See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/374134473

Transfer Learning in Satellite Imagery Classification: A Comparative Study of

Custom CNN and Pre-trained Models

Article · September 2023

CITATIONS READS

0 190

2 authors, including:

Issam Lagoiti
Faculté des Sciences Ain Chock - Casablanca
1 PUBLICATION 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by Issam Lagoiti on 23 September 2023.

The user has requested enhancement of the downloaded file.

Transfer Learning in Satellite Imagery Classification: A Comparative Study of
Custom CNN and Pre-trained Models
Issam Lagoiti a, *, Hassan Rhinanea
Geosciences Laboratory, Department of Geology, Faculty of Sciences, University Hassan II, Casablanca, Morocco

ARTICLEINFO ABSTRACT

Keywords: In the realm of satellite imagery analysis, automated and accurate image classification remains pivotal for
various applications, spanning from environmental monitoring to urban development. This study delves into
Satellite Imagery, the efficacy of transfer learning, a potent deep learning technique, in enhancing the classification performance
Transfer Learning, of satellite images. We juxtaposed a custom Convolutional Neural Network (CNN) against renowned pre-
CNN, trained architectures, namely ResNet, MobileNetV2, and InceptionV3, fine-tuning them for our specific task.
ResNet, Our findings elucidate that while the custom CNN yielded commendable results, the pre-trained models,
MobileNetV2, benefiting from the generic features learned on vast datasets like ImageNet, showcased superior performance.
InceptionV3, Among the pre-trained contenders, MobileNet stood out, achieving an accuracy of approximately 98.93%. This
Comparative Study, research underscores the potential of transfer learning in satellite image classification, offering a pathway for
ImageNet, future endeavors seeking optimized performance with limited labeled data.
Deep Learning.

However, as the volume and complexity of satellite imagery grew, these

1. Introduction traditional methods faced scalability and adaptability challenges. The advent
Satellite imagery has long been a cornerstone for myriad applications, of deep learning marked a turning point. Convolutional Neural Networks
ranging from climate studies and environmental conservation to urban (CNNs), with their ability to automatically extract hierarchical features from
planning and disaster management [1] [2]. These images, often captured from images, soon became the gold standard for image-based tasks [10].
Earth's orbit, provide an unparalleled perspective, allowing researchers and Foundational architectures like AlexNet and VGG were adapted and fine-tuned
policymakers to observe large swathes of the planet's surface in intricate for the nuances of satellite images. The challenges of obtaining vast labeled
detail. However, the sheer volume and complexity of these images’ present satellite datasets also birthed the widespread adoption of transfer learning.
challenges in terms of processing and analysis. Manual classification, once the Researchers began adapting models, initially pre-trained on large-scale
norm, is now impractical given the data deluge, thus prompting the need for datasets like ImageNet, to satellite image tasks, thereby benefiting from the
automated tools that can swiftly and accurately classify vast amounts of rich feature representations these models had already learned [11].
satellite imagery [3].
The dataset landscape for satellite imagery has also expanded. While our
In recent years, deep learning, particularly Convolutional Neural Networks research harnesses the RSI-CB256 dataset, the broader community has
(CNNs), has emerged as a potent tool for image classification tasks, showing engaged with a variety of datasets, including Sentinel-2, PlanetScope, and
remarkable success in various domains, from medical imaging to autonomous WorldView. These datasets, each with its unique characteristics, have
vehicles [4]. Yet, training these networks from scratch demands extensive facilitated diverse research explorations [12].
computational resources and vast amounts of labeled data, posing barriers for
Yet, despite the advancements deep learning has brought to the table,
specific applications like satellite image classification.
challenges in satellite image classification persist. Notably, satellite images
Enter transfer learning, a technique that leverages pre-existing models, often exhibit class imbalances, with certain land-use categories being
trained on large datasets, and adapts them to new tasks with relatively limited underrepresented. The dynamic nature of landscapes, influenced by seasonal
data. This approach holds promise, especially for domains where data changes or human interventions, can also introduce variability that models
acquisition and labeling can be challenging. This study seeks to investigate the must contend with. Furthermore, the burgeoning frequency of satellite
potency of transfer learning in the realm of satellite imagery classification, imaging necessitates solutions that can efficiently process and analyze vast
contrasting the performance of a custom-built CNN against that of several data streams in near real-time.
renowned pre-trained architectures, gauging the advantages and potential
Several researchers have taken up the mantle to address these challenges.
pitfalls of each approach [5].
Techniques such as data augmentation, to artificially enhance dataset diversity,
Through this research, we aspire to provide insights that can guide future ensemble learning, to combine the strengths of multiple models, and semi-
endeavors in satellite image processing, offering a blueprint for achieving high supervised learning, to leverage unlabeled data, have been explored.
accuracy even when confronted with limited labeled data [6].
In the grand tapestry of satellite image classification research, our work
aims to weave a narrative that evaluates multiple deep learning architectures,
2. Related Work: offering a holistic view of their strengths, weaknesses, and potential
Satellite image classification, a research focus for several decades, has applications.
undergone significant transformations in methodology and approach [6]. In
the early days, classification relied heavily on techniques such as spectral 3. Transfer Learning vs. Deep Learning from Scratch:
analysis, where the spectral signature of pixels was used as a primary feature
In the ever-evolving domain of satellite image classification, the debate
to classify them into categories [7]. Texture analysis emerged as another
between leveraging transfer learning and training a deep learning model from
popular method, where patterns and textures in images were recognized to
scratch continues to be a central point of discussion [11]. The allure of training
distinguish different regions [8]. Rule-based systems, informed by domain
a model entirely on satellite data is undeniable, as it ensures the learned
expertise, also found their place, setting predefined rules for image
features are specifically tailored to the nuances of such data [13]. Moreover,
categorization [9].
designing an architecture from the ground up offers researchers unparalleled
flexibility, allowing for customized configurations suited to unique dataset
characteristics [14].
 However, the challenges posed by this approach are equally compelling.
The demand for vast amounts of labeled satellite imagery, coupled with the
intense computational requirements for training deep models, often makes
this method less feasible [6]. Enter transfer learning, a paradigm that has
transformed how researchers approach satellite image classification [11]. By
leveraging models pre-trained on large-scale datasets like ImageNet, transfer
learning offers a head start. These models, having already distilled significant
features from diverse images, can be fine-tuned to satellite data, often with
fewer labeled examples [15]. The efficiency and often superior performance of
this approach have made it an attractive alternative for many in the field [16].
In our research, the choice between these methodologies wasn't merely
academic; it was pivotal. Recognizing the inherent benefits of transfer learning,
especially in the context of available data and computational constraints, we
gravitated towards this approach. As we delve deeper into our methods and
results, the advantages of this choice, from efficiency to performance, will
become increasingly evident. Our exploration stands not just as an
investigation into satellite image classification but also as a testament to the
power and potential of transfer learning in this domain.

