Carbon Dioxide Emission Prediction for Trend Analysis using CAViaR Pine Cone Optimization Algorithm Enabled Attention-Based LSTM

Climate change, which is driven mainly by the rise in greenhouse gases, like CO₂, is one of the most urgent global challenges today. Human activities, such as burning fossil fuels, deforestation, industrial processes, agriculture, urbanization, transportation, mining, and waste disposal release large amounts of CO₂ into the atmosphere, which is the leading cause of global warming (IPCC, 2021) [1]. This increase in CO₂ leads to rising temperatures, environmental damage, more frequent extreme weather events, and serious risks to human health and ecosystems [2]. To address this, global greenhouse gas emissions must be reduced by at least 50% by 2050, with a strong emphasis on cutting CO₂ emissions, even as energy demands continue to grow [1]. Essential strategies include adopting renewable energy, improving energy efficiency, and deploying carbon capture technologies [3]. Accurate forecasting and continuous monitoring of emissions support effective policymaking and track progress towards targets [1]. Although natural sources, such as organic decomposition, ocean release, and volcanic activity contribute some CO₂, their impact is small compared to anthropogenic emissions [2]. Addressing these challenges is essential for a resilient and sustainable future; moreover,

This work is open access and licensed under the Creative Commons CC BY 4.0 License.

Satellite imagery significantly contributes by offering critical geographical data. Thus, it supports environmental monitoring, disaster management, and informed decision-making in response to climate change and other global environmental issues [4].

Many studies have investigated global CO₂ emissions across both developed and developing nations using a variety of statistical methods. These studies often identify a strong connection between rising CO₂ levels and increased energy use driven by economic growth and development. Scarcity of long-term data is a key challenge, which is essential for accurate trend analysis over several decades. Short data records and small sample sizes hinder the ability to detect meaningful changes and interpret trends reliably. Furthermore, detailed historical and sector-specific emissions data are limited [2]. Understanding long-term CO₂ emission patterns is vital for forecasting economic growth and creating effective strategies to reduce emissions while safeguarding public health and the environment [5] [3]. While several forecasting techniques exist, such as econometric models, input-output models, and surveys, trend analysis is widely favored for its simplicity and dependence on available data [3]. Moreover, trend analysis was used to project energy-related CO₂ emissions for the top 25 countries and global total between 1971 and 2007. Emission data from the International Energy Agency were analyzed using least squares regression to ensure statistical reliability [3]. Hence, it provides dependable projections to aid in policy formulation and environmental planning.

The main intention is to develop a novel approach for predicting CO 2 emissions on Trend Analysis in Satellite Image Time Series using CPCO_ALSTM. For trend analysis, the satellite image of Query region is considered for each interval as input, and a Kalman filter accomplishes preprocessing. Then, segmentation of satellite image is done using DRRP-Net, wherein DRRP-Net is developed by merging DRINet and RP-Net. Then, estimation of carbon emissions is performed. Moreover, estimated carbon emission amount is compared with original CO 2 emission from database for accuracy. Here, Trend Analysis is performed using ALSTM, which is trained using CPCO. Here, the integration of PCO and CAViaR develops CPCO.

Major contribution of this research is mentioned below.

• • Devised CPCO_ALSTM for CO2 emission prediction on Trend Analysis: An innovative model called CPCO_ALSTM is devised to predict CO2 Emission. In this work, prediction of CO₂ emissions using trend analysis is performed using ALSTM. Here, ALSTM is trained using CPCO, the integration of PCO and CAViaR develops CPCO.

2.    Motivation

• 2.1.    Literature Review

The layout of the following sections is as follows: A review of the literature on existing methods for CO₂ emission prediction is provided in Section 2; Methodology and architecture of the devised method are detailed in Section 3; Evaluation outcomes and ablation study are detailed in Section 4; Additionally, in Section 5, conclusion and future work are summarized.

Development of a new model for predicting CO₂ emissions using trend analysis is driven by the demand for precise, long-term forecasts that can inform effective climate policies. Many current models are constrained by limited data availability or require complex assumptions. By using a straightforward, data-focused trend analysis, more reliable predictions can be made even with minimal data. This method seeks to improve policy decisions related to emission reduction and environmental management.

