HE-DFNETS: A Novel Hybrid Deep Learning Architecture for the Prediction of Potential Fishing Zone Areas in Indian Ocean Using Remote Sensing Images

The Indian subcontinent is known for its larger coastline spanning, over 8100 km and is considered the habitat for many millions of people. The livelihood of their habitat is purely dependent upon the fishing activities. Often, the search for fish requires more time for catching and more resources, thus increasing the operational cost leading to low profitability. With the advent of artificial intelligence algorithms, designing intelligent algorithms for an effective prediction of fishing areas has reached new heights in terms of high accuracy (Acy) and less time. But still, predicting the location of potential fishing zones (PFZs) is always a daunting task. To reduce these issues, this work presented the novel hybrid prediction architecture of PFZs using remote sensing images. The proposed architecture integrates the deep convolutional layers and flitter bat optimized long short-term memory (FB-LSTM)-based recurrent neural networks (RNN). These convolutional layers are utilized to remove the various color features such as chlorophyll, sea surface temperature (SST), and GPS location from the satellite images, and FB-LTSM is utilized to predict the potential locations for fishing. The extensive experimentations are carried out utilizing the satellite data from Indian National Centre for Ocean Information Services (INCOIS) and implemented using TensorFlow 1.18 with Keras API. The performance metrics such as prediction Acy, precision (Pscn), recall (Rcl) or sensitivity (Sty), specificity (Sfy), and F1-score and compared with other existing intelligent learning models. From our observations, the proposed architecture (99% prediction Acy) has outperformed the other existing algorithms and finds its best place in designing an intelligent system for better predicting of PFZs.


Introduction
e coastal oceanic atmosphere assumes a fundamental part in India's economy by excellence of their assets, useful living spaces and wide biodiversity. India has a long coastline of 7517 km incorporating islands which is a significant region both for investigation and misuse of common assets by the exclusive economic zone (EEZ) of 2.5 million km 2 . Marine fisheries area assumes a critical part in the economy as far as giving work to more than 14 million individuals and unfamiliar trade profit through export. e yearly marine fisheries creation in India is about 2.94 million tons against the harvestable capability of 3.93 million tons [1]. However, there is still a problem in finding the fishing areas that fishermen should visit [2,3].
Prediction of fishing zone has been done utilizing the sea parameters derived either from satellite images or ground truth primary data [4,5]. However, most of the research framework's utilized ocean graphic parameters such as chlorophyll and sea surface temperature (SST) feature for the prediction [6][7][8]. e prediction of fishing zones has now reached to new dimensions by the usage of machine and deep learning algorithms. Several algorithms such as long short-term memory (LSTM) [9], Markov models [10,11], Naïve Bayes (NB) classifiers [12], support vector machines (SVM) [13,14], and deep neural networks (DNN) [15][16][17] are used for prediction of fishery area based on different oceanographic parameters. However, an accurate prediction for PFZs still remains on the darker side of the research. To solve the aforementioned problem, this study proposes the novel hybrid model HE-DFNETS (hybrid ensemble DEEPFISHNETS) which integrates the double tier convolutional neural layers (DTCN) and flitter bat optimized LSTM (FOLSTM) for an efficient prediction of PFZs (PEZ) using remote sensing images. To the best of our knowledge, this work is the first of its kind utilized for the prediction of PEZ. e main contribution of the paper is as follows: (1) e ensembled convolutional layers are implemented to handle the different remote sensing images which comprise of sea surface temperature maps and sea surface chlorophyll (SSC) maps (2) Traditional training network is replaced with optimized long short-term memory (LSTM) for better performance. (3) erefore, flitter bats are implemented to optimize the hyperparameters of the LSTM for a higher prediction rate.

Related Works
Rahul et al. introduced fishery information revelation dependent on help vector machines and fluffy principles to recognize fish stock, and the utilization of undersea innovation and GPS to develop programmed fishery examination frameworks are probably the most recent patterns being embraced around the world to improve, investigate, and grow the monetary fishing zones over the seas. is system helps for long-term forecast of the PFZ [18]. Su et al. utilized random forest (RF) and gradient boosting decision tree (GBDT) AI techniques to precisely infer saltiness inconsistency data in the worldwide subsurface and more profound sea (0-2000 m). As indicated by the outcomes, the RF model can well recover the SSA and beat the GBDT model. Besides, the exactness of the two models for the most part decline with profundity under 500 m [19].
Chacko et al. exhibit the helpful use of satellite information in the assessment of OHC (ocean heat content) with better spatial and temporal inclusion.
is structure has assessed OHC700 in the Indian Ocean utilizing satellitedetermined SST, (sea surface height anomalies) SSHA and OHC700 clim by utilizing the ANN procedure. e outcomes recommend the utility of the ANN strategy in assessing OHC700 with sensible exactness on a close continuous premise [20]. Wang et al. proposed an AI forecast strategy joined with wavelet change. is interaction utilizes information from upper sea perception floats put in the Arctic Ocean (AO.) to anticipate the sensor simple of chlorophyll-A (C.A.) in the upper expanse of the AO. A model joining SAE (stacked auto encoder) Bi (bidirectional) LSTM and wavelet change is proposed. From the experiment, these frameworks provide better results in regards to root mean square error (RMSE) and mean absolute error (MAE) [21].
Heyn et al. present a procedure to screen the ice condition continuously through assessment of boundaries that describe the appropriation of frame speed increases. It is shown how a Kullback-Leibler disparity measure can arrange ice condition among a bunch of pretrained conditions. e examination shows that the factual order techniques, planned by measure information, give steadier and more solid outcomes [22]. e data collection unit collects from the satellites, image preprocessing and augmentation, segmentation with feature maps extractions using convolutional layers and finally flitter bat optimized LSTM networks.

