessed IoT dataset be represented as (Eq. (9))

$\mathbf{X}=\left\{ {{\mathbf{x}}_{1}},{{\mathbf{x}}_{2}},\ldots ,{{\mathbf{x}}_{N}} \right\}\in {{\mathbb{R}}^{N\times d}}$    (9)

where, $N$ denotes the number of samples and $d$ represents the original feature dimensionality. The objective of dimensionality reduction is to transform. $\mathbf{X}$ into a low-dimensional latent space $\mathbf{Z}\in {{\mathbb{R}}^{N\times {d}'}}$, where ${d}'\ll d$, while preserving agronomically relevant information.

3.5.1 Autoencoder-based Feature Compression (AFC)

AFC is employed as the first dimensionality reduction stage to capture nonlinear correlations among agricultural sensor features. An autoencoder consists of an encoder–decoder pair trained to reconstruct the input data with minimal error. The encoder maps the input vector. ${{\text{x}}_{i}}$ into a low-dimensional latent representation ${{\text{h}}_{i}}$ as (Eqs. (10)-(12))

${{\text{h}}_{i}}={{f}_{\text{enc}}}\left( {{\text{x}}_{i}} \right)=\sigma \left( {{\text{W}}_{e}}{{\text{x}}_{i}}+{{\text{b}}_{e}} \right)$    (10)

where, ${{\text{W}}_{e}}$ and ${{\text{b}}_{e}}$ denote encoder weights and biases, respectively, and $\sigma \left( \cdot  \right)$ is a nonlinear activation function. The decoder reconstructs the input as

$\hat{\mathrm{x}}_i=f_{\text {dec }}\left(\mathrm{h}_i\right)=\sigma\left(\mathrm{W}_d \mathrm{~h}_i+\mathrm{b}_d\right)$    (11)

where, ${{\text{W}}_{d}}$ and ${{\text{b}}_{d}}$ are decoder parameters. The autoencoder is trained by minimizing the reconstruction loss.

${{\mathcal{L}}_{\text{AFC}}}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathop \sum }}\,\parallel {{\mathbf{x}}_{i}}-{{\mathbf{\hat{x}}}_{i}}{{\parallel }^{2}}$    (12)

Through this optimization, AFC learns compact nonlinear latent features that preserve essential agronomic patterns such as soil–climate interactions and crop growth trends, while significantly reducing dimensionality.

3.5.2 Independent component analysis (ICA)

The ICA is a technique in machine learning. Even though AFC is effective at compressing data, it is possible that the latent features still have statistical dependencies. To further bring out features, the latent space generated by AFC is subjected to ICA. The output of AFC will be represented by Rf in (Eqs. (13)-(16)):

$\text{H}={{[{{\text{h}}_{1}},{{\text{h}}_{2}},\ldots ,{{\text{h}}_{N}}]}^{T}}$    (13)

ICA presupposes that its linear combination of statistically independent source components S:

$\text{H}=\text{AS}$    (14)

where, A is a mixing matrix, unknown. ICA is aimed at estimating an unmixing matrix W, where

$\text{S}=\text{WH}$    (15)

and the elements of Sare are as independent as possible. This is done by maximizing non-Gaussianity, which is commonly quantified by negentropy:

$J\left( \mathbf{s} \right)=H\left( {{\mathbf{s}}_{\text{gauss}}} \right)-H\left( \mathbf{s} \right)$    (16)

where, $H\left( \cdot  \right)$ denotes entropy and ${{\mathbf{s}}_{\text{gauss}}}$ is a Gaussian random variable with the same covariance. ICA can effectively separate independent agronomic factors (moisture stress, temperature variability, nutrient dynamics, et cetera) and increase the interpretability and robustness.

