Semantic Segmentation of Paddy Parcels Using Deep Neural

Networks Based on DeepLabV3

Syazwani Basir

, Nurul Aina Abdul Aziz

and Nurshafiza Zanal Abiddin

Malaysian Space Agency (MYSA), No. 13, Jalan Tun Ismail, 50480 Kuala Lumpur, Malaysia

Keywords: DeepLabV3, Paddy, Agriculture, Food Security, ArcGIS Pro.

Abstract: Paddy parcels are frequently converted to other structures which contributes significantly to changes in paddy

cultivation areas and a decrease in rice production. Determining the current land use status for paddy parcels

annually is quite challenging; thus, the Paddy Geospatial Information System (MakGeoPadi) has been

developed to determine the precise Malaysian paddy cultivation regions in order to provide a sufficient food

supply for the entire country. Deep convolutional neural network (DCNN) algorithms such as DeepLabV3

are used in this study to accurately estimate paddy yield of 12 granaries. The objective of this study is to

enhance the DeepLabV3 paddy parcel detection model to generate data that can be relied upon for reliable

decision-making. Deep-learning applications based on the DeepLabV3 model were classified into four

classes: active paddy parcel (PA), miscellaneous paddy parcel (PP), permanent structures (SK) and permanent

crop (TK) using ResNet50 in ArcGIS Pro version 2.9. DCNN has been utilised to perform semantic

segmentation. The DCNN architecture known as DeepLabV3 is trained using the 16,000 datasets in the

experiment, with Pleiades satellite images scaled at 224 x 224-pixel sizes. Following the training phase, the

DeepLabV3 model achieved the highest successful training accuracy, scoring 91.6%.

1 INTRODUCTION

In Malaysia, paddy parcels are widely cultivated

extensively in 12 paddy granaries across the country

which are planted twice a year; however, records

indicate that Malaysia has the least total area of paddy

cultivation in Southeast Asia, with an estimate of

600,000 hectares (Firdaus et al. 2020). Due to its

inability to meet the nation’s yearly rice consumption,

Malaysia continues to import rice from Vietnam and

Thailand (Tey & Radam, 2011). Nevertheless, given

that most commodities are currently experiencing

sharp price increases, this scenario is viewed as

highly risky if the supply nation ends the transaction

or increases the price of rice. The Russia-Ukraine war

is one of the factors affecting the import sector and

resulting in low supply and high demand (Wicaksana

& Ramadhan, 2022; Jagtap et al. 2022; Lin et al.

2023).

Through the National Agro-Food Policy 2021-

2030 (DAN 2.0), there is a need to enhance the

https://orcid.org/0000-0002-1203-537X

https://orcid.org/0009-0003-0776-0045

https://orcid.org/0009-0005-8597-8845

resilience of the national food system, particularly in

light of the current global crisis. This will entail

optimising network performance across all domains,

including production, processing, distribution,

nutrition, and food safety.

To deal with food security issues, the Malaysian

government aims to achieve 70% self-sufficiency

(SSL) for local rice production under the 12

Malaysian Plan (RMK-12). To assess target

attainment, precise and up-to-date spatiotemporal

information on paddy cultivation status is required.

Therefore, the Malaysian Space Agency (MYSA)

collaborated with the Malaysian Department of

Agriculture (DOA) to develop the Paddy Geospatial

Information System (MakGeoPadi), which pinpoints

the ideal paddy planting zones in Malaysia to assist

the government in monitoring to ensure that the

country has an adequate supply of rice so that it is

always prepared to deal with shortages in the global

market or an increase in grain prices. The main role

of the MakGeoPadi system is to identify the region of

Basir, S., Aziz, N. and Abiddin, N.

Semantic Segmentation of Paddy Parcels Using Deep Neural Networks Based on DeepLabV3.

DOI: 10.5220/0012698200003696

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 10th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2024), pages 173-180

ISBN: 978-989-758-694-1; ISSN: 2184-500X

173

12 granaries located across Malaysia. In the

meantime, land use changes often affect the actual

area of paddy production, accurate paddy

segmentation is essential. In order to fulfil the

requirements of the National Crop Cutting Survey

(CCS), an annual reporting requirement that has an

impact on the subsidies provided to the agriculture

sector. Under current practice, field observers

conduct the surveying process manually, which is an

expensive and time-consuming procedure.

