Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 18

CHAPTER FOUR

DATA ANALYSIS AND DISCUSSION


4.1 Introduction

This chapter presents a detailed analysis and discussion of the data collected for improved

intelligence gathering from satellite images using deep learning techniques. The analysis

involves examining the characteristics of the dataset, exploring the preprocessing and annotation

procedures, discussing the model-building process, applying transfer learning techniques, and

evaluating the performance of the developed models. The findings obtained from this analysis

provide insights into the effectiveness and potential of the proposed approach.

4.2 Dataset Description

To begin, we provide a comprehensive description of the satellite image dataset used in this

study. The dataset comprises a diverse collection of satellite images captured from various

sources and covering different geographical locations. The images exhibit a wide range of

resolutions, capturing both low-resolution scenes and high-resolution scenes with fine-grained

information. The dataset is carefully curated to include a variety of environments, such as urban

areas, rural regions, and natural landscapes, ensuring a representative sample for intelligence-

gathering tasks.

4.3 Data Preprocessing and Annotation

In this section, we discuss the preprocessing steps applied to the satellite images before model

training. The preprocessing techniques aim to enhance the quality of the images, remove noise,

and normalize the pixel values to facilitate effective model learning. Additionally, we describe

the annotation process employed to create a ground truth for training and evaluation purposes.
The annotation process involves labeling objects of interest, enabling the deep-learning models

to learn to recognize and classify these features accurately.

4.4 Model Building

The model-building process is a crucial step in developing an effective intelligence-gathering

system. In this section, we outline the architecture of the deep learning models employed for this

study. The models are based on convolutional neural networks (CNNs), which have proven to be

highly effective in image analysis tasks. We provide details about the layers, activations, and

parameters utilized in the models to extract relevant features from the satellite images. The

model-building process involves training the models using the annotated dataset and optimizing

the model's performance through iterative experimentation. Table 4.1 shows the summary of the

model

Table 4.1 Model Summary

Layer (type) Output Shape Param #

conv2d (None, 254, 254, 16) 448

max_pooling2d (None, 127, 127, 16) 0

conv2d_1 (None, 125, 125, 32) 4,640

max_pooling2d_1 (None, 62, 62, 32) 0


conv2d_2 (None, 60, 60, 16) 4,624

max_pooling2d_2 (None, 30, 30, 16) 0

flatten (None, 14,400) 0

dense (None, 256) 3,686,656

dense_1 (None, 1) 257

Total params: 3,696,625

Trainable params: 3,696,625

Non-trainable params: 0

This table represents the layers of a neural network model along with their corresponding output

shapes and the number of parameters (Param #) associated with each layer. Here's an explanation

of each row:
 conv2d: This is a convolutional layer that takes an input of shape (None, 254, 254, 16)

(where None represents the batch size). It applies a convolution operation and produces

an output of shape (None, 254, 254, 16). It has 448 parameters.

 max_pooling2d: This is a max-pooling layer that takes the output from the previous

convolutional layer as input, which has a shape of (None, 254, 254, 16). It performs a

max-pooling operation and produces an output of shape (None, 127, 127, 16). This layer

has no trainable parameters (Param # is 0) since it performs a fixed operation.

 conv2d_1: This is another convolutional layer that takes the output from the previous

max-pooling layer as input, which has a shape of (None, 127, 127, 16). It applies a

convolution operation and produces an output of shape (None, 125, 125, 32). It has 4,640

parameters.

 max_pooling2d_1: This is another max-pooling layer that takes the output from the

previous convolutional layer as input, which has a shape of (None, 125, 125, 32). It

performs a max-pooling operation and produces an output of shape (None, 62, 62, 32).

Similar to the previous max-pooling layer, it has no trainable parameters.

 conv2d_2: This is yet another convolutional layer that takes the output from the previous

max-pooling layer as input, which has a shape of (None, 62, 62, 32). It applies a

convolution operation and produces an output of shape (None, 60, 60, 16). It has 4,624

parameters.
 max_pooling2d_2: This is another max-pooling layer that takes the output from the

previous convolutional layer as input, which has a shape of (None, 60, 60, 16). It

performs a max-pooling operation and produces an output of shape (None, 30, 30, 16).