3.5.3 Integrated AFC–ICA dimensionality reduction strategy

The final reduced feature representation is obtained by sequentially applying AFC followed by ICA (Eq. (17)):

$\mathbf{Z}=\text{ICA}\left( \text{AFC}\left( \mathbf{X} \right) \right)$    (17)

3.6 Feature extraction

The importance of feature extraction is to convert dimension-reduced agricultural IoT data into informative representations that can well characterize crop growth dynamics and yield variability. Raw sensor signals can be insensitive to distributional properties, temporal variability, or nonlinear agronomic responses, even when they are reduced in dimensionality. Hence, based on the proposed framework, multi-level feature extraction is applied to obtain first-order trends, dispersion trends, higher-order statistical processes, and extreme-event sensitivities, which soil-crop-climate relationships are supposed to exhibit. This stratified feature learning approach will make sure that not only stable growth trends are covered, but also, seldom, yet significant stressors (e.g., drought spells or heat waves) are also present. The reduced feature vector at time t may be denoted as z(t), i.e., as z1 (t), z2 (t), ..., z d(t), where d is the reduced dimensionality achieved by the combined AFC-ICA step. The extraction of the features is done across sliding temporal windows to maintain agronomic continuity and time context.

3.6.1 Statistical feature extraction

The statistical characteristics are obtained to describe the attributes of central tendency and dispersion of agricultural sensor signals with time. These characteristics offer a condensed overview of the main tendencies in the change of soil moisture level, temperature changes, and crop physiology, which are highly linked with the formation of yields. The average value of a feature over a time window of length T is calculated as (Eqs. (18)-(21)):

$\mu =\frac{1}{T}\underset{t=1}{\overset{T}{\mathop \sum }}\,z\left( t \right)$    (18)

which represents the average agronomic condition over the observation period. Variability around this average is captured using variance, defined as

${{\sigma }^{2}}=\frac{1}{T}\underset{t=1}{\overset{T}{\mathop \sum }}\,{{(z\left( t \right)-\mu )}^{2}}$    (19)

These measures are essential for identifying stable versus volatile environmental conditions. Additionally, the standard deviation

$\sigma =\sqrt{{{\sigma }^{2}}}$    (20)

Allows for obtaining scale-sensitive dispersion measures affecting temporal neuron firing stability in the TA layer. These statistically descriptive features allow us to construct the fundamental feature set that describes the baseline behavior of crop growth under different environmental conditions. In addition to dispersion, the energy of the signal is calculated to give cumulative agronomic intensity:

$E=\underset{t=1}{\overset{T}{\mathop \sum }}\,{{z}^{2}}\left( t \right)$    (21)

which is specifically applied in the capture of long-term stress situations like long-duration high temperature or constant lack of moisture. These statistical properties guarantee that the representation extracted retains both magnitude-based and variability-based information that is required in yield forecasting. Higher-order statistical features are extracted using t-statistical methods with application to the responses

3.6.2 Higher-order statistical feature extraction

The responses are subjected to statistical methods of extracting higher-order statistical features. Although both first and second order statistics explain the overall behavior and dispersion, they do not have the capacity to explain the asymmetry and tail behavior of the agricultural sensor distribution. Nonlinear and extreme events often affect crop yield, i.e., sudden increases of rainfall or short-term heat stress, and thus demand higher-order statistical measures. The first computed skewness is to estimate the asymmetry of the distribution (Eqs. (22) and (23)):

$\text{Skewness}=\frac{1}{T}\underset{t=1}{\overset{T}{\mathop \sum }}\,{{\left( \frac{z\left( t \right)-\mu }{\sigma } \right)}^{3}}$    (22)

indicating whether sensor readings are biased toward lower or higher values. Similarly, kurtosis quantifies the peakedness and tail heaviness of the distribution:

$\text{Kurtosis}=\frac{1}{T}\underset{t=1}{\overset{T}{\mathop \sum }}\,{{\left( \frac{z\left( t \right)-\mu }{\sigma } \right)}^{4}}$    (23)

These higher-order features are particularly effective in detecting abrupt agronomic stress conditions that may not significantly alter the mean but drastically affect yield outcomes.