Figure 1: In the above image, the first diagram represents training

a model from scratch while the second diagram represents using a
model already trained on cats and dogs to classify the different class
of vehicles, thereby representing transfer learning.
https://www.freetimelearning.com/images/interview_questions/tra Figure 2: Schematic diagram illustrating the research
nsfer_learning.jpg methodology.
4.3. Data Pre-processing
In the initial stage of our study, we embarked on a critical preprocessing
4. Materials and Methods phase designed to prime the RSI-CB256 dataset for efficient and effective
4.1. Methodology Schematic: model training. This phase commenced with shuffling and splitting steps where
To provide a holistic overview of our research methodology, Figure 2 offers we randomized the file list of each class to avert any inherent order bias,
a comprehensive schematic that encapsulates every step, from data collection thereby facilitating a randomized distribution of images. Following this, we
to model comparison. strategically partitioned the dataset into training and testing subsets [17]. A
significant portion, determined by the train ratio, was earmarked for the
4.2. Data Collection and Description training set, serving as a rich resource for the models to learn and identify
In this study, we chiefly rely on the RSI-CB256 dataset, a well-curated distinguishing features pertinent to each class. The remainder constituted the
collection of satellite images specifically designed for land use classification. testing set, allocated for gauging the performance of the models on unseen
This dataset serves a critical role in identifying the two most predominant data.
types of land uses captured in each geospatial image, thereby facilitating a
nuanced analysis of various terrains. These images, characterized by a To further streamline the data loading process during the forthcoming
resolution of 256x256 pixels, are procured from the Sentinel-2 satellite, which training and evaluation stages, we established separate directories for the
guarantees a high level of detail in the representation of different land areas. training and testing data, wherein images were systematically organized
according to their respective classes, simplifying class-wise data retrieval.
The dataset is meticulously organized into two main directories, with the Simultaneously, we undertook data normalization to address the variance in
primary focus being on the RSI-CB256 directory, which harbors the principal pixel intensities found across satellite images, scaling pixel values to a range
set of images utilized in this study. This structured organization not only fosters between 0 and 1. This normalization process not only fostered quicker
an efficient approach to training but also augments the model's ability to convergence during training but also maintained consistency across the
classify distinct landscapes with a higher degree of accuracy. A noteworthy dataset [18].
feature of the RSI-CB256 directory is its segmentation into ten distinct classes,
each representing a unique category of land use such as 'Water', 'Green area', Additionally, to bolster the model's generalization capabilities and to
'Desert', and 'Cloudy'. This classification into specific categories further counteract potential overfitting due to limited data in some classes, we
enhances the potential of the model in distinguishing and categorizing various employed data augmentation techniques on the training set. These techniques
land uses, thereby serving as a robust tool in the study of geospatial imagery encompassed common transformations such as rotation, shearing, zooming,
and land use classification. and horizontal flipping, introducing a layer of variability and enabling the
model to discern land use patterns under a variety of conditions. As we
concluded the preprocessing phase, the dataset stood systematically organized
and optimized, perfectly poised for the ensuing stages of model training and
evaluation, thereby promising a robust and comprehensive analysis.
2
 4.4. Dataset Exploration: 4.6. Model Creation:
In the preliminary stage of our study, we meticulously scrutinized the Our research essentially pivoted around the adept utilization of deep
composition and inherent challenges present in the RSI-CB256 dataset. Our learning frameworks to facilitate precise classification of satellite images, a
first step in this analysis involved a detailed assessment of class distribution rapidly evolving field in recent years [19]. This endeavor involved harnessing
within the dataset. We undertook the task of quantifying the number of images the power of both custom and pre-trained models, thereby promising a robust
available for each land use class to pinpoint any notable imbalances. This and comprehensive analysis of their performance on our dataset.
quantitative analysis was vividly illustrated through pie and bar charts, which
effectively conveyed the proportional and exact distribution of images across In the context of pre-trained models, we turned to the substantial feature
different classes, thereby offering a comprehensive view of the dataset's extraction capabilities that established architectures could offer. In our pursuit
structure. to leverage these, we integrated three widely acknowledged models into our
study. First, the MobileNetV2, renowned for its remarkable efficiency, was
In addition to this, we engaged in sample visualization, where we curated tailored to suit our dataset [20]. We fine-tuned the upper layers while retaining
a selection of 20 random images from the dataset to display. Each image in this the weights of the foundational layers, which were cultivated through
selection was accompanied by its respective label, serving as a window into extensive training on the ImageNet database, a rich source of pre-trained
the diverse range of land use types encapsulated within the dataset. This visual models [21]. Following this, we turned our attention to the InceptionV3 model,
exploration proved instrumental in furnishing vital insights into the dataset's celebrated for its depth and complexity. Through the fine-tuning of its upper
characteristics and nuances, thereby laying a solid foundation for informed layers, we sought to capitalize on its adeptness in intricate feature extraction,
