Analyzing Deforestation Dynamics in Romania Using Random Forest

Algorithm and Google Earth Engine

Andrei Varan

, Adriana Mihaela Coroiu

and Liviu-Mihai Iacob

Department of Computer Science, Babes-Bolyiai University,Cluj-Napoca, Romania

Department of Computer Science, Technical University of Cluj-Napoca, Cluj-Napoca, Romania

Keywords: Landsat, Machine Learning, Gee, Deforestation, Sustainability.

Abstract: Despite the vital roles that forests play in reducing erosion and filtering out CO2, illegal logging persists

globally. Due to deforestation, agricultural practices, and infrastructure development, Romania, a country

with an abundance of natural resources and forests, is facing significant deforestation. In this research, we

proposed an approach that uses Google Earth Engine, machine learning, and satellite images to overcome this

problem. By combining new technologies, the current Landsat 9 deployment enhances Earth Engine's

capabilities and enables improved forest monitoring and analysis. The study uses NASA-provided Landsat

images, filtered out for Romania’s surface with an applied reducer and machine learning techniques, both

being used in the Google Earth Engine editor, to have a better visualization of Romania's deforestation.

1 INTRODUCTION

Romania's crucial forested areas, home to diverse

ecosystems, are facing severe deforestation,

threatening ecological integrity and growth (Kozak,

Ostapowicz, Bytnerowicz, and Wyżga, 2013). To

help with the detection of the deforested areas we

have used a Random Forest algorithm in Google

Earth Engine to monitor forest loss and gain across

Romania over multiple years. One benefit of having

a visualisation of the deforested areas is that it helps

organizations to develop targeted reforestation

initiatives on the affected areas and to emphasize the

alarming rates of deforestation. Moreover, it can help

environmental agencies to detect deforestation

patterns effectively and take the necessary action,

since they rely on accurate and up-to-date forest

monitoring systems.

The most recent satellite in the Landsat family,

Landsat 9, has strengthened the Google Earth Engine

dataset by integrating the newest technologies into its

bands, sensors, and lenses. To take use of these new

technologies, Earth Engine provides certain machine

learning models.

https://orcid.org/0009-0003-7924-4074

https://orcid.org/0000-0001-5275-3432

https://orcid.org/0009-0003-6488-1855

The goal of this article is to provide a visual

representation of an area of interest using Google

Earth Engine with the help of machine learning. In

this approach, we have followed a structured

workflow depicted in Figure 1. First, we import

Landsat image collections from different satellites,

which provide high-resolution and multi-temporal

data for monitoring land cover changes. The data is

then pre-processed with cloud masking and scaling

factors to ensure a better image quality for the forest

classification. Next, we have extracted the relevant

features, such as Normalized Difference Vegetation

Index (NDVI) and spectral bands, which serve as

input variables for classification. By using a Random

Forest classifier from Google Earth Engine, we have

categorized pixels as either forest or non-forest. The

classified results are then displayed into deforestation

maps, which visualize areas of forest loss over time.

Finally, we conduct analysis and validation to assess

the accuracy of our classification and ensure the

reliability of the results.

Earth Engine’s client-side compute capacity is

restricted, therefore, we have limited the area of

interest to Romania's surface. As a result, all the

experiments in this article are carried out merely on

Varan, A., Coroiu, A. M. and Iacob, L.-M.

Analyzing Deforestation Dynamics in Romania Using Random Forest Algorithm and Google Earth Engine.

DOI: 10.5220/0013480400003928

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

In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 843-847

ISBN: 978-989-758-742-9; ISSN: 2184-4895

843

the Romanian map, employing information from four

Landsat satellites as listed in Table 1, with their

timeframe from the launched date until the end of

service. Older satellites from the Landsat family have

provided spatial images. However, their quality is

inferior to the newer ones and because of that our

starting deforestation year is 1985.

Table 1: Landsat satellites family used in this research and

their timeframe.

Satellite Timeframe

Landsat 5 March 1984 - Januar

2013

Landsat 7 A

ril 1999 - Present

(

2025

)

Landsat 8 February 2013 - Present (2025)

Landsat 9 September 2021 - Present (2025)

Figure 1: Workflow for forest classification and

deforestation monitoring using Google Earth Engine and

Random Forest.

Our enhancements include leveraging the GAUL

dataset, Landsat images, and Google Earth Engine to

develop the needed algorithm specifically for

Romania's map. However, since GAUL provides

surface polygons with most of the world countries,

the application can be extended to any country.

Additionally, we have demonstrated that reducers and

cloud masking contributed to the improvement of

Romania’s forest cover classification photos.

Furthermore, within the Google Earth Engine

platform, we have displayed images of forest gain and

loss from various years to provide an interactive map

to further enhance the findings of deforested areas.