Materials and Methods (Data Collection Unit).
A variable environmental database was accumulated consisting of SST, SSC, and GPS co-ordinates (latitudes and longitudes). Given their importance as environmental predictors of fishing zones [23], three variables are mostly used in modelling the proposed architecture. As mentioned in [24][25][26], SSC information gives data on sea's usefulness and are significant for recognizing fronts and vortexes that are not generally obvious in SST maps. Table 1 shows the source of the different environmental satellite image data along with their characteristics and specification.
Nearly 10 years of image data (Jan 1 2011 to April 2021) has been downloaded to train the proposed architecture. Figure 2 shows the sample remote sensing images which represents the SSL and SSC parameters, respectively.

Image Preprocessing and Augmentation.
e preprocessing technique is used to remove noise pixels, lowquality pixels which affects the prediction ratio. Image histogram methods are employed for enhancing the image quality even though the resolutions of obtained images are high resolutions.

Image Augmentation.
To overcome the problem of overfitting, the image augmentation process is incorporated in the proposed framework.
ough we have downloaded the ten years of remote sensing images, these image data are considered to be a limited quantity to train the network. Hence, data augmentation is adopted to tackle this problem. To perform the data augmentation, offline transformations [27] is applied on the series of each image which produces the large amount of newly corrected training image samples. e obtained augmented samples have the same correlations with original images, and this step is most widely used for preventing the overfitting problems as shown in Figure 1.

2
Computational Intelligence and Neuroscience

Proposed Network Training.
is section details about the working mechanism of the proposed double-tier convolutional layers and flitter bat optimized LSTM.

Proposed Ensemble Convolutional Layers.
As mentioned in [28], the ensemble convolutional neural networks are used for an effective segmentation and feature extraction which is presented in Figure 3. e first-tier convolutional layers (Table 2) are used to segment the SSL remote sensing images in which the features are extracted and stored as separate feature maps. e similar fashion of convolutional layers (Table 2) are employed to extract the feature maps from the SSC maps.
Kim et al. [9] mainly focus on the assumption on temperature rise in the water by using the latest methods like "LSTM and deep learning" approach along with the "HWT" approach. So, the loss of all sea species can be prevented.

Optimized LSTM Training for Prediction.
en, the feature maps are extracted and ensembled for training the network.
e proposed system replaces the traditional neural network training network with the flitter bat optimized long short-term memory. e proposed LSTM training network's working mechanism is presented in the preceding section.
3.5.1. Hyperparameter Optimized LSTM Network. As mentioned in [17], though LSTM plays an important role in the prediction, performance of the network degrades when it handles larger datasets [27]. In the existing frameworks, there is computational complexity when the dataset scale  Computational Intelligence and Neuroscience gets increased. Motivated by this drawback, the proposed LSTM training must be aware of the computational complexity whose hyperparameters such as epochs, learning rate, and hidden layers are optimized by the bio-inspired flitter bat algorithms [29]. is approach will yield a better prediction rate compared to traditional network.

Flitter Bat Optimized LSTM.
Flitter bat algorithm is used to optimize the hyperparameters of the LSTM. e low complexity and less computational time of flitter bats than other bio-heuristic algorithms such as PSO and GA [30,31] has inspired us to implement the flitter bats to optimize the hyperparameters of LSTM. In this case, no of epochs, learning rate, and hidden nodes are taken as the input bat population whereas the fitness function is calculated by where A(n) is the A cy at initial stage, A(n + 1) is the A cy at preceding section, and n is the number of iterations. e working mechanism of optimized LSTM is presented in Algorithm 1.

Experimental Setup
e proposed HE-DFNETS are implemented in TensorFlow 3.18 with Keras API which runs on "Windows PC10 with i7 CPU, 4 GB NVIDIA Geo-force GPU, 16 GB RAM and 2.5 GHZ".