decision-making in the subsequent modeling stages. Through this in-depth particularly for satellite imagery. Lastly, we adopted the ResNet model, which
analysis, we were able to cultivate a foundational understanding of the RSI- is distinguished for utilizing residual blocks to counteract the vanishing
CB256 dataset, which significantly informed and guided our approach to the gradient problem, thus serving as a sterling candidate for our project. We
forthcoming modeling phases, promising a well-rounded and nuanced adapted its architecture to align with our specific classification objectives, fine-
analysis. tuning it to match our dataset's characteristics.
Alongside these pre-trained models, we also ventured into developing a
custom Convolutional Neural Network (CNN). This model was conceived from
the ground up, with a primary focus on optimizing it to complement the unique
traits of our dataset. The decisions pertaining to the selection of layers,
activation functions, and the overall depth of the network were grounded in
preliminary experiments and a keen understanding of our data's nature.
As we approached the compilation phase, each model, be it pre-trained or
custom-developed, was meticulously compiled with designated loss functions,
optimizers, and evaluative metrics. This critical step served to prime the
models for the impending training phase, guaranteeing their readiness to
adeptly learn from our meticulously augmented and preprocessed dataset,
thereby setting the stage for a potentially successful satellite image
classification.
4.7. Model Training:
Our endeavor to guarantee proficient classification of satellite images by
our models necessitated the implementation of a comprehensive training
regimen, a critical step in ensuring the reliability and performance of deep
learning models. This phase played a pivotal role in finely tuning the model
parameters to adeptly recognize the complex patterns encapsulated within our
dataset.
The initial step involved feeding the training data. In this step, we utilized
the meticulously preprocessed and augmented training dataset to expose the
models to a variety of visual inputs, an approach which has been shown to
significantly enhance model performance [21]. This was facilitated through the
Figure 3: Random images from the dataset. use of ImageDataGenerator instances, which provided a steady stream of
images exhibiting varied transformations, thereby significantly enhancing the
4.5. Data Augmentation:
learning process.
During our data preparation phase, TensorFlow's ImageDataGenerator
class played a pivotal role in refining and augmenting the dataset. This strategic Subsequently, we delineated the training process into a specified number
decision was aimed at increasing the model's exposure to varied visual of epochs, where each epoch encapsulated a complete cycle of forward and
patterns, setting a solid foundation for effective model training. backward passes of all the training instances, a process well-documented for
its role in optimizing deep learning algorithms [22]. The batch size, indicative
For the training data, we orchestrated two distinct generator instances.
of the number of training examples processed in a single iteration, was
The primary instance was meticulously configured to administer a range of
judiciously determined to balance memory constraints and convergence
augmentative transformations, such as rotation, shearing, zooming, and
trends.
shifting. By doing so, we ensured the model encountered a richer, more diverse
set of visual cues during its learning phase. To augment the efficiency of the training process, we employed a series of
callbacks, a technique which has been highlighted in literature for enhancing
Parallelly, a secondary generator was dedicated exclusively to the task of
the training process [23]. This included the model checkpointing mechanism,
rescaling. Its objective was to normalize the pixel values of training images,
which allowed us to periodically save the model weights, ensuring that we had
ensuring they resided between 0 and 1. This normalization was integral to
consistent access to the best-performing iterations. Additionally, to forestall
consistent training and fostering optimal model convergence.
overfitting and to streamline computational resources, we utilized early
In contrast to the training set, the test data was insulated from such stopping which keenly monitored validation loss, ceasing the training process
augmentative transformations to retain its original integrity. Yet, it was if a stagnation in improvement was detected over a specified number of
essential that it followed the same input scale as the training data. To this end, epochs, a strategy known for preventing model overfitting [24].
the test images were rescaled using a separate ImageDataGenerator instance,
ensuring uniformity during the model evaluation phase.
3
 In parallel, the optimization phase was undertaken, wherein the In the initial stages of this analysis, we scrutinized various metrics, as
backpropagation algorithm, in tandem with the chosen optimizer, dynamically suggested by contemporary research in the field [28]. The precision metric, in
adjusted the model parameters based on the evaluated gradients. This particular, illustrates the fraction of correctly predicted positive observations
iterative process aimed at progressively reducing the loss function, steering the over all predicted positives, highlighting the model's skill in class identification
model towards achieving optimal performance. without false positives. Concurrently, we examined the recall metric, which
showcases the fraction of actual positives correctly classified, serving as an
Throughout the training trajectory, we maintained a vigilant watch over indicator of the model's prowess in capturing a majority of positive samples for
both training and validation metrics. This bi-faceted monitoring approach a specific class, a methodology well-detailed in literature. Furthermore, we
facilitated a deeper insight into the models' generalization capabilities and the engaged with the F1-score, a metric representing the harmonic mean of
potential risk of overfitting, thereby informing necessary recalibrations in the precision and recall, providing a balanced view, especially when the class