2 RELATED WORK

Article (Walquer Huacani, Meza, Aguirre, Sanchez,

and Luque, 2022) examines the deforestation of the

Apurimac region's forest cover from 2001 to 2020

using the Google Earth Engine, GEE platform. The

research employs a supervised classification method

based on a decision tree developed by the University

of Maryland. The authors utilize Landsat 7 and 8

satellite images band channels, processed and

updated to determine the gross forest cover loss data

for the period 2001-2020. Descriptive statistics are

applied to analyse the variables and establish

potential correlations. The study reveals a deforested

area of 3958.231 hectares, with an annual

deforestation rate of 109.15% in 2017 and a recovery

rate of -67.05% in 2018. However, while the study

provides valuable insights into deforestation trends, it

does not incorporate any specific accuracy metrics or

validation measures.

The study by (Teodorescu and Voicu, 2021) uses

the Google Earth Engine API to classify the forest and

analyse the deforestation. Satellite pictures may be

used for forest analysis, deforestation tracking, water-

covered regions, land use change, land cover, land

health evaluation, and other purposes. The Moderate

Resolution Imaging Spectroradiometer (MODIS) is a

sensor-equipped instrument that has been collecting

images since 2000, daily images, surface refraction

adjusted BRDF for 15 days, reflection factor, which

is the proportion of light reflected on the surface of a

material, and by-products such as indications of

vegetation or snow cover (Schreier, Ghazaeyan, and

Dubovyk, 2021). While Teodorescu and Voicu,

(2021) focused on mapping specific areas in

Romania, such as Bucharest and the Carpathian

Mountains, our research expands this analysis to a

significantly larger area, encompassing the entire

country to provide a more comprehensive assessment

of deforestation patterns.

3 AIDED (ANALYZING

DEFORSTATION DYNAMICS)

3.1 Defining the AOI (Area of Interest)

AOI is the short form of area of interest. In Google

Earth Engine, there are two ways to define the area of

interest, which is either by manually drawing the

surface, represented by a polygon, or by importing an

already defined polygon with its coordinates. Since

the area of interest in our case was the Romania’s

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

844

map, we have used an already existing dataset for

defining countries boundaries.

The dataset which contains all the countries

information is called FAO GAUL (the Global

Administrative Unit Layers). GAUL's objective

(FAO GAUL, 2015) is to provide global layers with

a comprehensive and up-to-date collection of units at

the first and second administrative levels.

Based on (Giurca and Dima, 2022) Romania’s

officially recognized land area is 238,397 square km

and the imported GAUL Romania area leads to

237,542 square km. This discrepancy of 855 square

km highlights potential differences in data sources,

projection systems, or boundary delineation methods

used in global datasets compared to official national

records. As it represents only about 0.36 percent of

Romania’s total area, we have considered this to be a

strong dataset to be used for filtering the images to

Romania’s map.

3.2 Dataset

Regarding the dataset which contains the satellite

images, Landsat, provided by NASA, has been used

to help in detecting the deforestation in Romania. To

enhance the forest classification, the dataset was pre-

processed before feeding it to a Random Forest

algorithm. The pre-process steps included scaling

factors and computing the NDVI band. To cover a

large timeframe, we have used four Landsat datasets

(Google Developers, n.d.), Landsat 5 from 1984-

2012, Landsat 7 from 1999-2021, Landsat 8 from

2013-Present and Landsat 9 from 2021-Present.

The 10 percent scaling factor for Cloud or Snow

Cover to all images from Landsat’s Image

Collections, implies that if the target pixel has a

percentage more than 10 of cloud or snow cover, it

will be filtered out to improve the forest

classification. The value was chosen based on

established remote sensing practices for atmospheric

interference and cloud contamination in optical

satellite imagery. (Huete, Didan, Miura, Rodriguez,

Gao, and Ferreira, 2002). Furthermore, we computed

the NDVI band (Rouse, Haas, Schell, and Deering,

1973) for the images with the goal to improve forest

classification by storing the Normalized Difference

Vegetation Index in a band for each image.

The scaling factor of 0.0000275 was used to

multiply the pixel value of 17,321 and add an

additional offset of -0.2. This calculation results in a

reflectance value of 0.19952 once the scale factor is

incorporated (U.S. Geological Survey, 2021) into the

Landsat 8 dataset. This technique is used to convert

raw digital numbers into top-of-atmosphere

reflectance values, ensuring consistency in spectral

data interpretation.