• J. D. L. Brazidec, et al. [6] developed a CO₂ emission prediction method named Convolutional neural networks (CNN). This method effectively modified CO₂ plumes from power, although it struggled to account for outliers, such as plants with exceptionally high emissions. A. Danjou, et al. [7] introduced a method for predicting CO₂ emissions namely, Decision tree learning. The model accurately segmented the images for estimation purposes; However, it failed to collect more spatial variability of CO₂ emissions. B. F. Andrade, et al. [8] presented a Cross-sectional mass balance approach for predicting CO₂ emission. It eliminated the comparison between bottom-up and top-down evaluations but overlooked local-scale emissions, leading to inaccuracies in the overall emission estimation. S. N. Wren, et al. [9] developed FNO-DenseNet based on Fourier neural operator (FNO) for carbon emission prediction. It was easily scalable to larger regions, but it lacked the ability to predict Soil Organic Carbon (SOC) at fine resolutions.

• 2.2.    Challenges

Main challenges experienced by traditional methods on CO 2 Emission prediction are mentioned as follows.

• •    The CNN model presented in [6] effectively addressed missing data due to cloud cover. Still, it did not incorporate the crucial classification of high-frequency regions, such as urban areas.

• •    In [7], the developed model improved insights into urban CO₂ emissions; however, it omitted the use of satellite image-based emission estimation.

• •    Although the model in [8] showed promise for monitoring CO₂ emissions from specific power plants, the limited

spatial resolution of satellite imagery restricted its ability to capture detailed local emission patterns.

• •    Reliable and accurate CO2 emission data is essential for effective prediction, yet data availability and quality often pose challenges, particularly in developing regions with limited monitoring infrastructure.

3.    Proposed Conditional Pine Cone Optimization enabled Attention-Based Long Short-Term Memory for CO2 Emission prediction using Trend Analysis

• 3.1.    Satellite image Acquisition from Bhoonidhi and Climate Watch dataset

Analyzing satellite images has become an essential method for tracking environmental changes, urban growth, deforestation, and agricultural developments over time. While traditional models have been commonly applied for trend detection, they often face challenges due to the complex, nonlinear, and noisy nature of satellite data. These conventional techniques usually rely on assumptions of stationarity, struggle with missing data and seasonal effects, and have limited ability to detect subtle or rapid changes in land conditions. Hence, innovative model is developed for CO 2 Emission prediction on Trend Analysis in Satellite Image Time Series using CPCO_ALSTM. In this work, CO 2 is not directly detected from satellite images. Instead, a specific geographic area is cropped from the satellite image, and the corresponding CO 2 emission values for that region are obtained from the dataset. For trend analysis, at every time interval , and , the satellite image of Query region is considered as an input and is preprocessed by Kalman filter. After that, segmentation of satellite image is done using DRRP-Net, wherein DRRP-Net is developed by merging DRINet and RP-Net. The segmentation is done into major land-cover classes, such as buildings, roads, vegetation, water bodies, and open land. Then, carbon emissions estimation is performed. Moreover, carbon emissions are estimated and compared with original CO 2 from the database for accuracy. Here, Trend Analysis is performed using ALSTM, which is trained using CPCO. Here, the integration of PCO and CAViaR develops CPCO. Additionally, schematic structure of developed CPCO_ALSTM to predict CO 2 emission on Trend Analysis is illustrated in Fig. 1

Fig. 1. Schematic structure of developed CPCO_ALSTM for CO 2 emission prediction on trend analysis

The Bhoonidhi dataset [10] is a comprehensive collection of high-resolution satellite imagery used for land use and land cover mapping in India. Here, at the dataset , the images at the interval are given as,

H = { H 1 f , h 2 ,..., H^f f ,..., H }

where, the total images in the dataset H at f th interval is indicated as H f , randomly selected qth sample from dataset is implied as H? . Similarly, at time interval f +1 and f - w the randomly chosen image is considered as H^+1 and Hq w respectively. Additionally, the Climate Watch dataset [11] offers extensive worldwide information on greenhouse gas emissions, climate policies, and environmental metrics to aid in climate research and decision-making, and at f th interval, it is expressed as, p={ Pif, P2f,..., pq,..., pf}                                             (2)

Here, Climate Watch dataset is signified as P , total images in the dataset is represented as P^f , and the q^th image that is chosen randomly from dataset is implied as P_q^f . Likewise, it is processed for every time interval. In this work, the sample at time H ^f is selected for further analysis.