training strategy. distribution is uneven, as corroborated by numerous studies [29].
Upon culmination of this stage, our models stood well-prepared, having Further deepening our analysis, we utilized confusion matrices, a vital tool
assimilated the requisite features and patterns from the dataset, thereby that facilitates a visual representation of predictions for each class against their
promising proficient performance in the satellite image classification task. actual counterparts. This detailed visualization served as a guide, pointing us
4.8. Model Evaluation: towards potential areas that might be experiencing difficulties and suggesting
Following the training phase, it was of paramount importance to paths that might benefit from additional data or model adjustments, a strategy
undertake a meticulous evaluation of our models' performance on data that echoed in recent research [30].
had not been previously seen during the training process, a practice that is As we proceeded, a comparative study across the models furnished a
pivotal in understanding the real-world applicability of deep learning models. comprehensive view of the capabilities harbored by each model. This
This critical phase afforded us a deeper insight into the capacity of the models juxtaposition not only highlighted the model showcasing superior
to generalize and classify satellite imagery with a high degree of accuracy, performance but also unveiled nuanced insights, such as differential
transcending the patterns and features encountered during training. performance across various classes.
At the outset of this phase, we utilized the test dataset, a repository of Identifying potential areas for improvement, we pinpointed consistent
images that had remained isolated and untouched during the training timeline, misclassifications and inherent biases in the models. This led to the
as the primary resource for this evaluation. These images, having undergone formulation of recommendations, including strategies like accruing more data,
essential preprocessing steps, were then introduced to the models to generate modifying model architectures, or implementing class-specific augmentation
predictions. techniques, aimed at ameliorating identified shortcomings.
To provide a rounded assessment of model performance, we employed a At the conclusion of our analytical journey, we had not only quantified the
series of metrics. Firstly, the accuracy metric served as a straightforward effectiveness of our models but also gathered qualitative insights that would
indicator, quantifying the proportion of images that were classified correctly steer future research trajectories and foster advancements in model
from the overall test image pool, an approach often utilized in machine development.
learning for performance evaluation [25]. Simultaneously, the loss metric
illustrated the deviation between the predicted probabilities and the actual 4.10. Model Comparison:
labels, shedding light on the model's confidence and precision in predictions. In our research, the comparison of models was a pivotal step to
understand their relative performance and merits. We evaluated a custom
However, to venture beyond the basic parameters of accuracy and loss, we Convolutional Neural Network (CNN) alongside three renowned pre-trained
adopted a more nuanced approach to analyzing the models' performance architectures: MobileNetV2, InceptionV3, and ResNet. The criteria for
intricacies. This encompassed the utilization of metrics such as precision, comparison included key performance metrics, the inherent architectural
recall, and the F1-score, which furnished a granular perspective of the models' complexities of each model, and their respective strengths and weaknesses. A
efficacy, particularly in situations where the class imbalances might skew the systematic approach was employed to ensure an objective assessment,
accuracy readings, a phenomenon well documented in the literature [26]. focusing on the potential benefits of each architecture for satellite image
These metrics underscored the models' proficiency in accurately distinguishing classification. The outcome of this comparison, detailed in the results section,
both positive and negative instances within each class. Furthermore, we offers valuable insights into the selection and implementation of the most
employed confusion matrices to visually delineate the dispersion of true and suitable model for specific satellite imagery tasks.
predicted labels, thereby illuminating areas where the models excelled,
alongside pinpointing potential zones of misclassification.
5. Results:
To facilitate a cohesive and fair comparison, all models, encompassing the
Upon rigorous evaluation of the four distinct architectures, we elucidated
custom CNN and the pre-trained frameworks, were scrutinized under the same
their effectiveness in satellite image classification. The models, each with their
evaluative lens. This uniform approach to evaluation not only underscored the
unique architectural underpinnings and training paradigms, produced results
respective strengths of each model but also flagged potential avenues for
that are critical for our understanding of the application of deep learning in this
enhancement.
domain.
As we drew this phase to a close, we were equipped with a nuanced
5.1. Model Performance Metrics:
understanding of each model's performance when faced with real-world,
For a succinct representation, the performance metrics of each model
unseen data. This critical insight paved the way for well-informed conclusions
have been tabulated below:
and strategic recommendations for future endeavors in this domain.
Training
4.9. Performance Analysis: model Loss Accuracy Epochs
Time
Upon the culmination of the model evaluation phase, we embarked on an
intensive exploration to dissect the acquired results, thereby gaining an
MobileNet 0.038769 0.989352 10 min 10
understanding of the respective strengths and weaknesses exhibited by each
model. This exploration was grounded in an analytical framework, tailored to
delineate the nuances of the models' performance and their potential Inceptionv3 0.063439 0.98669 11 min 10
implications in real-world applications [27].
ResNet 0.066325 0.984028 11 min 10