Therefore, the application of remote sensing and

deep learning techniques is proposed for an

automated surveying system that classifies the paddy

fields' satellite photos semantically based on the

target class. Since we are handling high-resolution

satellite imagery with geolocation and spatial layer

with attribute table information, we decided to test the

capability of ArcGIS Pro's Deep Learning toolset.

ArcGIS Pro tools consume models built to

recognize certain characteristics in third-party deep

learning frameworks (such as TensorFlow, CNTK,

and PyTorch) and provide features or class maps.

(ESRI, 2020). This toolset contains several type of

modules including object detection, pixel

classification and image translation for raster

datasets. Pixel classification module that are already

equip in the Deep Learning library is UNET, Pyramid

Scene Parsing Network, DeeplabV3, BDCN Edge

Detector, HED Edge Detector and Change Detector.

Sharifzadeh, S., Tata, J., Sharifzadeh, H., & Tan,

B. (2020) previously use farm area segmentation in

satellite images using deeplabv3+ neural networks.

The DeepLabv3+ model is utilised for semantic

segmentation of farm pixels after categorising the

picture patches containing farm areas. To transfer the

learned characteristics for the new farm segmentation

problem, four distinct pretrained networks are used:

resnet18, resnet50, resnet101, and mobilenetv2. The

model trained with resnet50 achieved the maximum

semantic segmentation accuracy.

Therefore, this research aims to develop the

DeepLabV3 paddy rice detection model in order to

produce data suitable for reliable decision-making.

To attain higher accuracy, Deep-learning applications

based on the DeepLabV3 model were classified into

four categories: active paddy parcel (PA),

miscellaneous paddy parcel (PP), permanent

structures (SK) and permanent crop (TK). In this

study, the DeepLabV3 deep learning workflow was

completed using ArcGIS Pro and the Deep Learning

Toolbox. The images in this study were generated

from Pleiades satellite imagery in three-band RGB

colour.

2 METHODOLOGIES

2.1 Study Area

The research was carried out in IADA Barat Laut

Selangor (BLS), the largest paddy cultivation region

in Malaysia, which covers an area of 17,741 hectares

(Sistem MakGeoPadi, 2023). The rice production

from the BLS granary is expected to contribute to the

nation’s goal of self-sufficiency in terms of food

security (SSL) (Omar et al. 2019). The study region

is in the district of Kuala Selangor in the state of

Selangor and is depicted in Figure 1. This region is

noteworthy for its well-designed irrigation network,

which makes it easier to cultivate short-term crops

twice a year. Known as one of the most productive

regions in Malaysia for producing rice, this lush area

was crucial in producing an anticipated 155,631

metric tonnes of rice in 2021.

Figure 1: (a) The MakGeoPadi system map shows the

location of ten paddy granary areas in Peninsular Malaysia

and (b) the enlargement of a satellite image of the study area

in IADA Barat Laut Selangor (BLS) overlaid with irrigation

blocks.

The site's flat geography supports many agricultural

crops such as rice, oil palm, vegetables, and fruits.

(a)

(b)

GISTAM 2024 - 10th International Conference on Geographical Information Systems Theory, Applications and Management

174

The ditches, rice field boundary, and roads that divide

the fields which are often small clearly show their

boundaries. Since most of the year is cloudy and

rainy, there are two planting seasons: Main Season

(August through February) and Off Season (March

through July). As a result, the soil moisture content is

rather high. The irrigation system that contributes to

the varied conditions of the paddy lot planting stages

is one of the study area's unique features. Since there

are currently four irrigation phases, all planting stages

mainly ploughing, irrigating, planting, and harvesting

can be monitored simultaneously on the day that the

picture satellite is acquired.

In the last few years, there has been an annual

decline in the number of active paddy parcels in this

area of 200 hectares due to fast land use changes in

recent years (Malaysian Department of Agriculture,

2022). This condition necessitates effective

monitoring action to determine the present status of

the specific lot each season.