3.3 Creating the Training and Testing

Dataset

To create the training and testing datasets, we

manually pointed on the Google Earth Engine’s map,

two set of points. The forest points which contain the

label forest and have the property landcover 1 and the

‘no forest’ points with the label of no_forest and

property landcover 0, this landcover property is used

for the classifier to successfully classify the pixels as

forest or not forest. After the machine learning

algorithm successfully categorized the pixels of the

two datasets (e.g., 2020 and 2021), we have computed

the difference of pixel property between the first and

second year:

• -1 indicates that the pixel was not forest in 2020

but is forest in 2021, resulting in forest gain.

• 0 indicates that the pixel was forest in 2020 and

the same in 2021, resulting that the forest remained.

• 1 indicates that the pixel was forest in 2020 but

not forest in 2021, resulting in forest loss.

3.4 Random Forest Classifier

For the machine learning algorithm, we have used the

existing model provided by the Google Earth Engine,

named smileRandomForest. The train and test data

were split randomly by 80 and 20 rule. Since we were

using Landsat images, a scale of 30 was used among

with the following bands Red SR_B4, Green SR_B3,

Blue SR_B2, Near-Infrared SR_B5 and NDVI. This

principle is similar to a connected graph, where the

leaves are cut. We have displayed the difference of

the classified images before the reducer Figure 2. and

after applying the reducer Figure 3. We can observe

that the reducer improved significantly in filtering out

the misclassified pixels or the isolated pixels. The

validation

points

the

same

properties

the

Figure 2: Forest classification on Romania without the

reducer applied.

Analyzing Deforestation Dynamics in Romania Using Random Forest Algorithm and Google Earth Engine

845

forest and non-forest points, but they were manually

selected for each specific year to validate the presence

of forested and non-forested areas.

To remove the isolated pixels we have applied a

reducer, by using the focalMode built-in Earth Engine

function with the kernel of 3x3, resulting in isolated

or incorrectly classified pixels being merged or

ignored. This function clears these inconsistencies by

replacing each pixel’s value with the most frequent

class in its surrounding neighbourhood, ensuring

better spatial overview. This has improved the

detection of forested areas, since it occurs in clusters,

rather than as isolated patches.

Figure 3: Forest classification on Romania with the isolated

pixels filtered out.

4 EXPERIMENTS

To evaluate the results of the classifier, in this article

we have used the following metrics:

• accuracy, which has been calculated as

proportion of correctly classified pixels over every

correctly classified pixels

𝑂𝐴 

𝑎𝑙𝑙 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠

𝑎𝑙𝑙 𝑝𝑖𝑥𝑒𝑙𝑠

Cohen's Kappa coefficient (Cohen, 1960) is a

widely utilized metric for evaluating the consistency

between two datasets. It is referred as more reliable

than basic percentage agreement, because it adjusts

for agreements that may happen randomly (Vieira,

Kaymak, and Sousa, 2010).

Because the Google Earth Engine environment is

limited with the Machine Learning models, we could

only choose (at the time when we have implemented

the algorithm) between SVM, Random Forest and

Naïve Bayes. With the help of article (Tamiminia,

Salehi, Mahdianpari, Quackenbush, Adeli, and

Brisco, 2020) where it states that the best Image

Classification algorithm for forest detection is

Random Forest by the results shown in the article, we

tested only the smileRandomForest classifier. In

Table 2. are shown the best results of each Landsat

dataset for the best years.

Table 2: The results metrics of the best studied years.

Year Landsat

Model

Train

Cohe

n’s

Kapp

Valid

ation

Accur

acy

Validation

Cohen’s

Kappa

2000

Landsat

5 & 7

99.29

78.80

52.9308

2001 Landsat

5 &

99.43

72.37

40.6656

2002

Landsat

5 & 7

99.43

72.44

40.9654

2015

Landsat

5 & 7

99.61

91.67

79.5336

2016

Landsat

5 & 7

99.60

90.58

76.6478

2017

Landsat

99.56

92.07

80.6247

2018

Landsat

99.60

91.94

80.2464

2021

Landsat

8 & 9

99.49

94.71

87.3172

2022

Landsat

8 & 9

99.43

94.92

87.8405

We can observe a big difference between Landsat

5 & 7 and Landsat 8 & 9, the best Cohen’s Kappa

accuracy on validation dataset performed from

Landsat 5 & 7 is 79.53% in year 2015, while the best

validation Cohen’s Kappa accuracy from Landsat 8 &

9 was in the previous year (2022). This big difference

is due to the latest satellites sensors technologies

which improved the band and NDVI images.

The training scores exceeds 99%, which indicates

that the model indicates that the model could be

overfit on the train data. We are aware that this issue

likely arose from the reflectance of sun in the images

which impacted the NDVI, and other bands used in

the classifier.