CNN 0.292482 0.877551 21 min 20

Table 1: Performance Metrics of Evaluated Models

4
 The results highlight the superior accuracy of the MobileNet model, with 5.3. Detailed Model Performance Analysis:
InceptionV3 and ResNet models following closely. The custom CNN, while To further elucidate the performance of each model, we have included
expressly crafted for the dataset, yielded results that did not surpass its pre- confusion matrices, which provide a visual representation of the model's
trained peers. performance on the test data. Each matrix displays a breakdown of the true
positives, true negatives, false positives, and false negatives obtained by the
5.2. Evolution of Training: respective model, offering detailed insight into the classification prowess and
Loss and accuracy curves for each model were plotted to offer an insight areas where the model may falter. These matrices are instrumental in
into the training trajectory, possible overfitting, convergence, and the overall identifying classes that are commonly misclassified, thereby highlighting
stability of the learning process, Figure 4 and Figure 5. potential areas for model improvement Figure 6.

Figure 4: Comparative Analysis of Training and Validation Accuracy Across Epochs for Various Deep
Learning Models.

Figure 5: Comparative Analysis of Training and Validation Loss Across Epochs for Various Deep Learning
Models.

5
 Confusion Matrix for CNN Confusion Matrix for MobileNet
models models

Confusion Matrix for ResNet models Confusion Matrix for inceptionv3 models

Figure 6: Confusion Matrix for each model

By analyzing these matrices, we were able to pinpoint specific areas where Notably, when we applied this function to our custom CNN model, we
each model excels and where they struggle, providing a roadmap for future observed a few instances of incorrect predictions, a testament to the
optimizations and adjustments. Furthermore, the confusion matrices serve as challenges associated with perfecting the classification process. These
a transparent tool for readers to grasp the real-world implications of the visualizations serve as a valuable tool for intuitively gauging the model's
models' performances, fostering a deeper understanding and facilitating performance, allowing us to identify potential areas where the model may
subsequent research and development in this domain. benefit from further optimization and refinement.
5.4. Visual Analysis of Model Predictions: As we continue to refine our models, this visual tool will remain a
In our endeavor to offer a more tangible perspective of each model's cornerstone in our analytical arsenal, assisting us in continually assessing and
predictive prowess, we developed to visually illustrate the models' predictions enhancing our models' predictive accuracy and reliability. Moreover, it offers
on a random set of images from the dataset. The essence of this function is to readers a tangible insight into the complex process of image classification
select a random batch of images and portray them alongside both the through deep learning architectures.
predicted and true labels, thereby offering a vivid illustration of the model's
classification capabilities.
As depicted in Figure 7, the function operates by selecting a random batch
of images from the dataset and predicting their respective labels using the
trained model. Following this, it plots the images in a grid format, wherein each
image is accompanied by its predicted label (along with the associated
confidence percentage) and its true label. The color coordination in the labels
serves as an intuitive indicator of the prediction's accuracy, with green
denoting correct predictions and red highlighting incorrect ones.

6
 CNN
MobileNet
Inceptionv3
ResNet

Figure 7: Visual Representation of Model Predictions on Random Image.

