Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1914

Geospatial Land Classification Via

Advanced Image Processing using CNN
Minzy M1; Divya Mohan2; Asha D.3; V. Balamurugan 4; Aryamol S.5
Electronics and Communication Engineering Department
Ahalia School of Engineering and Technology Palakkad, India

Abstract:- High resolution analysis of remote sensing images is II. RELATED WORKS
pivotal for various classification including land use
determination, environmental detection, environmental planning  This paper focuses on image processing within the XDB
and geospatial object recognition. This paper introduces a robust
plant disease database to enhance the classification of 20
method for categorizing satellite images into distinct groups,
facilitating accurate classification for global geographical areas. distinct diseases affecting 10 plant species. The processing
It includes image compression, image preprocessing, image involves two pre-processing steps—image selection and
segmentation and feature extraction. This innovative approach resizing—and employs modified VGG architectures
enables precise identification and understanding of different (VGG16 and VGG19) with pre-trained weights from the
areas, contributing to optimize resource allocation and improved ImageNet database. A comparative study based on
land management practices in agriculture. CNN is the classifier classification metrics, including accuracy, precision, recall,
that is employed in this experiment. The outcome demonstrates and F1-score, reveals that, for this specific database, a pre-
that our suggested strategy offers excellent accuracy, trained CNN with a depth equal to or smaller than VGG16
outperforming many recently published publications.
captures disease-sensitive features, enhancing the
Keywords:- CNN. recognition of plant pathologies. This justifies the superior
performance of VGG16 over VGG19 across training,
I. INTRODUCTION validation, and test data.
 The dynamic landscape evolution has led to an increased
The paper focuses on the satellite images collected from utilization of previously untapped lands, providing
UC merced land use dataset. The proposed system begins with opportunities for diverse purposes such as agriculture and
the collection of images from UCMerced land use dataset urban infrastructure development. This paper introduces an
consisting of 21 land use classes such as forests, deserts, automated approach to detect unused land spaces, utilizing
highways, beaches, agriculture, and rivers. Each image satellite image processing techniques. The dataset
measures 256x256 pixels. It captures high resolutions, comprises remote sensing Earth images, with
multispectral imagery of the earth surface, allowing for preprocessing steps involving grayscale conversion,
detailed observations of landscapes, ecosystems and changes compression, and noise reduction.Segmentation
over time. Here the land is classified into four main categories distinguishes between used and unused land regions, and
such as agriculture, forest, water, buildings. One of the local binary feature extraction captures edges, flat areas,
growing fields for image processing is agricultural image and corners.Various classification algorithms label and
processing. Hence we are placing a greater emphasis on categorize remote sensing Earth images, including the
agricultural land. Here we access agricultural land and utilization of a Convolutional Neural Network (CNN) for
identify the specific crop cultivating in that area. automatic classification and labeling. The application of
Random Forest for segregating landscapes into used and
By offering a scalable and automated method for land unused land demonstrates superior accuracy compared to
classification and crop detection, the project aims to address existing systems.
issues with manual labour, inefficiency, and resource waste in  This paper introduces an automated approach to detect
agricultural monitoring and management. Using satellite or unused land spaces, utilizing satellite image processing
aerial imagery, a convolutional neural network (CNN) model techniques. The dataset comprises remote sensing Earth
that can precisely identify and classify various forms of land images, with preprocessing steps involving grayscale
cover and crop detection on agricultural fields is being conversion, compression, and noise reduction.
developed with an emphasis on agricultural land and crop Segmentation distinguishes between used and unused land
detection. To help with agricultural management and regions, and local binary feature extraction captures edges,
monitoring, the process of identifying and categorizing flat areas, and corners. Various classification algorithms
different crops and land features will be automated. label and categorize remote sensing Earth images,
including the utilization of a Convolutional Neural
Network (CNN) for automatic classification and labeling.
The application of Random Forest for segregating
landscapes into used and unused land demonstrates
superior accuracy compared to existing systems.

IJISRT24MAR1914 www.ijisrt.com 2150

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1914

 This paper focuses on agricultural remote sensing segmentation to recognise borders and objects based on color
monitoring using HJ satellite imagery, presenting an or grayscale discontinuities.After extracting pertinent elements
efficient processing methodology. The workflow begins from the photos, the model is used to train and images are
with image radiation correction, atmospheric correction, classified.
and the computation of over 10 vegetation indices using
ENVI/IDL. Subsequently, the integration of ArcSDE with
the Oracle Database facilitates rapid image storage.
Finally, thematic agricultural remote sensing images are
swiftly published on ArcGIS Server. This study offers a
comprehensive solution for expediting the preprocessing,
storage, and dissemination of remote sensing images for
agricultural monitoring, ensuring timely and accurate data
access.

III. PROPOSED METHODOLOGY

One of the hottest areas of research in computer vision is

image classification. In other image application domains,
image classification is the fundamental system that is broken
down into three key components: classifier, picture
preprocessing, and image feature extraction.

In this paper we have divided land into several groups

including forests, agriculture and buildings. Since agriculture
plays a big part in the field of image processing,we categorize
agricultural land into several groups and determine which crop
is being cultivated as well as the amount of land utilized for
cultivation for future analysis.