• 3.2.    Image Preprocessing at Every Time Interval Using Kalman Filter

At every time interval, images undergo preprocessing to improve their quality and make them suitable for further analysis. Here, image H_q^f is preprocessed using Kalman filter as it effectively reduces noise and provides optimal estimates in dynamic and time-varying images.

A. Kalman Filter

The Kalman filter [12] is used for pre-processing. The cropped satellite image region is first converted into a structured representation, and the Kalman filter is used as a recursive noise-reduction estimator to smooth measurement inconsistencies across spectral bands and acquisition noise. It is calculated using below expression,

4= whih( bwhih + oh )—1

*

uh = uh + Ah (vh - I huh)

WhS =(1 — Ah Ih) Wh(5)

where, Kalman gain is indicated as A * , forecast error covariance matrix is signified as W^z , observation error covariance matrix is implied as O_h , analysis error covariance matrix is indicated as W s , observation operator is denoted as I ^ , prediction error covariance matrix is implied as W^z , predicted value is represented as u^z , analysis value is implied as u ^a , and observation is signified as v . Then, the obtained preprocessed image at f ^th interval is indicated as R_q^f , similarly, R S ^{+ 1} at ( f + 1 ) ^th interval, and R ^f ^w at ( f - _w ) ^th interval.

• 3.3.    Satellite Image Segmentation at Every Time Using DRRP-Net

At each time interval, satellite images are segmented to identify and separate distinct regions, enabling precise analysis of land surface variations. Here, the preprocessed image R_q^f is allowed for segmentation of satellite image using DRRP-Net, whereas DRRP-Net is developed by integrating DRINet with RP-Net. The prototype-learning mechanism of RP-Net is adopted to improve class representation and boundary discrimination in heterogeneous land-cover regions (e.g., buildings, roads, vegetation, and water). The DRINet is employed to enhance multi-scale feature extraction, which is important because satellite imagery contains objects of highly varying spatial scales. The residual and dense connections improve gradient flow and feature reuse, reducing information loss. Additionally, network structure of DRRP-Net is illustrated in Fig. 2.

Fig. 2. Network structure of DRRP-Net

In RP-Net [13] model, the input image H_q^f is utilized as an input. Additionally, the outcome m^f is expressed as,

( a^ , X )

m1 f ( ,e) = argX max msoft where, spatial position of binary mask is indicated as (a, в, x) , soft max is signified as msoft. Then, the output of RP-Net model m f along with preprocessed image Rqf is fed into DRRP-Net layer, and is formulated below.

^V 3          *                         ( a , вх )

m 2 = ² ^ ^ ^ ^Rq ( MS , £ ) ^ ( MS , ^ - arg X max m soft

M = 1 S = 1 ^£ = 1

where, the preprocessed image’s size is indicated as 3 and v , weight is implied as T, and Rq implies the preprocessed image. In the DRINet Model [14], the resultant of DRRP-Net layer mf with input image Hqf is proceeded, and its expression is evaluated below, f m

( Ф — 3 ) V

= ^ ² ££ x R l M = 1 S = 1 ^s = ¹

q q ⁽ ms^ )

( a , в , X )

* ^T ( vs^) ^{- ar}g x max m soft

Thus, the resultant segmented image obtained from DRRP-Net at f th interval is indicated as m f . Likewise, the outcome attained at (f +1)th interval is mf+1, and at (f - w)th interval is mf

.

• 3.4.    Carbon Estimation at Regular Time Intervals for CO₂ Emission Analysis

Carbon estimation [15] at regular time intervals enables accurate analysis and monitoring of CO₂ emissions over time. Carbon emission estimation calculates the greenhouse gases produced by transport, energy consumption, and industrial processes. This process supports environmental monitoring and sustainable decision-making.

C = i N -ⁿ d [ JCO d = 1

Here, CO₂ emissions derived from classified land areas are indicated as C , the average values of different land use types are represented as N_d and П^. This involves applying specific carbon emission coefficients and accounting for

JCO   44

the number of occurrences of each land use category, where, ---- 2 = — = 3 667 . The result generated from this stage

JC   12

reflects the overall carbon balance. Then, the obtained carbon estimated outcome is P q at f^th interval, similarly, P q ^{+ 1} at ( f + 1 ) ^th interval, and P q ^w at ( f - _w ) ^th interval.