5.5. Conclusive Insights:

The evident supremacy of pre-trained models, following fine-tuning, can 6. Discussion:
potentially be attributed to the robust feature extraction capabilities these The intricate results drawn from our rigorous methodology provide a
models have acquired from extensive training on datasets like ImageNet. While fertile ground for extensive discussion, reflecting upon the nuances of each
the custom CNN did not match the performance zenith of its counterparts, it model's performance, the overarching themes observed, and the broader
showcased the potential of custom-tailored models for specific datasets, implications of our findings within the realm of satellite imagery classification.
emphasizing the importance of domain specificity in design.
6.1. Deep Dive into Model Performances:
To encapsulate, the results elucidate the capabilities and nuances of deep In the course of our study, we meticulously analyzed a series of models,
learning models in the realm of satellite image classification. The variations in each exhibiting distinct characteristics and results. One such model, the
model performances proffer opportunities for further research, potentially in MobileNet, particularly stood out, epitomizing excellence in accordance with
architectural refinements, tailored training strategies, and adaptive data its foundational principles. This model is ingeniously crafted to maintain a
augmentation techniques. harmonious equilibrium between computational efficiency and classification
accuracy. Its prowess in adeptly handling satellite images hints at an inherent
versatility, capable of adapting to a wide array of visual domains, making it a
promising tool in the realm of satellite image classification.
7
 Venturing further, we witnessed the remarkable performances of the Moreover, the task of acquiring labeled satellite imagery poses a considerable
InceptionV3 and ResNet models, both of which are deep architectures challenge, given the extensive resources and time it demands. This
renowned for their respective design philosophies. These models predicament brings to the fore the necessity of harnessing the immense
substantiated the widely held scientific belief that the incorporation of depth reservoirs of unlabeled data that lie at our disposal. Implementing learning
and intricate connectivity patterns within neural networks can foster a richer paradigms such as semi-supervised or unsupervised approaches could
feature extraction process. This attribute is particularly beneficial when dealing potentially revolutionize the field, paving the way for groundbreaking
with complex datasets such as satellite images, where the ability to discern advancements in satellite image classification. These paradigms would allow
finer details can be a pivotal factor in achieving higher accuracy levels. Their us to capitalize on the rich information embedded in unlabeled data, fostering
impressive outcomes in our experiments reaffirm their stature as powerful a more comprehensive and nuanced approach to image classification. By
tools in the field of image classification, capable of navigating the complexities navigating these untapped pathways, we aspire to push the boundaries of
inherent in satellite imagery. what's achievable in satellite image classification, steering towards a future
where our models are both innovative and adept in interpreting the complex
Parallelly, our journey also saw the inception of a custom Convolutional narratives woven within satellite imagery.
Neural Network (CNN), a venture into the realm of domain-specific model
development. Although this model didn't eclipse its pre-trained counterparts 6.6. Synthesizing the Implications:
in terms of performance, it emerged as a testament to the immense potential The broader implications of our study extend beyond the academic realm.
embedded in domain-tailored designs. The results it garnered serve as a With the exponential growth in satellite imagery data, there's a pressing need
beacon, illuminating the path towards further research and refinement in this for automated, accurate, and efficient classification systems. Our research,
direction. It underscores the necessity to delve deeper into the nuances of through its findings and discussions, paves the way for more refined
creating models that are finely tuned to the specific characteristics of the methodologies, emphasizing the need for a symbiotic relationship between
dataset at hand, potentially unlocking new avenues for innovation and domain-specific knowledge and advanced machine learning techniques.
accuracy in satellite image classification.
6.7. Concluding Thoughts:
6.2. Transfer Learning: A Game-Changer? In the grand tapestry of satellite image classification, our research adds a
Our experiments underscored the transformative power of transfer nuanced narrative, emphasizing the intricate interplay between data, models,
learning. Harnessing knowledge from comprehensive datasets, the pre-trained and domain knowledge. As we reflect upon our findings, the journey ahead
models rapidly converged to high accuracies. This not only highlights the promises further explorations, challenges, and innovations, ensuring that
importance of transfer learning in scenarios constrained by data or satellite image classification remains a vibrant and evolving field of study.
computational resources but also suggests a broader paradigm shift in how we
approach niche classification tasks. 7. Conclusions and Recommendations:
6.3. Challenges and Nuances: In the conclusion of our study, we underscore the instrumental role of
The analysis of satellite imagery presents two primary challenges that transfer learning in satellite image classification. The application of knowledge
warrant further exploration. Firstly, the inherent class imbalances in the gleaned from extensive datasets such as ImageNet has empowered our models
dataset can lead to biased model predictions. A detailed class-wise analysis to achieve significant accuracy levels, highlighting the importance of utilizing
could unveil strategies to mitigate these biases and facilitate a more balanced pre-trained architectures, particularly when faced with limited domain-specific
learning process. Secondly, the dynamic nature of satellite images, influenced data.
by seasonal changes, weather patterns, and human interventions, poses a
significant challenge. Investigating the models' ability to adapt to these Moreover, our research illustrates a crucial balance between complexity
temporal variabilities is crucial to enhance their accuracy and reliability in and performance. While deeper architectures like ResNet exhibited superior
classifying ever-changing satellite imagery. performance, efficiency-centric designs like MobileNetV2 illustrated the
possibility of attaining noteworthy results with reduced computational
6.4. Bridging with Existing Literature: demands. This balance is critical in real-world applications where
The performance metrics achieved in our study align with the overarching computational resources may be limited. Additionally, the venture into crafting
trends observed in contemporary satellite image classification research. Pre- models specific to satellite imagery is promising, though it demands
trained models, especially those hinging on transfer learning, are increasingly meticulous design and iterative refinement to match or surpass the capabilities
becoming the gold standard. Our findings resonate with this sentiment, adding of pre-trained counterparts.
empirical weight to the burgeoning body of evidence supporting this approach.
In terms of practical implications, the ability to accurately classify satellite
6.5. Future Research Directions: images can revolutionize domains such as environmental conservation, urban
As we move forward in this arena of satellite image classification, it is crucial development, and disaster management. Particularly in scenarios requiring
to venture into uncharted territories that promise potential advancements in real-time image classification, models optimized for efficiency, like
the field. One such avenue involves the exploration of alternative architectures MobileNetV2, may be more suitable, offering a balance of speed and accuracy.
that are built to adapt seamlessly to the intricate characteristics of satellite
imagery. Models like EfficientNet stand as potential candidates, representing Looking ahead, we recommend future research to focus on exploring
the next frontier in this domain. Despite their promising adaptability and datasets beyond RSI-CB256 to ensure robustness and generalization of models
scaling dynamics, their full potential remains largely untapped in the context across different geographical and temporal contexts. The advent of newer
of satellite image classification, indicating a rich area for future research. neural architectures like Transformer-based models and Capsule Networks
beckon exploration in the satellite image classification domain. Incorporating
Furthermore, the incorporation of advanced augmentation paradigms could attention mechanisms might enhance classification accuracy by allowing
be instrumental in elevating the efficacy of our models. Moving beyond the models to focus on critical regions within images. Moreover, with the
boundaries of traditional data augmentation, embracing techniques such as exponential growth of satellite data, federated learning paradigms, which
Generative Adversarial Networks (GANs) offers a promising prospect. These facilitate decentralized model training across multiple devices or nodes, could
networks hold the potential to synthetically enhance our dataset, thereby be revolutionary.
fostering the development of more robust and resilient models capable of
navigating the complex landscape of satellite imagery with greater proficiency. To sum up, our research provides a glimpse into the boundless potential
of deep learning in satellite imagery classification. The results, methodologies,
and discussions presented serve as a foundational platform for future studies
to build, refine, and innovate upon. As technology and data continue to evolve,
the intersection of satellite imagery and deep learning promises
transformative breakthroughs, reshaping our perspective of the world from a
bird's eye view.