3.6.3 Adaptive weighted skewness feature

Traditional skewness treats all observations equally, which may dilute the influence of critical growth-stage variations. To overcome this limitation, an adaptive weighted skewness feature is proposed, assigning higher importance to sensor observations occurring during yield-sensitive growth phases. Let $w\left( t \right)$ denote an adaptive weight derived from temporal relevance or sensor reliability. The adaptive weighted skewness is defined as (Eq. (24)):

$\text{AWS}=\frac{\mathop{\sum }_{t=1}^{T}w\left( t \right){{\left( \frac{z\left( t \right)-\mu }{\sigma } \right)}^{3}}}{\mathop{\sum }_{t=1}^{T}w\left( t \right)}$    (24)

The weights are dynamically set to reflect the volatility of the environments and the agronomic significance, and ensure that the deviations in flowering or grain-filling phases have a stronger effect on the feature representation. This expression allows the model to highlight asymmetric behavior in times of critical behavior as opposed to averaging it in the whole season. Therefore, adaptive weighted skewness is much more sensitive to phase-specific patterns of stress that have a direct influence on the formation of the yield.

3.6.4 Adaptive weighted kurtosis feature

Similar to skewness, conventional kurtosis fails to account for the temporal significance of extreme observations. The proposed adaptive weighted kurtosis feature addresses this limitation by amplifying the contribution of rare but agronomically critical events. It is formulated as (Eq. (25)):

$\text{AWK}=\frac{\mathop{\sum }_{t=1}^{T}w\left( t \right){{\left( \frac{z\left( t \right)-\mu }{\sigma } \right)}^{4}}}{\mathop{\sum }_{t=1}^{T}w\left( t \right)}$    (25)

This adaptive formulation guarantees that there are no sudden events in the temperature, unforeseen events of rainfall, or sudden depletion of soil moisture that will be repressed in the process of aggregation. The presence of high adaptive weighted kurtosis implies heavy-tailed distributions in the case of extreme stress conditions, which allows the TA-CRN model to acquire strong representations of yield-disruptive events.

3.7 Proposed technique: Temporal Attention–Enhanced Convolutional Recurrent Network

The nature of crop growth and yield formation is controlled by cumulative interactions between the environment, over longer time scales, and both short-term variations (varying rainfall or heat stress) and long-term ones (like soil moisture storage and seasonal climate variations) are important. Traditional deep learning architectures tend to be incapable of both local feature interactions and long-range temporal interactions. To overcome this weakness, the suggested framework combines convolutional feature extraction, recurrent temporal modeling, and an attention mechanism into one framework given in Figure 2.

Figure 2. Proposed technique

The TA-CRN works in series with the optimized feature vectors acquired following BIANO-based feature selection. Convolutional layers then identify spatial and local time patterns in the multivariate sequences of IoT sensors, which represent a response to an interaction of soil, weather, and crop variables. Recurrent units are then used to process the extracted feature maps to predict temporal dependencies across the stages of crop growth. Lastly, a temporal attention process dynamically assigns more weight to yield-relevant time steps so that the network concentrates on the influential agronomic events and gives less priority to irrelevant fluctuations. The integrated design enables the model to be adaptive to a change in the environment without forgetting contextual memory throughout the entire cropping season.