• 3.5.    Trend Analysis

Trend analysis of carbon emission estimation involves examining emission data over time to identify consistent patterns or changes in greenhouse gas outputs. In this work, trend analysis is carried out using an ALSTM model, which is trained with the CPCO. Furthermore, CPCO is a hybrid approach that integrates CAViaR with PCO to enhance training performance.

• A.    ALSTM

The ALSTM model [16] employs a neural network to predict the nonlinear seismic behavior of bridges by incorporating a temporal attention mechanism into a traditional encoder-decoder framework. In this ALSTM model, the CO₂ emissions estimated outcome P_q^f is fed as an input. This attention mechanism enables the network to concentrate on the most important features within time-series data. The architecture comprises six main components: an input, encoder, attention, decoder, and output phase. Encoder as well as decoder contains LSTM layers followed by two fully connected layers, which are utilized by attention block with activation function. ReLU activation and dropout layers (rate of 0.2) are applied to prevent overfitting and vanishing gradients. The ALSTM model’s structure is illustrated in Fig. 3.

Fig. 3. Structure of ALSTM

Moreover, the output attained from ALSTM network is indicated as V q at f^tth interval, likewise, V q ⁺ 1 at ( f + 1 ) ^th interval, and V ^ - w at ( f - w ) ^th interval.

• B.    Tuning of ALSTM using CPCO

Tuning is performed to train the network for increased accuracy and reliability of CO₂ emission estimation. Here, ALSTM is trained using CPCO, which is the combination of CAViaR with PCOA [17]. PCOA is based on the behavior of pine cones to solve complex optimization issues. The conditional autoregressive structure of CAViaR is adapted as a dynamic updating mechanism to control the optimization search behavior under varying error distributions during hyperparameter tuning. Specifically, it is used to model conditional variations in the loss function, thereby improving stability and preventing extreme parameter fluctuations during optimization.

• •    Solution Encoding: It is used to represent a potential solution by defining its position within a specified search space. Additionally, it is expressed as so ln = [ 1 x1 ] , where hyperparameter of ALSTM is indicated as.

• •    Evaluation of fitness: It is the discrepancy between predicted value and actual target value to determine optimal

result. The calculation is carried out as follows,

n fitness = ~^(Vgf * - Vqf ) n qr

where, ц indicates entire images in database, targeted outcome is denoted as V qf * and actual outcome is indicated as V _q ^f . Additionally, CPCO method’s algorithmic steps are as follows.

Step 1. Population Initialization

During initialization, pine trees and cones are used to divide the problem space into segments. In this space, trees are located at the middle of each segment and cones are initialized around them within defined boundaries

Step 2: Exploitation Phase

In this phase, the search space is partitioned into super-cubes and this process imitates the behavior of cones dispersing as they fall under the influence of gravity.

Step 3: Exploration Phase

To mimic the pollination of pine tree, PCOA uses two distinct mechanisms modeling agents from male cones reaching female cones within or between trees, represented by the following formulation.

yB neW = yB. .+ 0.5 x^  x trbStt - - tB„ J + 0.5 x^ x trbSst - - vB. Д                 (11)

i, j             i, j                   ^1               i         ^1, j                     ^3               i          ^3, j where, position of cone is indicated as yB , index of cone is represented as i , best position of tree is indicated as trbst , index of tree is implied as j , and chance of successful pollination is indicated as.

By CAViaR, pt bb-J = ^0 +Ё ^7 +L Cp p( yBj)

=1

Let p = t = 2, yB-: = £^+£ B-1 + £ 2bB--2 + £ P yBn-1) + £ 2p( yB"- 2 )(13)

i,j         0        1 i,j          2 i,j           1             i,j            2

Substituting Eq. (13) in Eq. (11), yB-ew = (t0 + £ yyBf-1 + £ yBn--2 + £ p( yB"-1) + £ 2p( yB"- 2)) i,j             0        1 i,j          2 i,j           1             i,j            2

+0.5 x^ x ttrbst- - IB. J + 0.5 x^ x trrbst- - yB., J ^1              i        ^1, J                   ^3              1s where, is a constant, position of cone is indicated as yB , index of cone is represented as i , best position of tree is indicated as trbst , index of tree is implied as j , and successful pollination’s chance is implied as.

Step 4: Animals dispersing pine cones

Inspired by the natural behavior of animals dispersing pine cones, the algorithm employs four operators. One of these involves relocation of a cone by a squirrel from a random location to an optimal position, which is guided by quadratic programming to enhance solution quality.