8
8. Acknowledgments [12]. Hughes, G. “On the Mean Accuracy of Statistical Pattern
Recognizers.” IEEE Transactions on Information Theory, vol. 14, no.
The author would like to extend their sincere gratitude to the creators and
1, Jan. 1968, pp. 55–63, https://doi.org/10.1109/tit.1968.1054102.
contributors of the Satellite Image Classification Dataset-RSI-CB256. This
dataset, comprised of diverse classes sourced from Sensors and Google map [13]. Castelluccio, Marco, et al. “Land Use Classification in Remote
snapshots, has been instrumental in facilitating our research endeavors. Sensing Images by Convolutional Neural Networks.”
ArXiv:1508.00092 [Cs], 1 Aug. 2015,
Additionally, our appreciation extends to our peers and reviewers who
https://arxiv.org/abs/1508.00092.
have provided invaluable feedback and insights throughout the course of this
research. [14]. Lopez, Jorge. “Combining Satellite Imagery and Machine Learning
to Predict Poverty.” Medium, 23 Sept. 2022,
Lastly, we acknowledge the institutions and organizations that have
https://towardsdatascience.com/combining-satellite-imagery-and-
supported our work. Their belief in the potential of our research has been a
machine-learning-to-predict-poverty-884e0e200969. Accessed 10
driving force behind our endeavors.
Sept. 2023.