3.7.1 End-to-end operational workflow of the proposed TA-CRN–BIANO framework

The overall operational sequence of the suggested framework is planned as a highly linear and dependency-based pipeline where each module is going to prepare the input representation that is needed by the next one given in Figure 3. To begin with, the heterogeneous agricultural measurements are obtained using the IoT-enabled sensory and field data that cover soil moisture, soil temperature, ambient humidity, air temperature, rainfall, and crop-condition descriptors. Such raw multivariate streams are subsequently processed by harmonizing data quality, such as missing-value imputation, data outliers, data normalization, and time alignment to convert sensor readings at various times into synchronized time-based streams. This step generates a single agricultural data tensor where a single value set of environmental and crop-state conditions is available at each time. AFC is then used to compress the synchronized sensor matrix to a small nonlinear latent representation. AFC is to reduce redundancy by maintaining higher-order inter-variable dependencies, which are usually completely lost by linear reduction methods on their own. The resultant compressed latent vectors are, in turn, fed to ICA, which further decomposes statistically independent latent sources and minimizes residual correlation between compressed components. That is, nonlinear manifold compaction is done by AFC followed by post-compression source disentangling and decorrelation by ICA. The AFC-ICA cascade is not random, and instead, there is deliberate staging in the way nonlinear information preservation is succeeded by the enhancement of statistical independence. Of such refined latent components, statistical descriptors and higher-order descriptors are obtained, such as the mean, variance, entropy, adaptive weighted skewness, and adaptive weighted kurtosis, thus giving rise to a feature space, which captures agronomic fluctuations of central tendency as well as asymmetric extremity.

Figure 3. Workflow-oriented schematic diagram

The obtained feature pool is then subjected to the proposed BIANO algorithm that does joint feature subset selection and TA-CRN hyperparameter optimization. At this point, the subsets of candidate features and parameter configurations are viewed as developing nutrient-carrying objects, and their fitness is measured based on classification accuracy, compression, and computational efficiency. The best configuration identified by BIANO is then applied to produce temporally ordered multivariate sequences in line with the crop development. These optimized sequences are eventually passed to the TA-CRN classifier, which consists of convolutional layers that learn local short-term temporal interactions, recurrent units that learn long-range agronomic dependencies, and temporal attention that enhances yield-critical phenological intervals. The network gives an end result of a four-class crop-yield categorization (LY, MY, HY, VHY), assessed by performance measures of class-based and macro-based.

3.7.2 Convolutional feature extraction layer

An optimized sequence of input features can be written as

$\mathbf{X}=\left\{ {{\mathbf{x}}_{1}},{{\mathbf{x}}_{2}},\ldots ,{{\mathbf{x}}_{T}} \right\}$    (26)

where, $T$ represents the number of temporal observations and ${{\mathbf{x}}_{t}}\in {{\mathbb{R}}^{{{d}'}}}$ denotes the reduced feature vector at time $t$. A one-dimensional convolution operation is applied to extract local temporal patterns:

${{\mathbf{c}}_{t}}=f(\underset{k=0}{\overset{K-1}{\mathop \sum }}\,{{\mathbf{W}}_{k}}{{\mathbf{x}}_{t-k}}|\mathbf{b})~$    (27)

where, ${{\mathbf{W}}_{k}}$ represents convolutional kernel weights of size $K$, $\mathbf{b}$ is the bias vector, and $f\left( \cdot  \right)$ denotes a nonlinear activation function. This convolutional operation captures short-term dependencies and interactions among agronomic variables, producing discriminative feature maps resilient to sensor noise.

3.7.3 Recurrent temporal dependency modeling

The convolutional feature maps are passed to a recurrent layer in order to capture long-term temporal dependencies during the stages of crop growth. Let c_t denote the convoluted value at time t. The repeated concealed condition is calculated as:

${{\mathbf{h}}_{t}}=\phi \left( {{\mathbf{W}}_{h}}{{\mathbf{c}}_{t}}+{{\mathbf{U}}_{h}}{{\mathbf{h}}_{t-1}}+{{\mathbf{b}}_{h}} \right)$    (28)

where, ${{\mathbf{W}}_{h}}$ and ${{\mathbf{U}}_{h}}$ are learnable weight matrices, ${{\mathbf{b}}_{h}}$ is the bias vector, and $\phi \left( \cdot  \right)$ is a nonlinear activation function. This formulation allows the network to preserve the important historical data and also adjust to slow environmental changes, which could be in vegetative and reproductive phases of growth.