2.2 Preparation of Dataset

2.2.1 Satellite Image Processing

The following four key paddy-planting activities are

included in the multi-temporal Pleiades satellite

images that were obtained: cultivated paddy lands,

which contain dwellings, roads, and other land

usages; and uncultivated paddy regions. The High-

Resolution Imager (HiRI) on board the Pléiades

constellation delivers very-high optical resolution

(0.5 m resolution) with a swath of 20 km (Airbus,

2011). To improve colour presentation and contrast,

digital enhancement procedures were applied to the

satellite imagery in three-band RGB colour of

compressed unsigned 8-bit dataset with tiff format.

Image enhancement or the practice of modifying

digital images is the next compulsory step so that they

are more suited for computer vision procedure. The

adjustment of brightness, sharpen, and haze reduction

must be performed to enhance the actual feature.

Since the study area has its administrative boundaries,

the subsets of the enhanced image were then created

in order to remove areas that weren’t granary from the

scene.

2.2.2 Ground Truth Preparation

The authorised body in land surveying, the

Department of Survey and Mapping Malaysia

(JUPEM), provided the lots for the National Digital

Cadastral Database (NDCDB). Four (4) classes were

manually created out of these lots: by superimposing

the multi-temporal Pleiades satellite images over the

cadastral lot, the following paddy parcels can be

identified: active paddy parcel (PA) (ploughing,

irrigating, planting, and harvesting), miscellaneous

paddy parcel (PP), permanent structures (SK) and

permanent crop (TK). To generate ground truth

samples for deep learning training, the segmentation

data was generated in a standard shapefile GIS

format. As the authorized department, the IADA

Barat Laut Selangor conducted an on-site verification

process to confirm the accuracy of the segmentation.

The Pleiades image and the ground truth are seen in

Figure 2.

Figure 2: Images of paddy parcels samples: (a) satellite image

and (b) their ground truth labels. PA: active parcel, SK:

building label; PP: vegetable and TK: oil palm.

2.2.3 Extraction of Training Datasets

Before a deep learning model can detect

characteristics and classify the pixel, it must be

trained to recognise those items. It is our

responsibility to gather and supply input data and

training samples, after which you must train the

model to identify those characteristics or objects.

(b)

(a)

Semantic Segmentation of Paddy Parcels Using Deep Neural Networks Based on DeepLabV3

175

(ESRI,2020). The items in an image can be

interactively identified and labelled, and the training

data can be exported as the image chips, labels, and

statistics needed to train a model. The Export

Training Data for Deep Learning geo-processing tool

can be used to create the training data required for the

subsequent step if you already have labelled vector or

raster data.

A ground truth polygon shapefile and an unsigned

8 bit RGB TIF satellite picture were used to create the

training dataset. This combination enabled the

construction of the dataset. Four distinct class labels

were applied to the segmented polygon: PA, PP, SK

and TK, and the images were split into pairs of tiles

that matched the same geographic area. Each tile

represented the RGB values of the input features. A

total of sixteen thousand samples with 224 x 224 pixel

images from ground reality were utilized in the

sample.

2.2.4 DeepLabV3 Module Training

For semantic segmentation, fully convolutional

neural networks (FCNs) are frequently employed.

Using FCNs on images for segmentation tasks is

problematic since the input feature maps get smaller

as they pass through the network's pooling and

convolutional layers. As a result, information is lost

and output with fuzzy object borders and low

resolution predictions is produced. DeepLab uses a

technique called multiple pooling layers, or spatial

pyramid pooling (SPP), to handle multi-scale

pictures. With a fixed integer representing the input

image size, it partitions the feature maps produced

from the convolutional into spatial bins. DeepLabV3

employs atrous convolution with SPP to extend the

field of view of filters, which aids in integrating larger

contexts without adding more parameter (Ahmat

Imran et al. 2020).