Fig 2: Proposed Framework

A. CNN Classifier
CNN stands for Convolutional Neural Network. It is a
Fig 1: Classification Scheme of Land Use kind of neural network that is frequently applied to problems
involving computer vision and image processing. An input
The UCmerced landscape data set is used to capture picture filter that expands the input volume to its greatest size
satellite imagery. There are twenty-one classes of land use, is taken into consideration by CNN. The dot product is found
including agriculture, buildings, and forests. Pixel size for each between the filter and a portion of an image because of a spatial
image is 256x256. From the training data set, 21 images— position. The filter computes the dot product at each pixel and
referred to astest images—were chosen. The photos are trained is directed and complex at each pixel. The process is repeated
using CNN as the machine learning model. Images are with more filters, each of which looks through the input for a
downsized to lower computational demands and memory distinct set of template forms. A series of convolution layers
complexity because they are usually huge, which results in with activation functions strewn in between make up a
longer processing times and more memory convolutional neural network. It makes use of a RELU
consumption.Preprocessing module input consists of scaled activation function. Each layer's output serves as the
photos from the previous phase. Preprocessing of the supplied subsequent layer's input. There are numerous filters in each
photos is done to improve final results. Converting the image layer, and they all result in various activation functions.
to RGB and grayscale representations is part of this
preprocessing.Next, in an effort to improve representation and High-level and middle-level filters examine high-level
ease analysis, digital images are segmented, which entails features like corners and edges, while the filters at the earlier
dividing them into discrete portions. Here, region-based layer typically reflect low-level features.Thus, it progresses
segmentation is used, especially with thresholding from basic to more intricate aspects.

IJISRT24MAR1914 www.ijisrt.com 2151

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1914

 Advantages of CNN

 Reduced overfitting due to fewer characteristics requiring

adjustment than with a typical CNN
 Computationally cheaper due to fewer calculations,
makingit more suitable for mobile applications.
 Preserves the spatial structure of the image.
 Enables the identification of both low-level and high-level
features by using a variety of filters to search the input for
various template forms.

Fig 5: Segmented Image

Fig 3: Basic CNN Architecture

Each convolutional layer in the chain that processes the

input image applies a different set of filters to the image in
order to extract features. A non-linear filter known as a RELU
is applied to the output following each convolutional layer in
order to add non-linearity to the model. After passing through
a pooling layer, the output of the RELU layer shrinks the
feature maps' size by down sampling them. A fully connected
layer receives the pooling layer's output after it has been
flattened into a 1D vector. After being flattened, the output is
sent to a fully linked layer, which applies logic to determine
the type of image it is.

IV. EXPERIMENTATION RESULTS

 The Depiction of Complete Experiment is as Follows: Fig 6: Boundary Image

Fig 4: Input Image

Fig 7: Detected Land Image

IJISRT24MAR1914 www.ijisrt.com 2152

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1914

V. CONCLUSION

Requiring a robust approach to categorize the enormous

dataset of high-resolution photos, the performance of deep
learning-based remote sensing classification has shown their
feasibility in solving real-world problems and assists in
various statistical surveys. A thorough and comprehensive
definition of land recognition and classification is provided in
this study. The study has investigated agricultural image
processing and related application areas, and it has established
a more comprehensive model for efficient picture recognition.
CNN, the classifier which yields the highest accuracy results
is used in this experiment.

REFERENCES

[1]. V Rezende, M Costa, A Santos,"Image processing with

convolutional neural networks for classification of
Fig 8: BGR to RGB Conversion plant diseases "- 2019 8th Brazilian conference, 2019 -
ieeexplore.ieee.org.
[2]. Akshay S,Manohar N,Mytravarun T K,Pranav M
A,"Satellite Image Classification for Detecting Unused
Landscape using CNN",Proceedings of the
International Conference on Electronics and
Sustainable Communication Systems (ICESC 2020)
IEEE Xplore Part Number: CFP20V66-ART; ISBN:
978-1-7281-4108-4.
[3]. Shan Dong,Yin Zhuang,Zhanxin Yang,Long pang,He
chen,Teng long "Land Cover Classification From VHR
Optical Remote Sensing Images by Feature Ensemble
Deep Learning Network", IEEE Geoscience and
Remote Sensing Letters - Volume: 17, Issue: 8, August
2020.
[4]. Yu Haiyang, Liu Yanmei,Yang Guijun,Yang
Xiaodong,"Quick image processing method of HJ
satellites applied in agriculture monitoring",2016
World Automation Congress (WAC) - August 2016.
[5]. G. Cheng, L. Guo, T. Zhao, J. Han, H. Li, J. Fang,
"Automatic landslide detection from remote-sensing
Fig 9: Cultivated Area and Crop Type imagery using a scene classification method based on
BoVW and pLSA”, Int. J. Remote Sens., vol. 34,no. 1,
pp. 45-59, 2013.
[6]. W. Zhao, Z. Guo, J. Yue, X. Zhang, and L. Luo, ”On
combining multiscale deep learning features for the
classification of hyperspectral remote sensing
imagery”, International Journal of Remote Sensing.,
vol. 36, no. 13, pp. 3368-3379. 2015.
[7]. Scott, J. Grant, England, M. R., Starms, W. A.,
Marcum, A. Richard,and Davis, H. Curt,” Training
deep convolutional neural networks for landcover
classification of high-resolution imagery,” IEEE
Geoscience and Remote Sensing Letters., vol. 14, no.
4, pp. 549-553, 2017.
[8]. Y. Chen, T. J. Yang, J. Emer, and V. Sze, ”
Understanding the limitations of existing energy-
efficient design approaches for deep neural
networks,”Energy., vol. 2, no. L1, pp. L3, 2018.

Fig 10: Accuracy

IJISRT24MAR1914 www.ijisrt.com 2153