3.7.4 Temporal attention mechanism

Although the temporal information is represented by the recurrent units, all the time steps are not equally valuable in the prediction of the yield. The yield critical periods are brought in by the introduction of a temporal attention mechanism.

${{e}_{t}}={{\mathbf{v}}^{T}}\text{tanh}\left( {{\mathbf{W}}_{a}}{{\mathbf{h}}_{t}}+{{\mathbf{b}}_{a}} \right)$    (29)

where, ${{\mathbf{W}}_{a}}$, ${{\mathbf{b}}_{a}}$, and $\mathbf{v}$ are learnable attention parameters. The normalized attention weight ${{\alpha }_{t}}$ is obtained using a softmax function:

${{\alpha }_{t}}=\frac{\exp \left( {{e}_{t}} \right)}{\mathop{\sum }_{i=1}^{T}\text{exp}({{e}_{i}})}.$    (30)

Context vector c_i, which is the weighted temporal summary, is estimated as:

$\mathbf{c}=\underset{t=1}{\overset{T}{\mathop \sum }}\,{{\alpha }_{t}}{{\mathbf{h}}_{t}}$    (31)

The attention-based aggregation enables the model to focus dynamically on important developmental stages of the crop, such as flowering and grain filling, which have a significant contribution to overall crop production. The final result of the crop y is obtained as a linear transformed version of the attention-weighted version of the context vector:

$\hat{y}={{\mathbf{W}}_{o}}\mathbf{c}+{{\mathbf{b}}_{o}},$    (32)

where, ${{\mathbf{W}}_{o}}$ and ${{\mathbf{b}}_{o}}$ denote output layer parameters. The network parameters are optimized using gradient-based learning to minimize prediction error while ensuring stable convergence.

To improve sensitivity to temporally localized agronomic stress events, two adaptive higher-order descriptors are introduced in the feature extraction stage: adaptive weighted skewness (AWS) and adaptive weighted kurtosis (AWK). Unlike conventional skewness and kurtosis, which treat all temporal observations equally, the proposed descriptors assign greater importance t time steps that are agronomically more informative for yield formation. Let a temporal variable sequence for a given sensing channel be denoted by ${{x}_{t}}$, $t=1\ldots ,T$, and let ${{w}_{t}}$ represent the adaptive importa0, andweight associated with e ${{w}_{t}}\ge 0$ and $\mathop{\sum }_{t=1}^{T}{{w}_{t}}=1$. The weighted mean is defined as ${{\mu }_{w}}=\mathop{\sum }_{t=1}^{T}{{w}_{t}}{{x}_{t}}$, and the weighted standard deviation is ${{\sigma }_{w}}=\sqrt{\mathop{\sum }_{t=1}^{T}{{w}_{t}}{{({{x}_{t}}-{{\mu }_{w}})}^{2}}}$.

Based on these quantities, the adaptive weighted skewness is computed as $AWS=\frac{\mathop{\sum }_{t=1}^{T}{{w}_{t}}{{({{x}_{t}}-{{\mu }_{w}})}^{3}}}{\sigma _{w}^{3}+\varepsilon }$, and the adaptive weighted kurtosis is computed as $AWK=\frac{\mathop{\sum}_{t=1}^{T}{{w}_{t}}{{({{x}_{t}}-{{\mu}_{w}})}^{4}}}{\sigma _{w}^{4}+\varepsilon }$, where $\varepsilon$ is a small stabilizing constant to avoid numerical instability. The adaptive weights are not manually assigned; rather, they are derived from agronomic relevance signals estimated from temporal variability and yield sensitivity. Specifically, the weight at each time step is proportional to the normalized product of (i) local temporal deviation magnitude, (ii) growth-stage importance coefficient, and (iii) feature–yield association strength estimated on the training partition. Thus, ${{w}_{t}}\propto {{\alpha }_{t}}\cdot \mid {{x}_{t}}-\mu \mid \cdot {{\rho }_{t}}$, where ${{\alpha }_{t}}$ is a stage-priority coefficient and ${{\rho }_{t}}$ represents normalized correlation-based yield relevance. This design allows the proposed descriptors to emphasize stress-sensitive intervals such as flowering and grain-filling phases, where asymmetric or heavy-tailed fluctuations are often more predictive of final yield outcomes than uniform temporal statistics.