Once again, it has no trainable parameters.

 flatten: This layer takes the output from the previous max-pooling layer, which has a

shape of (None, 30, 30, 16), and flattens it into a 1D tensor of shape (None, 14,400). It

has no parameters.

 dense: This is a fully connected (dense) layer that takes the flattened output from the

previous layer, which has a shape of (None, 14,400), and produces an output of shape

(None, 256). It has 3,686,656 parameters.

 dense_1: This is another fully connected layer that takes the output from the previous

dense layer, which has a shape of (None, 256), and produces a final output of shape

(None, 1). It has 257 parameters.

At the bottom of the table, the total number of parameters in the model is provided, which is

3,696,625. Both the total params and trainable params are the same in this case, indicating that

all
4.4.1 Model Training

The model training process involved training a neural network model over 20 epochs. The model

was trained using a dataset, and the training progress was monitored for loss and accuracy. The

outcome of the training process, including loss, accuracy, validation loss, and validation

accuracy for each epoch, is shown in Table 4.2 and Figure 4 1, 4.2.

Table 4.2 Training Process

Epoch Loss Accuracy Val Loss Val Accuracy

1 0.4391 0.7560 0.3649 0.8151

2 0.3485 0.8162 0.3134 0.8438

3 0.3151 0.8475 0.3238 0.8516

4 0.2625 0.8690 0.2753 0.8828

5 0.2193 0.9018 0.3455 0.8724

6 0.1827 0.9226 0.2697 0.8724


7 0.1656 0.9338 0.1998 0.9115

8 0.2600 0.8951 0.2582 0.8932

9 0.2150 0.9144 0.2723 0.9036

10 0.1661 0.9449 0.2206 0.9141

11 0.1330 0.9531 0.2201 0.9167

12 0.1137 0.9635 0.2114 0.9219

13 0.0997 0.9643 0.1654 0.9401

14 0.0587 0.9821 0.2110 0.9453

15 0.0425 0.9881 0.2322 0.9245


16 0.0273 0.9911 0.3021 0.9427

17 0.0254 0.9903 0.2718 0.9349

18 0.1187 0.9561 0.1934 0.9167

19 0.0755 0.9725 0.2625 0.9297

20 0.0421 0.9859 0.1695 0.9479

The table shows the training progress of a model over 20 epochs. Each row corresponds to an

epoch and provides the following information:

 Epoch: The epoch number.

 Loss: The training loss, indicating the discrepancy between predicted and actual values.

 Accuracy: The training accuracy, representing the proportion of correctly classified

samples.

 Val Loss: The validation loss, indicating the performance on a separate validation dataset.

 Val Accuracy: The validation accuracy, representing the accuracy on the validation

dataset.

 These metrics allow us to assess the model's performance and its ability to generalize to

unseen data.
To determine the epoch that performs better, we look at the validation accuracy metric. In this

case, the highest validation accuracy achieved is 0.9479, which occurs at Epoch 20. Therefore,

Epoch 20 performs better in terms of validation accuracy compared to the other epochs. The

figure below shows the graphical representation of the above table.

Figure 4.1 Accuracy

Figure 4.2 Loss


4.4.2 Model Performance Evaluation

To evaluate the performance of the developed deep learning models, we employ rigorous

evaluation metrics and techniques. The evaluation metrics include accuracy, precision, and recall

which provide insights into the models' capability to detect and classify high and low-resolution

in satellite images accurately. Table 4.3 shows the outcome of the accuracy, precision, and recall

Table 4.3 Precision, Recall, And Accuracy

Precision 91%

Recall 97%

Accuracy 94%

The table shows the following:


 Precision: Precision measures the proportion of true positive predictions out of all

positive predictions. In this case, a precision of 91% indicates that 91% of the positive

predictions made by the model are correct.

 Recall: Recall, also known as sensitivity or true positive rate, measures the proportion of

true positive predictions out of all actual positive samples. A recall of 97% suggests that

the model correctly identifies 97% of the actual positive samples.

 Accuracy: Accuracy is the overall correctness of the model's predictions. It measures the

proportion of correct predictions (both true positives and true negatives) out of all

predictions. An accuracy of 94% indicates that the model correctly classifies 94% of all

samples.

These metrics provide insights into different aspects of the model's performance. High precision

indicates a low rate of false positives, high recall suggests a low rate of false negatives, and high

accuracy indicates overall good performance in classification.

4.5 Pre-trained Model on Land Cover Datasets

The pre-trained model for high and low satellite image data was utilized to train and predict land

cover datasets. After 20 epochs of training, the model achieved an impressive accuracy of 90%.