The DeepLab model addresses this challenge by

using Atrous convolutions and Atrous Spatial

Pyramid Pooling (ASPP) modules, as shown in

Figure 3. This architecture has evolved over several

generations. Atrous Convolution is introduced in

DeepLab as a tool to adjust or control the effective

field-of-view of the convolution. It modifies field-of-

view using a parameter called ‘atrous or dilation rate’.

It is a simple yet powerful technique to make the field

of view of filters larger without impacting

computation or the number of parameters. Atrous

convolution is similar to traditional convolution

except the filter is up sampled by inserting zeros

between two successive filter values along each

spatial dimension. r - 1 zeros are inserted where r is

atrous/dilation rate. This is equivalent to creating r −

1 holes between two consecutive filter values in each

spatial dimension. In the diagram below, a filter of

size 3 with a dilation rate of 2 is applied to calculate

the output. We can visualize filter values separated by

one hole since the dilation rate is 2. If the dilation rate

r is 1, it will be standard convolution (Chen et al.

2016).

Figure 3: Atrous Spatial Pyramid Pooling in DeepLabV3

Model Architecture (ESRI, 2020).

Using the ArcGIS Pro Train Deep Learning

Model the model parameter which is DeepLabV3, the

maximum epoch, the batch size, model argument and

backbone model variable need to be determined.

In principal, the increasing of the batch size can

improve tool performance; however, as the batch size

increases, more memory is used. In this study, the

batch size used is 8. The batch size is a hyper

parameter of gradient descent that controls the

number of training samples to work through before

the model’s internal parameters are updated while

number of epochs is a hyper-parameter of gradient

descent that controls the number of complete passes

through the training dataset. (Brownlee, J., 2022)

Three type of backbone model was tested which

is Resnet34, Resnet50, and Resnet101 to test the best

model fit the study.

This backbone selection enhances

the model's efficacy and efficiency, especially with

regard to the use of computational resources and

precision in capturing complex information. The chip

size is 224 which is suite for 0.5-meter sample image

resolution.

2.2.5 Validation of DeepLabV3 Model

The trained model is validated to assess its

performance and ensure that it can effectively

generalise to a new or previously unexplored dataset.

The study uses a training dataset with 10% validation.

Test accuracy indicates that the trained model

recognises independent images that were not used in

training, whereas training accuracy indicates that the

same images are used for both testing and training.

When training a deep learning model for imagery, the

output from the Train Deep Learning Model tool

includes a file named model_metric.html. This file

contains information on your trained model, such as

GISTAM 2024 - 10th International Conference on Geographical Information Systems Theory, Applications and Management

176

the learning rate, training and validation loss, and the

average precision score (ESRI, 2020).

3 RESULT AND DISCUSSION

3.1 DeepLabV3 Training Accuracy

Several metrics are available when using Deep

Learning to help us determine how effectively our

model is performing. These features include the

smoothness of the curve, its convergence, and how

generalizable the learning rate is. (Ibrahim, M.,

2023). With this knowledge, we might deduce more

details about our models. In this study, we will be

concentrating on accuracy and loss. They are both

crucial values to consider while we are training our

models. Loss is a value that represents the sum of all

errors in our model. It gauges how well our model is

performing.

A model that has a low accuracy but a high loss

would indicate that it makes significant mistakes in

the majority of the data. However, low accuracy and

loss indicate that the model produces modest errors in

the majority of the data. The best scenario would be

for the model to make tiny errors on a small portion

of the data if the accuracy is high and the loss is low.

Riva, W. (2023).

The result of training accuracy from three

different backbone model are depicted in Figure 4.

Based on the graph 4(a) and 4(b), the high variance of

the model performance is an indicator of an

overfitting problem. The training time of the model or

its architectural complexity may cause the model to

overfit. If the model trains for too long on the training

data or is too complex, it learns the noise or irrelevant

information within the dataset. At 8000 batches

processed in epoch 10, the model has to be stopped.

Figure 4(c) depicts the optimal Restnet-50

backbone since the learning curve converges to a

point where additional training does not yield

significant improvements. This indicates that the

model has maximised its learning from the training

set and is operating at peak efficiency. Naturally,

there may be little ups and downs along the route, but

sharp, abrupt jumps in the curve may indicate that

something is off. A smooth learning curve indicates a

steady and reliable learning process for the model.