The suggested TA-CRN model is a synergistic convolutional feature extraction, recurrent temporal modelling and attention-based weighting that are designed to capture short-term agronomic variations and long-term dependence of crop growth. The TA-CRN framework focuses on yield-sensitive time steps and reduces irrelevant variability in a dynamic manner, improving predictive performance, sensitivity to noisy IoT measurements, and interpretability. This renders the proposed model completely applicable in real-time precision agriculture systems of scalable, reliable, and data-efficient crop yield forecasting.

3.8 Feature selection and hyperparameter adaptation using Bio-Inspired Adaptive Nutrient Optimization

In the context of IoT-based crop-yield prediction, feature optimisation is one of the critical steps, especially since despite a dimensionality-reduction step, the latent variables that have survived might still contain partially redundant, weakly discriminative, or time-varying features that eventually degrade the performance of generalisation. To address this problem, the suggested framework will include a BIANO algorithm which is a population-based meta-heuristic inspired by the nutrient acquisition, allocation, depletion, and adaptation mechanisms that can be observed in biological ecosystems and in plant-soil interactions. Within the abstraction used, each candidate solution is represented as a virtual nutrient organism that exists in two coupled states: (i) a binary feature-selection mask that determines the set of retained features, and (ii) a continuous hyper-parameter vector that parameters the TA-CRN architecture (e.g. learning rate, number of convolutional filters, number of recurrent hidden units, dropout, dimensionality of attention). The biological analogy here is that biological organisms that are rich in nutrients are the candidate solutions that are able to garner more predictive value out of the available agronomic environment, and those species that are nutrient-poor are out of competitive relevance and are progressively eliminated. In this way, BIANO can combine feature-subset refinement and hyper-parameter adaptation into a single optimisation loop, as opposed to considering them as separate processes.

Let the $i$- The candidate at iteration $t$ be represented as $X_{i}^{\left( t \right)}=\left\{ M_{i}^{\left( t \right)},H_{i}^{\left( t \right)} \right\}$, where $M_{i}^{\left( t \right)}\in {{\{0,1\}}^{d}}$ denotes the binary mask over $d$ extracted features and $H_{i}^{\left( t \right)}$ denotes the continuous hyperparameter vector. The nutrient state $N_{i}^{\left( t \right)}$ quantifies the adaptive fitness memory of the candidate and is updated according to its predictive effectiveness. For each candidate, the TA-CRN is trained on the selected feature subset and evaluated on the validation set. A multi-objective fitness function is then computed to simultaneously maximize predictive quality and minimize unnecessary complexity:

$\mathcal{F}_{i}^{\left( t \right)}=\alpha \cdot \text{Ac}{{\text{c}}_{i}}+\beta \cdot F{{1}_{i}}+\gamma \cdot \text{Spe}{{\text{c}}_{i}}-\lambda \cdot \frac{\mid {{M}_{i}}\mid }{d}-\mu \cdot \frac{{{\text{ }\!\!\Theta\!\!\text{ }}_{i}}}{{{\text{ }\!\!\Theta\!\!\text{ }}_{\text{max}}}}$    (33)

where, $\text{Ac}{{\text{c}}_{i}}$, $F{{1}_{i}}$, and $\text{Spe}{{\text{c}}_{i}}$ denote validation accuracy, F-measure, and specificity, respectively; $\mid {{M}_{i}}\mid /d$ penalizes excessive feature retention; ${{\text{ }\!\!\Theta\!\!\text{ }}_{i}}$ is the parameter count of the instantiated TA-CRN model; and $\alpha ,\beta ,\gamma ,\lambda ,\mu$ are weighting coefficients satisfying $\alpha +\beta +\gamma =1$ for the predictive component. This formulation is intentionally designed to avoid the common failure mode in metaheuristic model tuning, where accuracy improves only by increasing feature dimensionality or network size.