This demonstrates the effectiveness of leveraging pre-existing knowledge from satellite imagery

to accurately classify and identify different land cover types. By leveraging the learned features

from the pre-trained model, the adapted model was able to generalize well and make accurate

predictions on the land cover data. This approach offers a valuable tool for land cover mapping
and monitoring, enabling efficient and accurate analysis of large-scale satellite imagery for

various environmental and geographical applications.

4.6 Pre-trained Model on Disaster Dataset

The pre-trained model for high and low satellite image data was utilized to train and predict the

Disaster dataset comprising different types of disasters. After 20 epochs of training, the model

achieved an impressive accuracy of 92%. This demonstrates the effectiveness of leveraging pre-

existing knowledge from satellite imagery to accurately classify and identify different disaster

types. By leveraging the learned features from the pre-trained model, the adapted model was able

to generalize well and make accurate predictions on the land cover data. This approach offers a

valuable way of detecting and monitoring, enabling efficient and accurate analysis of large-scale

satellite imagery for various disaster applications.

4.7 Pre-trained Model to Detect Military Spot

The pre-trained model for high and low satellite image data was utilized to train and predict

military datasets. After 20 epochs of training, the model achieved an impressive accuracy of

90%. This demonstrates the effectiveness of leveraging pre-existing knowledge from satellite

imagery to accurately classify and identify different military bases. By leveraging the learned

features from the pre-trained model, the adapted model was able to generalize well and make

accurate predictions on the military data. This approach offers a valuable tool for monitoring,

enabling efficient and accurate analysis of large-scale satellite imagery for various military

bases.

4.8 Discussion and Summary


The model-building process is a critical step in developing an effective intelligence-gathering

system. In this study, we focused on leveraging deep learning models based on convolutional

neural networks (CNNs) for analyzing satellite images. CNNs have demonstrated remarkable

performance in various image analysis tasks, making them well-suited for our purposes.Our

model architecture consisted of multiple layers designed to extract meaningful features from the

satellite images. These layers included convolutional layers with filters of different sizes,

activation functions such as ReLU, and pooling layers to reduce spatial dimensions. The

parameters of the model were carefully selected and optimized to ensure efficient feature

extraction and classification.

To train our models, we utilized a dataset that was annotated with land cover information. The

training process involved iteratively adjusting the model's weights based on the annotated data to

minimize the loss and improve accuracy. We monitored the progress of the training process by

observing the loss and accuracy metrics. After training the model for 20 epochs, we analyzed the

performance of the model using evaluation metrics such as accuracy, precision, and recall.

Accuracy measures the overall correctness of the model's predictions, while precision and recall

provide insights into the model's ability to detect and classify high and low-resolution features in

satellite images. The results of our experiments showed promising performance. The model

achieved an accuracy of 90%, indicating its ability to accurately classify different land cover

types. Furthermore, precision and recall metrics of 91% and 97% respectively demonstrated the

model's capability to correctly identify relevant features in satellite images.


The utilization of pre-trained models for high and low satellite image data proved to be

advantageous in training and predicting land cover datasets. By leveraging the pre-existing

knowledge captured in the pre-trained model, we were able to effectively classify and identify

different land cover types with high accuracy. This approach offers a valuable tool for land cover

mapping and monitoring, enabling efficient and accurate analysis of large-scale satellite imagery

for various environmental and geographical applications. In addition to land cover datasets, we

also applied the pre-trained model to disaster datasets and military datasets. In both cases, after

20 epochs of training, the model achieved impressive accuracies of 92% for disaster

classification and 90% for military base detection. These results highlight the versatility and

effectiveness of the pre-trained model in various intelligence-gathering scenarios.

CHAPTER FIVE

SUMMARY, CONCLUSION, RECOMMENDATIONS, CONTRIBUTIONS, AND

FURTHER STUDIES

4.1 Summary

The aim of this study was to improve intelligence gathering for satellite images of varying

resolutions using deep learning techniques. Specifically, convolutional neural networks (CNNs)

were employed to extract meaningful features and classify the images. The study utilized pre-

trained models for high and low-resolution satellite images and further trained them on specific
datasets related to land cover, disasters, and military spots. The models were evaluated based on

accuracy, precision, and recall metrics. Results showed that the adapted models achieved

impressive accuracies, with the land cover model reaching 90%, the disaster model reaching

92%, and the military model achieving 90%. These findings demonstrate the effectiveness of

leveraging pre-existing knowledge from satellite imagery and adapting it to specific intelligence-

gathering tasks.