Step 5: Assigning cones to every tree

Each tree adjusts the positions of nearby cones to realistically mimic the natural pattern of falling. Over successive generations, this process attracts nearby cones and allows well-positioned trees to gather more cones

Step 6: Fitness Re-calculation

CPCO algorithm continues to iterate until optimal solution is achieved by Eq. (10).

Step 7: Termination

In termination step, CPCO algorithm concludes once optimal solution has been successfully obtained.

• 3.6.    Accuracy Calculation

4.    Result and Discussion

• 4.1.    Experimental Setup

• 4.2.    Dataset Description

• 4.3.    Evaluation Metrics

To evaluate the accuracy of CO2 emission estimates, the predicted values are compared against actual data obtained from a reliable database. This evaluation is performed at regular time intervals, allowing for continuous monitoring and prediction of CO2 emissions over time.

The devised CPCO_ALSTM for predicting CO 2 emission for Trend Analysis is evaluated, and its dataset description is mentioned in this part.

The CPCO_ALSTM model for CO₂ emission prediction and trend analysis is implemented using the Python tool.

In this study, the satellite imagery is obtained from the Bhoonidhi geo-portal, which is maintained by the Indian Space Research Organization (ISRO) [10]. Specifically, Level-2 products from the Resourcesat-2 mission are used, including LISS-3 and LISS-4 (MX70) sensors. The Resourcesat-2 LISS-3 (L2) data provide a spatial resolution of 23.5 meters and include four spectral bands: Green (0.52–0.59 µm), Red (0.62–0.68 µm), Near Infrared (0.77–0.86 µm), and Shortwave Infrared (1.55–1.70 µm). The Resourcesat-2 LISS-4 (MX70, L2) data offer a higher spatial resolution of 5.8 meters and include three spectral bands: Green (0.52–0.59 µm), Red (0.62–0.68 µm), and Near Infrared (0.77–0.86 µm) [11]. The combination of moderate (23.5 m) and high (5.8 m) spatial resolution imagery enables detailed land-cover segmentation of buildings, roads, vegetation, water bodies, and open land areas. The inclusion of Near Infrared and Shortwave Infrared bands enhances vegetation discrimination and surface material characterization, which are essential for extracting emission-related spatial proxies used in indirect CO₂ emission estimation.

To analyze the efficiency of devised CPCO_ALSTM for CO 2 Emission prediction, metrics like normalized MSE, accuracy, and Jaccard coefficient are utilized.

• A.    Normalized MSE

Normalized MSE is used to measure the difference between the predicted and actual continuous CO₂ emission values after normalization.

• B.    Accuracy

Accuracy is used to evaluate the correctness of emission range prediction after dividing the continuous CO₂ values into 10 emission bins. It measures the proportion of samples for which the predicted emission range matches the actual emission range.

• C.    Jaccard Coefficient

• 4.4.    Comparative Methods

• 4.5.    Comparative Assessment

The Jaccard Coefficient is used to measure the similarity between the predicted and actual emission classes after dividing the CO₂ values into 10 bins emission ranges. It is calculated as the ratio of correctly predicted samples to the total unique samples present in either the predicted or actual class.

CNN [6], Global adaptive-mesh model [7], Cross-sectional flux method [8], FNO-DenseNet [9], UNet-FCNN, and DRRP-Net are the comparative methods that are used to analyze the performance of CPCO_ALSTM.

The efficiency of the developed CPCO_ALSTM is evaluated using various metrics by varying delay and training data.

• A.    Assessment based on Delay

Evaluation of CPCO_ALSTM with comparative techniques, like CNN, Global adaptive-mesh model, Crosssectional flux method, FNO-DenseNet, UNet-FCNN, and DRRP-Net, by changing delay from 5 to 25 years, is illustrated in Fig. 4. In Fig. 4 a), Analysis of CPCO_ALSTM with respect to accuracy is illustrated. For delay of 25 years, accuracy of CPCO_ALSTM is 96.16%, while other traditional approaches have achieved an accuracy of 84.13%, 87.16%, 88.65%, 90.74%, 92.88%, and 94.06%. Fig. 4 b) illustrates the analysis of CPCO_ALSTM in normalized MSE. For delay = 25 years, conventional methods have reached normalized MSE of 0.431, 0.386, 0.326, 0.268, 0.202, and 0.115. Moreover, CPCO_ALSTM has achieved a low normalized MSE rate of 0.095. In addition to this, Fig. 4 c) shows the evaluation of CPCO_ALSTM in terms of Jaccard Coefficient. Here, CPCO_ALSTM has reached 0.955 Jaccard coefficient, whereas, traditional approaches have attained 0.854, 0.871, 0.883, 0.891, 0.913, and 0.932 of Jaccard Coefficient value.