The nutrient state is updated using a gain–decay adaptation rule that balances exploitation of high-performing solutions with controlled retention of diversity:

$\begin{gathered}N_i^{(t+1)}=(1-\rho) N_i^{(t)}+\eta \cdot \max \left(0, \mathcal{F}_i^{(t)}-\overline{\mathcal{F}}^{(t)}\right)- \delta \cdot \max \left(0, \overline{\mathcal{F}}^{(t)}-\mathcal{F}_i^{(t)}\right)\end{gathered}$    (34)

where, $\rho$ is the memory decay coefficient, $\eta$ is the nutrient gain factor, $\delta$ is the nutrient depletion factor, and $\overline{\mathcal{F}}^{(t)}$ is the population-average fitness at iteration? Those candidates with a performance above average are gaining nutrient deposits and, consequently, a greater chance of passing their segments of features and hyperparameters to the following generation.

On the other hand, weak performers are deprived of nutrients, which makes them more susceptible to substitution, mutation, or structural simplification. Such a nutrient-memory mechanism is what distinguishes BIANO as compared to traditional fixed meta-heuristics in that it incorporates a historical adaptive bias as opposed to using instantaneous fitness rankings. BIANO uses three operations inspired by biology to evolve the population: nutrient-guided assimilation, competitive pruning, and adaptive mutation. During nutrient-guided assimilation, feature indices and hyper-parameter components of elite candidates are inherited by sub-optimal candidates with a probability distribution proportional to normalised nutrient strength, further exacerbating exploration at high-fitness local optima.

In competitive pruning, low-contribution features are eliminated with a probability that is negatively proportional to their prior selection frequency and local significance to the current candidate, and this tends to favor parsimonious representations. In adaptive mutation, the feature mask and hyperparameter vector are mutated inversely proportional to the existing population diversity; when this diversity is low, a mutation strength is amplified to leave local maxima, and when it is high, a mutation is suppressed to maintain convergence stability. Termination of the algorithm can occur when the maximum number of iterations or global best fitness is not improving with a small tolerance over a specified patience window.

The downstream training and evaluation of the final TA-CRN model are then realised using the resulting optimised solution. In the methodological perspective, BIANO has an advantage over traditional wrapper-based feature selection in that it explicitly co-optimises the prediction performance, compactness of representation, and complexity that is concerned with deployability. It is especially relevant in the context of the agricultural IoT systems, where sensor redundancy, non-stationarity of the environment, and the impossibility of deep models to be deployed in real-time may compromise otherwise precise models. BIANO can be used to trade off nutrient-memory adaptation with a joint discrete-continuous optimisation to yield a feature-parameter configuration, which can be more discriminative and more stable, as well as operationally suited to precision agriculture settings.

3.9 Model training and validation

Following BIANO-based feature optimization, the retained feature vectors are reorganized into temporally ordered multivariate sequences aligned with crop growth progression. Although the original agronomic endpoint is continuous harvested yield measured in kg/ha, the present study formulates the prediction problem as a multi-class yield categorization task to support operational decision-making in precision agriculture. Specifically, the continuous yield values are discretized into four agronomically interpretable categories—Low Yield (LY), Medium Yield (MY), High Yield (HY), and Very High Yield (VHY)—using percentile-based thresholds estimated exclusively from the training partition to avoid information leakage. This formulation is practically meaningful because field interventions are typically triggered by categorical risk or productivity zones rather than by exact scalar yield values alone, especially in real-time advisory systems involving irrigation scheduling, nutrient prioritization, and stress mitigation.