4.2 Conclusion

In conclusion, the area of intelligence collection for satellite photos of various resolutions has

significantly advanced thanks to the use of deep learning techniques, particularly convolutional

neural networks (CNNs). The study achieved precise classification and identification of land

cover, catastrophes, and military sites by leveraging the capabilities of pre-trained algorithms and

honing them on particular datasets. The acquired high accuracy shows that deep learning has the

ability to offer insightful information for a variety of fields, such as environmental monitoring,

disaster management, and military surveillance applications. The process of obtaining

intelligence was greatly aided by the use of pre-trained models. These complicated patterns and

traits associated to high and low resolution photos were caught by these models, which were

originally trained on big datasets with a variety of satellite imagery. They were able to adapt and

specialize in effectively categorizing the desired land cover, disaster categories, and military sites

by fine-tuning the pre-trained algorithms using domain-specific information. The study's findings

have broad ramifications for environmental monitoring. Understanding and managing natural

resources, urban growth, and ecological changes may all be aided by accurate land cover

categorization. For the purposes of preparing for, responding to, and recovering from
catastrophes, the capacity to recognize and categorize various types of disasters using satellite

images offers invaluable insights. Accurate surveillance and identification of military locations

can improve security and strategic planning.

The interpretability and explainability of deep learning models in the context of intelligence

gathering should also be continually improved. Investigating methods for sensitivity analysis and

uncertainty assessment can improve the predictability and transparency of the models, fostering

confidence among stakeholders and end users. The study's results show, in summary, that deep

learning methods, especially CNNs, have enormous promise for enhancing intelligence

collection from satellite photos of various resolutions. The high accuracy and insightful results

obtained pave the path for improvements in applications for environmental monitoring, disaster

management, and military surveillance. Accurate categorization and identification of diverse

traits and events may be accomplished by utilizing pre-trained models and refining them on

particular datasets, leading to better decision-making and effective resource management.

4.3 Recommendations

Based on the findings of this study, the following recommendations are proposed:

 Expand the dataset: To further improve the models' performance, it is recommended to

gather more diverse and extensive datasets for land cover, disasters, and military spots.

This will enhance the models' ability to generalize and make accurate predictions.
 Explore transfer learning: Investigate the use of transfer learning techniques to leverage

pre-trained models from related domains. This could potentially enhance the models'

performance by incorporating knowledge from other fields such as remote sensing or

computer vision.

 Enhance data augmentation: Experiment with more advanced data augmentation

techniques to augment the training dataset. This can help in addressing class imbalance

and improving the models' robustness to variations in satellite imagery.

 Collaborate with domain experts: Collaborate with experts in the fields of land cover

mapping, disaster management, and military surveillance to gain domain-specific insights

and further refine the models' architectures and training methodologies.

4.4 Contribution

This study makes several contributions to the field of intelligence gathering for satellite images:

 Architecture design: The study outlines the architecture of deep learning models based on

CNNs, providing insights into the layers, activations, and parameters used for feature

extraction from satellite images.


 Performance evaluation: Rigorous evaluation metrics such as accuracy, precision, and

recall were employed to assess the models' capabilities in accurately classifying and

identifying features in satellite imagery.

 Adaptation of pre-trained models: The study showcases the effectiveness of adapting pre-

trained models for high and low-resolution satellite images to specific intelligence-

gathering tasks, resulting in improved accuracy.

4.5 Further Studies

Further studies can build upon this research in the following directions:

 Multi-modal intelligence gathering: Investigate the integration of satellite imagery with

other data sources such as aerial images, weather data, or social media feeds. This can

enable a more comprehensive and holistic intelligence gathering approach.

 Real-time analysis: Explore techniques for real-time analysis of satellite images, allowing

for timely detection and response to dynamic events or changes in the environment.

 Uncertainty estimation: Investigate methods to quantify and communicate the uncertainty

associated with the model's predictions. This can enhance decision-making processes by

providing confidence intervals and probability distributions.

 Privacy and ethical considerations: Address the privacy and ethical concerns associated

with intelligence gathering from satellite imagery. Explore ways to balance the benefits

of intelligence gathering with the protection of individual privacy and the adherence to

ethical guidelines.

You might also like