In term of accuracy, it compares the model's

predictions with the actual values in terms of

percentage, it evaluates how well our model predicts

Figure 4: Learning curve (a) RestNet-34 (b) RestNet-101 (c)

RestNet-50, respectively.

Figure 5 shows the training accuracy of RestNet-

50 module backbone. As the epoch increased, the

model correctly predicts the outcome of the next step

does the accuracy increased. Understanding these

dynamics allows us to see how, as the model

progressively gets better and makes more accurate

predictions, the accuracy curve displays discrete

jumps rather than a smooth trend. Final training

accuracy is 94.5%.

Figure 5: Accuracy of trained Module.

(a)

(c)

(b)

Semantic Segmentation of Paddy Parcels Using Deep Neural Networks Based on DeepLabV3

177

3.2 DeepLabV3 Segmentation Output

From the tested dataset, a total area of segmented

images (17,000 hectares) was successfully generated.

About 30 minutes were needed for the pixel

classification procedure, a far shorter time than the

hour needed for training. As shown in Figure 6, the

class of miscellaneous crop like vegetable, corn, and

banana tree tends to be incorrectly assigned as an

active parcel like segmentation output in (a) and (b).

This situation only happened during harvesting

season of those short-term crop which is bare land

that or the features of line from paddy straw which is

similar with vegetable ridges. The quality of the

image and the status of paddy planting during image

acquisition always become the main factors in a good

result. Half of the area in this image was captured

during the land preparation phase, when the features

were identical for the PP harvesting phase.

Otherwise, most classifications for active parcel

(PA) and permanent structure (SK) are very well

segmented based on the observations in (c) and (d).

Only certain permanent structure like house in a

smaller size will be generalized as (PA). Inaccurate

annotations, or "label noise," have a substantial

impact on performance and can occur throughout

labelling process. The model is unable to learn the

characteristics of exact object borders and, to some

extent, narrowly formed small objects due to relative

shift and segment error. ( Maiti, A et al. 2022).

3.3 Accuracy Assessment

To enhance the accuracy of the segmentation map,

ground-checking activities were carried out with the

agriculture agency. A total of 120 randomly selected

points roughly 30 points per class were used for the

accuracy assessment by using the ArcGIS Pro

compute confusion matrix tool. To store the class that

results from the classification and the matching

ground truth class for each point, respectively, the

tool accepts shape files with two fields designated

classified and ground truth (Jensen, 2016). The

confusion matrix produced by the accuracy

assessment tool is shown in Figure 7. The number of

points from each ground truth class that were found

on the classification output classes (columns) and the

number of points from each classification output class

that are part of the ground truth classes (rows) are

cross-referenced in the matrix. The major diagonal of

the matrix lists the number of correctly identified

points and the total number of correctly classified

points.

Figure 6: Enlargement of segmentation output.

User accuracy refers to how real the classified

map is on the ground. Thus, in this study, 91.6% of

the mapped area has the same features as the ground,

with a Kappa of 0.88. Producer accuracy refers to the

classification scheme. According to the findings, the

miscellaneous paddy parcel (C2) category and

permanent crop (C4), which are minority classes in

the area can be damaging to the learning process

because the number of samples is a bit lower than C1

and C3. Furthermore, there should be a significant

and uniform distribution of training samples with the

anticipated segmentation labels for every label. The

existence of various classes in one plot will increase

the complexity of pixel classification in the area.

GISTAM 2024 - 10th International Conference on Geographical Information Systems Theory, Applications and Management

178

Figure 7: DeepLabV3 user and producer accuracy.

3.4 Segmentation Post-Processing

Several post-processing procedures must be

completed on the DeepLabV3's raw classification

output, particularly when it comes to paddy

cultivation activities that are governed by individual

lot borders. For raster-to-vector conversion, the

ArcGIS Pro spatial analyzer tool is therefore

essential. To meet the requirements of the

MakGeoPadi database design, several post-

processing steps are compulsory to apply. After the

raster layer is converted to a polygon, the first step is

to smooth the polygon with Bezier curves, which will

be fitted between vertices using Bezier interpolation.