9. References [15]. Xie, Michael E, et al. “Transfer Learning from Deep Features for
Remote Sensing and Poverty Mapping.” ArXiv (Cornell University),
[1]. Abburu, Sunitha, and Suresh Babu Golla. “Satellite Image 1 Oct. 2015, https://doi.org/10.48550/arxiv.1510.00098 . Accessed
Classification Methods and Techniques: A Review.” International 21 Apr. 2023.
Journal of Computer Applications, vol. 119, no. 8, 18 June 2015, pp.
20–25, [16]. Sameera, Nerella, and M. Shashi. “Deep Transductive Transfer
pdfs.semanticscholar.org/6961/3390ca76bf103791ef251e1568de Learning Framework for Zero-Day Attack Detection.” ICT Express,
b5fe91dd.pdf, https://doi.org/10.5120/21088-3779. Mar. 2020, https://doi.org/10.1016/j.icte.2020.03.003.
[2]. Carlos Favis Ezequiel, et al. “UAV Aerial Imaging Applications for [17]. Joseph, V. Roshan, and Akhil Vakayil. “SPlit: An Optimal Method for
Post-Disaster Assessment, Environmental Management and Data Splitting.” Technometrics, 28 Apr. 2021, pp. 1–23,
Infrastructure Development.” International Conference on https://doi.org/10.1080/00401706.2021.1921037. Accessed 16
Unmanned Aircraft Systems, 27 May 2014, Dec. 2021.
https://doi.org/10.1109/icuas.2014.6842266.Accessed 30 Apr.
2023. [18]. Shorten, Connor, and Taghi M. Khoshgoftaar. “A Survey on Image
Data Augmentation for Deep Learning.” Journal of Big Data, vol. 6,
[3]. Tempa, Karma, and Komal Raj Aryal. “Semi-Automatic Classification no. 1, 6 July 2019, https://doi.org/10.1186/s40537-019-0197-0.
for Rapid Delineation of the Geohazard-Prone Areas Using Sentinel-
2 Satellite Imagery.” SN Applied Sciences, vol. 4, no. 5, 9 Apr. 2022, [19]. Bengio, Y., et al. “Representation Learning: A Review and New
https://doi.org/10.1007/s42452-022-05028-6. Accessed 4 May Perspectives.” IEEE Transactions on Pattern Analysis and Machine
2022. Intelligence, vol. 35, no. 8, Aug. 2013, pp. 1798–1828,
https://doi.org/10.1109/tpami.2013.50.
[4]. Hu, Yunfeng, et al. “A Deep Convolution Neural Network Method
for Land Cover Mapping: A Case Study of Qinhuangdao, China.” [20]. Howard, Andrew, et al. “Searching for MobileNetV3.”
Remote Sensing, vol. 10, no. 12, 17 Dec. 2018, p. 2053, Openaccess.thecvf.com,2019,
https://doi.org/10.3390/rs10122053. https://openaccess.thecvf.com/content_ICCV_2019/html/Howard
_Searching_for_MobileNetV3_ICCV_2019_paper.html.
[5]. Shallu, and Rajesh Mehra. “Breast Cancer Histology Images
Classification: Training from Scratch or Transfer Learning?” ICT [21]. Russakovsky, Olga, et al. “ImageNet Large Scale Visual Recognition
Express, vol. 4, no. 4, Dec. 2018, pp. 247–254, Challenge.” ArXiv.org, 2014, https://arxiv.org/abs/1409.0575 .
https://doi.org/10.1016/j.icte.2018.10.007.Accessed 31 Mar. 2020. [22]. LeCun, Yann, et al. “Deep Learning.” Nature, vol. 521, no. 7553, May
[6]. Cheng, Gong, et al. “Remote Sensing Image Scene Classification: 2015, pp. 436–444, www.nature.com/articles/nature14539,
Benchmark and State of the Art.” Proceedings of the IEEE, vol. 105, https://doi.org/10.1038/nature14539.
no. 10, Oct. 2017, pp. 1865–1883, [23]. Dahl, G. E., et al. “Context-Dependent Pre-Trained Deep Neural
https://doi.org/10.1109/jproc.2017.2675998. Networks for Large-Vocabulary Speech Recognition.” IEEE
[7]. Fauvel, Mathieu, et al. “Spectral and Spatial Classification of Transactions on Audio, Speech, and Language Processing, vol. 20,
Hyperspectral Data Using SVMs and Morphological Profiles.” IEEE no.1,Jan.2012,pp.30–42,
Transactions on Geoscience and Remote Sensing, vol. 46, no. 11, https://doi.org/10.1109/tasl.2011.2134090.Accessed 10 May
Nov. 2008, pp. 3804–3814, 2019.
https://doi.org/10.1109/tgrs.2008.922034. [24]. Wilson, Ashia C., et al. “The Marginal Value of Adaptive Gradient
[8]. Laleh Armi, and Shervan Fekri-Ershad. “Texture Image Analysis and Methods in Machine Learning.” ArXiv:1705.08292 [Cs, Stat], 21 May
Texture Classification Methods - a Review.” ArXiv (Cornell 2018, https://arxiv.org/abs/1705.08292 .
University), 13 Apr. 2019. Accessed 10 Sept. 2023. [25]. Sokolova, Marina, and Guy Lapalme. “A Systematic Analysis of
[9]. Bruzzone, L., and D.F. Prieto. “Automatic Analysis of the Difference Performance Measures for Classification Tasks.” Information
Image for Unsupervised Change Detection.” IEEE Transactions on Processing & Management, vol. 45, no. 4, July 2009, pp. 427–437,
Geoscience and Remote Sensing, vol. 38, no. 3, May 2000, pp. www.sciencedirect.com/science/article/pii/S0306457309000259,
1171–1182, https://doi.org/10.1109/36.843009. Accessed 24 May https://doi.org/10.1016/j.ipm.2009.03.002. Accessed 29 Aug.
2020. 2019.

[10]. Simonyan, Karen, and Andrew Zisserman. “Very Deep [26]. Chawla, Nitesh V., et al. “Editorial.” ACM SIGKDD Explorations
Convolutional Networks for Large-Scale Image Recognition.” Arxiv- Newsletter, vol. 6, no. 1, 1 June 2004, p. 1,
Vanity.com, 2014, www.arxiv-vanity.com/papers/1409.1556/. https://doi.org/10.1145/1007730.1007733.

[11]. Yosinski, Jason, et al. “How Transferable Are Features in Deep [27]. Smith, Linda, and Michael Gasser. “The Development of Embodied
Neural Networks?” ArXiv.org, 2014, Cognition: Six Lessons from Babies.” Artificial Life, vol. 11, no. 1-2,
https://arxiv.org/abs/1411.1792. Jan.2005, pp. 13–29, https://doi.org/10.1162/1064546053278973.

9
 [28]. Hohman, Fred, et al. “Visual Analytics in Deep Learning: An
Interrogative Survey for the next Frontiers.” IEEE Transactions on
Visualization and Computer Graphics, vol. 25, no. 8, 1 Aug. 2019,
pp.2674–2693,arxiv.org/pdf/1801.06889.pdf,
https://doi.org/10.1109/tvcg.2018.2843369.
[29]. Davis, Jesse, and Mark Goadrich. “The Relationship between
Precision-Recall and ROC Curves.” Proceedings of the 23rd
International Conference on Machine Learning - ICML ’06, 2006,
pages.cs.wisc.edu/~jdavis/davisgoadrichcamera2.pdf,
https://doi.org/10.1145/1143844.1143874.
[30]. Chawla, N. V., et al. “SMOTE: Synthetic Minority Over-Sampling
Technique.” Journal of Artificial Intelligence Research, vol. 16, no.
16,1June2002,pp.321–357,
www.jair.org/index.php/jair/article/view/10302,
https://doi.org/10.1613/jair.953.

10

View publication stats