The TA-CRN model is trained in a supervised classification setting using these yield categories as ground-truth labels. Convolutional layers first capture short-range temporal interactions among multivariate IoT variables, recurrent units then model long-term growth dependencies across phenological stages, and the temporal attention mechanism dynamically amplifies yield-critical intervals such as reproductive and grain-filling phases. The network parameters are optimized by minimizing the categorical cross-entropy loss:

${{\mathcal{L}}_{CE}}=-\frac{1}{N}\underset{i=1}{\overset{N}{\mathop \sum }}\,\underset{c=1}{\overset{C}{\mathop \sum }}\,{{y}_{ic}}\text{log}\left( {{{\hat{y}}}_{ic}} \right)$    (35)

where, ${{y}_{ic}}$ denotes the true one-hot encoded class label and ${{\hat{y}}_{ic}}$ denotes the predicted posterior probability for class $c$. Adaptive learning-rate optimization, early stopping, dropout regularization, and weight decay are used to ensure stable convergence and improved generalization under noisy agricultural sensor conditions.

Validation is conducted on unseen data to quantify classification robustness and reduce the risk of overfitting. Since the task is explicitly framed as categorical yield prediction, the primary evaluation metrics include accuracy, precision, sensitivity (recall), specificity, F-measure, macro-F1, and confusion-matrix-based class separability. In addition, to preserve agronomic continuity with the original scalar endpoint, the manuscript may optionally report an auxiliary regression-style consistency analysis by assigning class centroids to predicted categories and computing approximate RMSE or MAE; however, the principal experimental results and all benchmark comparisons in this study are interpreted within the multi-class classification framework Because the forecasting problem is formulated as a four-class agronomic yield categorization task (LY, MY, HY, and VHY), classification-oriented metrics are used as the primary evaluation criteria, and all metrics are reported in macro-averaged form where applicable to avoid dominance by majority classes (Table 3).

Table 3. Performance metrics analysis

Despite the fact that the framework proposed showed good and consistent predictive behaviour on the current dataset, the issue of generalisation should be viewed within the framework of the experimental design. The data corresponds to a single crop (rice) and a single production geographic location; the intra-domain robustness is, in turn, the only validation provided by the current studies, as opposed to inter-domain generalisation. The model was partially addressed by this limitation through a variety of train-test partitions, seasonal growth phases, noise levels, and repeated runs, all of which demonstrated low performance variance and consistent convergence. These tests suggest that the suggested TA-CRN-BIANO model does not over-fit a particular separation of the information, but is training identifiable temporal-agronomic tendencies within the obtainable area. Nonetheless, due to the strong dependence of the dynamics of crop yields on genotype, soil texture, irrigation plan, anomalies in seasonal rainfall, and the climate signals in the region, external validation on geographically diverse and non-temporal data is still necessary before a claim of universal transferability. The good within-domain accuracy demonstrates the viability of the offered solution in deploying precision agriculture in similar sensing and farming conditions in the future, yet the framework should be expanded by future studies to the federated or transfer-learning environments based on the data from several agro-climatic zones and the year of cultivation. This extension would enable the direct measure of domain -shift resilience and would also enhance the translational applicability of the proposed approach.

In addition to the existing scope of the experiment, there are numerous extensions that are significant towards translational maturity. First, to be able to measure resilience in the domain of domain shift and agro-climatic heterogeneity, multi-year, multi-region, and multi-crop multi-validation is necessary. Second, future work must consider the framework in explicit cross-domain transfer conditions, such as transfer learning, domain adaptation, and federated agricultural learning conditions, where local farms provide distributed sensor intelligence with no centralized pool of raw data. Third, although the present study demonstrates favorable inference efficiency, additional compression strategies such as pruning, quantization, and edge-aware neural architecture refinement could further improve deployment on resource-constrained farm gateways. These directions will help convert the current proof-of-concept methodology into a broadly scalable and operationally robust precision agriculture decision-support system.