The vertices of the input polygons are then traversed

by the resultant polygons. There is no tolerance needed

for this algorithm. The result will approximate Bezier

curves. Second, to match the segmentation result into

their lot boundary, the polygon needs to be merged

with the cadastral lot. The merge tool combines the

data from numerous sources and adds it to a new data

set. It combines attributes with the ability to match

fields from input datasets, so it's not only geometry.

As long as the layers have the same feature type, the

merge geoprocessing tool can combine two or more

of them. The type and arrangement of fields in

attribute tables must be checked especially the

duplicate fields. The area information of each

polygon needs to be calculated again in hectares’ unit.

The attribute table shows details about a chosen

layer's features. A feature (with or without geometry)

is represented by each row in the table with a new

unique code lot that represents each polygon.

Third, the minimum mapping unit was set to 0.01

hectare by using elimination tools. This tool is used to

remove a polygon by combining it with the polygon

that has the longest border with the surrounding

features. The tool is used in situations where topology

exists between feature classes.

3.5 Integration with MakGeoPadi System

The new segmented paddy lot area is ready to be

updated in the database. Following that, a map of the

paddy status activities was added to the MakGeo-Padi

database together with other auxiliary data including

the owner's profile, the farmer's revenue, and yield

details. An entity relationship diagram is created

using the information gathered to explain the general

link between tables and spatial data. Furthermore, the

logical structure is designed to guarantee that data is

kept in an orderly fashion (Siti Masayu Yahaya et.al.,

2015)

The segmented area also becomes the base data

for further analysis including analyzing planting

activities using SAR satellite. The latest information

on the status of paddy planting activities is crucial to

identify paddy parcels that are unable to comply with

the planting schedule in each season to enable

appropriate action by the agricultural agency,

including providing assistance and advisory services

to the farmers involved. Through MakGeoPadi the

actual cultivated area is determined thus the amount

of sustainable production can guarantee sufficient

national food supply in preparation for the increase in

the number of the country's population according to

the Malaysia National Transformation 2050 (TN50).

4 CONCLUSIONS

DeepLabV3 segmentation with backbone Resnet-50

was found to be able to semantically classify the

paddy cultivation area into active paddy parcel (PA),

miscellaneous paddy parcel (PP), permanent

strictures (SK) and permanent crop (TK) with training

accuracy of 91.6%. Future improvements to the

training data can be made to improve deep learning’s

semantic segmentation results mainly in providing

balanced training sample. The condition of the activities

and the precise and timely statistics of the paddy

planting area are necessary of managing agriculture and

formulating policies. This produces crucial data for

machinery. According to the DOA's efficacy research,

monitoring by satellite pictures monitors 100% of the

entire granary while saving 50% of the time required for

field surveys.

ACKNOWLEDGMENTS

We thank DOA and IADA Barat Laut Selangor for

their cooperation in providing the data. MYSA makes

this information available to authorized users for

operationalization reasons at http://makgeopadi.

mysa.gov.my/.

Semantic Segmentation of Paddy Parcels Using Deep Neural Networks Based on DeepLabV3

179

REFERENCES

Ahmed, I., Ahmad, M., Khan, F.A., Asif, M. (2020).

Comparison of deep-learning-based segmentation

models: Using top view person images. IEEE Access,

8, 136361-136373.

Akhmalluddin et al. (2023). U-NET Semantic

Segmentation of High-Resolution Imagery Using

ArcgGIS Pro Deep Learning Analyst. presented at the

Asian Conference on Remote Sensing (ACRS),

Nangang, Taiwan. 2023.

Brownlee, J. (2022) Difference between a batch and an

epoch in a neural network, Machine

LearningMastery.com. Available at: https://machine

learningmastery.com/difference-between-a-batch-and-

an-epoch/ (Accessed: 13 March 2024).

Chen, L., Papandreou, G., Kokkinos, I., Murphy, M.

(2016). A Yuille Semantic Image Segmentation with

Deep Convolutional Nets and Fully Connected CRFs,

arXiv:1412.7062 2016.