Performance Metrics and Evaluation
For the better classification, the images are resized to 256 × 256 × 3. Nearly 1, 06,700 image datasets were used for training. Table 3 depicts the partitioned datasets utilized for both training and testing the network. e hybrid combination of the CNN-FO-LSTM network is used in the proposed architecture whose hyperparameters are optimized by flitter bat algorithms. e sample images were used for training the proposed network. As the next step, the proposed architecture is tested with the images in which the ensembled convolutional layers extracts the image features and O-LSTM training networks classifies the appropriate categories. To prove the outstanding performance of proposed architecture, metrics such as "A cy , S ty , S fy , and F1-score" are calculated. Table 4 shows proposed framework's validation parameters.

Results and Discussion.
is section presents the significance of the proposed architecture over the other existing learning models in terms of various performance evaluations. e evaluation is carried out in tri-folded scenario. In the first scenario, prediction of PFZs for different areas are validated. Additionally, validation loss characteristics and receiver operating characteristics (ROC) are evaluated for the proposed architectures. In the next scenario, performance of the proposed architecture is compared with the other state-of-the-art learning models such as SVM [14], BILSTM [9], NB [12], gradient-boosted decision trees (GBDT) [21], artificial neural networks [17], and KNN-RF [32]. To overcome the imbalance problem, the proposed learning architecture is tested with random images.

Scenario-I. In this evaluation, prediction A cy of the proposed architecture is calculated for the Indian East
Coastal Areas from the random dates from Jan 1 2021 to May 1 2021 along with the other performance metrics. Table 5 shows the prediction performance of the proposed architecture. e validation of the proposed architecture is done by random data sets in order to avoid the imbalance problems [33,34]. From Table 5, it is found that the proposed architecture has predicted the PFZs accurately. e prediction A cy of proposed architecture is shown in Figure 4. e integration of ensembled convolutional features and optimized LSTM training has yielded the 99% prediction A cy which was observed from the month of January to May 2021. Table 6 presents the average performance metrics such as "sensitivity, P scn , S fy , and F1-score".
(1) Inputs: hidden layers, learning rate, and number of epochs (2) Outputs: A cy of prediction (3) Initialize the bat populations as hidden layers, learning rate and number of epochs (4) Initialize initial velocity, loudness, frequency, and distance using equation (7) and (8) (5) While n � 1 to max_iteration (6) Calculate the F(A) using equation (1) (7) If F(A) � � threshold (equation (1)) (8) Go to step 12 (9) Else (10) Update the velocity, loudness, frequency, and distance (11) Go to step 6 (12) End (13) End (14) End ALGORITHM 1: Optimized LSTM networks.  Computational Intelligence and Neuroscience e similar fashion of the performances (99% S ty , 99% P scn , 99% S fy , and 99% F1-score) has been found from Jan 2021 to May 2021. Tables 7 and 8 give the comparative analysis of proposed framework vs. other existing frameworks. Tables 7 and 8 show the comparative analysis between the proposed architecture and other learning models. It is found that integration of optimized LSTM with ensembled          Computational Intelligence and Neuroscience 2021 and also it outperformed the other learning models such as ANN (55%), NB (60%), KNN (62%), RF (63.5%), GBDT (68.5%), SVM (70%), and BILSTM (78%), respectively. Figure 5 shows the ROC characteristics of the proposed architecture for randomly chosen prediction zones. Figure 6 shows the characteristics of validation loss of the proposed architecture. It is found in Figure 6 that loss is very less than 0.001 which is considered to be more suitable for the prediction of PFZs.

Conclusion
In this paper, a novel HE-DFNETS is proposed for the prediction of PFZs areas which can be used by the fisherman community. e proposed algorithm works on the principle of ensembled convolutional layers and replaces the traditional neural network training with the optimized LSTM network. In the proposed framework, the hyperparameters are optimized by the flitter bat optimization technique. e datasets include the satellite images which comprise SST and SSC along with GPS coordinates. e datasets were downloaded from https://incois.gov.in/portal/ remotesensing/TERA_display.html. Extensive experimentations were accomplished using the above datasets, and validation metrics were calculated for different scenarios of the environment. e performance of the proposed architecture is validated randomly from the month of Jan 2021 to May 2021. It is found that the prediction A cy is maintained uniformly to 99% for every month, and it outperforms the other state-of-the-art learning models. e above results show the promising performance of the proposed architecture in predicting the PFZs and can be utilized for the betterment of the fisherman's community.
6.1. Future Enhancement. Hence, the proposed architecture requires further enhancement in terms of reduced computational complexity.
Data Availability e datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Conflicts of Interest
e authors declare no conflicts of interest.