Since the framework allows us to display maps in

such a way that users can interact by dragging and

zooming across the map, it was necessary to have a

visual representation of the forest situation in

Romania. Figure 4. shows the final map of forest loss

(red) and forest gain (green) in Romania from the year

2022. This was achieved using Landsat 8 and Landsat

9 images provided by Google Earth Engine.

ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering

846

Figure 4: Displayed map of forest loss/gain of Romania in

the year 2022.

5 CONCLUSION

The purpose of this article was to examine the

dynamics of deforestation in Romania utilizing the

Google Earth Engine platform and the Random Forest

algorithm. Climate change and deforestation being

the latest topics discussed worldwide every day, this

study aimed to better understand the patterns and

sources of deforestation, as well as to provide insights

for forest cover detection in Romania.

The use of Landsat images, with an applied cloud

masking of 10 percent, scaling factors and with the

computed Normalized Difference Vegetation Index

(NDVI) used with the Random Forest classifier to

analyse deforestation dynamics in Romania found

considerable deforestation patterns. The investigation

identified specific areas with high rates of

deforestation, underlining the importance of focused

conservation initiatives. Agriculture growth,

infrastructural development, and illegal harvesting

have all been cited as major sources of deforestation

in the nation. The Random Forest method was shown

to be successful in classifying the forested area across

Romania. Because of its capacity to handle

complicated interactions between data, it was

possible to accurately classify and forecast

deforestation regions. The utilization of Google Earth

Engine, with its large data store and cloud-based

computing capabilities, was critical in doing the

research at scale.

Restricted hardware resources and limitations of

Google Earth Engine client side made the purpose of

this research to be limited for Romania’s map. To

combat this, another approach is to utilize the Google

Earth Engine API to retrieve the data on the local side,

process the data using TensorFlow or Keras and then

load it back in the Google Earth Engine using S3

Buckets.

REFERENCES

Cohen, J. (1960). Coefficient of agreement for nominal

scales. Educational and Psychological Measurement,

37(1), 37–46.

FAO GAUL: Global Administrative Unit Layers. (2015).

First-Level Administrative Units | Earth Engine Data

Catalog | Google for Developers. Google for

Developers.

Giurca, A. and Dima, D.-P. (Eds.). (2022). The Plan B for

Romania’s Forests and Society. Transilvania

University Press.

Google Developers. (n.d.). Landsat Collections in Earth

Engine. Earth Engine Data Catalog.

Huete, A., Didan, K., Miura, T., Rodriguez, E., Gao, X., and

Ferreira, L. (2002). Overview of the radiometric and

biophysical performance of the MODIS vegetation

indices. Remote Sensing of Environment.

Kozak, J., Ostapowicz, K., Bytnerowicz, A., and Wyżga, B.

(2013). The Carpathian Mountains: Challenges for the

central and eastern European landmark. In J. Kozak, K.

Ostapowicz, A. Bytnerowicz, and B. Wyżga (Eds.), The

Carpathians: Integrating nature and society towards

sustainability (pp. 1–12). Springer-Verlag.

Rouse, J., Haas, R., Schell, J., and Deering, D. (1973,

December 10–14). Monitoring vegetation systems in

the Great Plains with ERTS. Proceedings of the Third

ERTS Symposium, NASA, Washington, DC, USA.

Schreier, J., Ghazaeyan, G., and Dubovyk, O. (2021). Crop-

specific phenomapping by fusing Landsat and Sentinel

data with MODIS time series. European Journal of

Remote Sensing, 54, 48-55.

Tamiminia, H., Salehi, B., Mahdianpari, M., Quackenbush,

L., Adeli, S., and Brisco, B. (2020). Google Earth

Engine for geo-big data applications: A meta-analysis

and systematic review. ISPRS Journal of

Photogrammetry and Remote Sensing, 164, 152–170.

Teodorescu, P. and Voicu, S.-N. (2021). Image

classification using machine learning algorithms in

Google Earth Engine environment. Informatica

Economica, 25(3), 5-16.

U.S. Geological Survey. (2021, January 7). How do I use a

scale factor with Landsat Level-2 science products?

Vieira, S. M., Kaymak, U., and Sousa, J. M. C. (2010).

Cohen’s kappa coefficient as a performance measure

for feature selection. International Conference on

Fuzzy Systems, Barcelona, Spain.

Walquer Huacani, Meza, N. P., Aguirre, F., Sanchez, D. D.,

and Luque, E. N. (2022). Analysis of deforested area

using Google Earth Engine in the period 2001–2020 in

the Apurimac region. Journal of Internet and e-

Business Studies, 2022, Article ID 988227S.

Analyzing Deforestation Dynamics in Romania Using Random Forest Algorithm and Google Earth Engine

847