(c)

Fig. 4. Comparative evaluation of CPCO_ALSTM based on delay, a) Accuracy, b) Normalized MSE, c) Jaccard Coefficient

• B.    Assessment based on Training Data

  Fig. 5 shows assessment of CPCO_ALSTM with traditional methods, such as CNN, Global adaptive-mesh model, Cross-sectional flux method, FNO-DenseNet, UNet-FCNN, and DRRP-Net, by changing delay. Moreover, 90% training data is considered in this assessment. Fig. 5 a) portrays the analysis of CPCO_ALSTM with respect to accuracy. The accuracy attained by CPCO_ALSTM is 96.35%, while baseline methods have achieved an accuracy of 85.11%, 87.37%, 88.78%, 90.93%, 92.94%, and 94.35%, respectively. Assessment of CPCO_ALSTM for normalized MSE is displayed in Fig. 5 b). Conventional methods have reached normalized MSE of 0.331, 0.301, 0.275, 0.221, 0.187, and 0.104. Additionally, CPCO_ALSTM has achieved a low normalized MSE rate of 0.087. Fig. 5 c) illustrates the evaluation of CPCO_ALSTM in accordance with Jaccard Coefficient. Here, CPCO_ALSTM has reached 0.957 Jaccard coefficient, however, traditional approaches have attained 0.855, 0.866, 0.874, 0.897, 0.908, and 0.936 of Jaccard Coefficient value.

  □ZL CNN                    [Л FNO-DenseNet m DRRP Net          I                         CXI CNN                    S FNO-DenseNet EZZ1 DRRP Net

  

  m CNN                   ГЛ1 FNO-DenseNet m DRRP-Net

  
  
  

  (c)

  
  

• 4.6.    Algorithmic Assessment

Fig. 5. Comparative evaluation of CPCO_ALSTM based on training data, a) Accuracy, b) Normalized MSE, c) Jaccard Coefficient

The algorithmic performance of CPCO_ALSTM is compared with Adam_ALSTM [18], RMSProp_ALSTM [19], and Bayesian Optimization_ALSTM [20] and the results are provided in Fig. 6. The results obtained on population size 50 are discussed in this section. Fig. 6 a) portrays the analysis with respect to accuracy. The accuracy attained by CPCO_ALSTM is 95.64%, while Adam_ALSTM, RMSProp_ALSTM, and Bayesian Optimization_ALSTM have achieved an accuracy of 92.74%, 93.63%, and 94.32%. Assessment for normalized MSE is displayed in Fig. 6 b). The Adam_ALSTM, RMSProp_ALSTM, and Bayesian Optimization have reached normalized MSE of 0.192, 0.150, and 0.112 and the CPCO_ALSTM has achieved a low normalized MSE rate of 0.090. Fig. 6 c) illustrates the evaluation in accordance with Jaccard Coefficient. Here, CPCO_ALSTM has reached 0.958 Jaccard coefficient, however, Adam_ALSTM, RMSProp_ALSTM, and Bayesian Optimization_ALSTM have attained 0.914, 0.929, and 0.940 of Jaccard Coefficient value.

10     15    20    25     30     35     40     45     50

Population Size

(a)                                                              (b)

(c)

Fig. 7. Algorithmic evaluation of CPCO_ALSTM based on population size, a) Accuracy, b) Normalized MSE, c) Jaccard Coefficient

Table 1. Comparative discussion of CPCO_ALSTM for CO 2 emission estimation

Variation	Metrics	CNN	Global adaptive-mesh model	Crosssectional flux method	FNO-Dense Net	UNet-FCNN	DRRP-Net	Proposed CPCO_ALSTM
Delay= 100s	Accuracy (%)	84.13	87.16	88.65	90.74	92.88	94.06	96.16
	Normalized MSE	0.431	0.386	0.326	0.268	0.202	0.115	0.095
	Jaccard Coefficient	0.854	0.871	0.883	0.891	0.913	0.932	0.955
Training Data 90%	Accuracy (%)	85.11	87.37	88.78	90.93	92.94	94.35	96.35
	Normalized MSE	0.331	0.301	0.275	0.221	0.187	0.104	0.087
	Jaccard Coefficient	0.855	0.866	0.874	0.897	0.908	0.936	0.957