Elharrouss, O., Akbari, Y., Almaadeed, N., & Al-Maadeed,

S. (2022). Backbones-review: Feature extraction

networks for deep learning and deep reinforcement

learning approaches. arXiv preprint arXiv:2206.08016.

ESRI (2020). How DeepLabV3 Work. Access at

https://developers.arcgis.com/python/guide/how-deep

labv3-works/

Firdaus, R.R., Leong Tan, M., Rahmat, S.R., Senevi

Gunaratne, M. (2020). Paddy, rice and food security in

Malaysia: A review of climate change impacts. Cogent

Social Sciences, Vol. 6(1), pp. 1818373

Ibrahim, M. (2023) A deep dive into learning curves in

machine learning, W&B. Available at:

https://wandb.ai/mostafaibrahim17/ml-articles/reports/

A-Deep-Dive-Into-Learning-Curves-in-Machine-Lear

ning--Vmlldzo0NjA1ODY0

Jagtap, S., Trollman, H., Trollman, F., Garcia-Garcia, G.,

Parra-López, C., Duong, L., Martindale, W., Munekata,

P.E., Lorenzo, J.M., Hdaifeh, A., Hassoun, A. (2022).

The Russia-Ukraine conflict: Its implications for the

global food supply chains. Foods, 11(14), p.2098.

Jensen, John R. (2016). Introductory Digital Image

Processing: A Remote Sensing Perspective. 4th ed.

Pearson.

Lin, F., Li, X., Jia, N., Feng, F., Huang, H., Huang, J., Fan,

S., Ciais, P., Song, X.P. (2023). The impact of Russia-

Ukraine conflict on global food security. Global Food

Security, 36, p.100661.

Maiti, A., Oude Elberink, S. J., & Vosselman, G. (2022).

Effect of label noise in semantic segmentation of high

resolution aerial images and height data. ISPRS Annals

of the Photogrammetry, Remote Sensing and Spatial

Information Sciences, 2, 275-282.

Omar, S.C., Shaharudin, A., Tumin, S.A. (2019). The status

of the paddy and rice industry in Malaysia. Khazanah

Research Institute. Kuala Lumpur. Malaysia.

Production, and Yield. Retrieved From http://www.

fao.org/3/ca6408en/ca6408en.pdf

Quan, B., Liu, B., Fu, D., Chen, H., & Liu, X. (202).

Improved deeplabv3 for better road segmentation in

remote sensing images. In 2021 International

Conference on Computer Engineering and Artificial

Intelligence (ICCEAI) (pp. 331-334). IEEE.

Riva, W. (2023) Interpretation of loss and accuracy for a

machine learning model, Baeldung on Computer

Science. Available at: https://www.baeldung.com/

cs/ml-loss-accuracy (Accessed: 10 March 2024).

Sharifzadeh, S., Tata, J., Sharifzadeh, H., & Tan, B. (2020).

Farm area segmentation in satellite images using

deeplabv3+ neural networks. In Data Management

Technologies and Applications: 8th International

Conference, DATA 2019, Prague, Czech Republic.

Sistem Maklumat Geospatial Tanaman Padi,Retrieved

August 18 2023. http://makgeopadi.mysa.gov.my/jelap

ang/_main

Siti Masayu Yahaya et.al (2015) "Remote Sensing and GIS

Web- Based System for Paddy Cultivation

Management in Malaysia," presented at the Asian

Conference on Remote Sensing (ACRS), Manila,

Philippines, 2015.

Tey, J.Y.S., Radam, A. (2011). Demand patterns of rice

imports in Malaysia: Implications for food security.

Food Security, 3, 253-261.

Wicaksana, K.S., Ramadhan, R.F. (2022). The Effect of the

Russia-Ukraine Crisis on Price Fluctuations and Trade

in Energy Sector in Indonesia. Jurnal Nasional

Pengelolaan Energi MigasZoom, 4(1), 6-18.

GISTAM 2024 - 10th International Conference on Geographical Information Systems Theory, Applications and Management

180