• 4.7.    Comparative Discussion

5.    Conclusions

Table 1 demonstrates CPCO_ALSTM method’s comparative discussion for CO 2 Emission estimation in which efficiency of CPCO_ALSTM is compared with competitive methods, like CNN, Global adaptive-mesh model, Crosssectional flux method, FNO-DenseNet, UNet-FCNN, and DRRP-N. It is performed by varying delay and training data. Here, conventional approaches have attained accuracy of 89.42%, 90.91%, 92.41%, and 94.27%, whereas CPCO_ALSTM has attained 96.35% of accuracy. Additionally, 0.331, 0.301, 0.275, 0.221, 0.187, and 0.104 are error values attained by traditional approaches. Moreover, CPCO_ALSTM has reached a low normalized MSE rate of 0.087. Additionally, CPCO_ALSTM has achieved a Jaccard Coefficient value of 0.957. However, other techniques have reached

Jaccard Coefficient values of 0.855, 0.866, 0.874, 0.897, 0.908, and 0.936. This demonstrates that CPCO_ALSTM for CO₂ emission estimation method has strong reliability, precision, and effectiveness in identifying emission patterns.

Estimating CO₂ emissions is crucial for forecasting pollution levels and supporting climate-related strategies and decisions. Conventional methods often struggle with poor prediction accuracy, lack flexibility, and are ineffective in managing rapidly changing emission patterns. Here, CPCO_ALSTM is designed to predict CO 2 emission on Trend Analysis in Satellite Image. At every time interval, the input image is acquired and is preprocessed using Kalman filter. Thereafter, segmentation of satellite image is done using DRRP-Net, wherein DRRP-Net is developed by merging DRINet with RP-Net. Then, carbon emissions estimation is achieved. Moreover, the estimated amount of carbon emissions is compared with the actual CO₂ emissions from the database to assess accuracy. Here, Trend Analysis is performed using ALSTM, which is trained using CPCO. Here, the integration of PCO and CAViaR develops CPCO. Additionally, CPCO_ALSTM has attained 96.35% of accuracy, 0.087 of normalized MSE, and 0.957 of Jaccard Coefficient. Future work lies in enhancing model accuracy, scalability, and adaptability by integrating advanced optimization to better capture dynamic and complex emission behaviors.

All the Declarations and StatementsAuthor Contributions Statement

Choudari Lakshmi: Conceptualization, Methodology, Algorithm Development, Software Implementation, Data Curation, and Writing – Original Draft Preparation.

All authors have read and agreed to the published version of the manuscript.

Conflict of Interest Statement

The authors declare no conflicts of interest.

Funding Declaration

This research received no external funding.

Data Availability Statement

The dataset used in this study is obtained from publicly available datasets, namely the Bhoonidhi geo-portal maintained by ISRO and the Climate Watch dataset. These datasets are accessible for research purposes.

Ethical Declarations

None

Acknowledgments

We would like to thank GITAM (Deemed to be University), Visakhapatnam, for providing the necessary support and facilities to carry out this research work. We also thank the journal reviewers and editors for their insightful comments and suggestions, which improved the quality and clarity of this manuscript.

Declaration of Generative AI in Scholarity Writing

The authors declare that no generative AI or AI-assisted technologies were used in the preparation of this manuscript. All content, including text, figures, and analysis, were produced directly by the authors.

Abbreviations

The following abbreviations are used in this manuscript:

CO₂ - Carbon Dioxide

CPCO - Conditional Pine Cone Optimization

ALSTM - Attention-Based Long Short-Term Memory

CPCO_ALSTM - Conditional Pine Cone Optimization enabled Attention-Based Long Short-Term Memory

DRRP-Net - Dense-Res Recurrent Prototypical Network

DRINet - Dense-Res-Inception Network

RP-Net - Recurrent Prototypical Network

PCO - Pine Cone Optimization

CAViaR - Conditional Autoregressive Value at Risk

CNN - Convolutional Neural Network

FNO - Fourier Neural Operator

MSE - Mean Squared Error

Appendix

None.