Exploring Spectral Data, Change Detection Information and

Trajectories for Land Cover Monitoring over a Fire-Prone Area

of Portugal

André Alves

, Daniel Moraes

2,3 b

, Bruno Barbosa

, Hugo Costa

2,3 d

, Francisco D. Moreira

Pedro Benevides

, Mário Caetano

2,3 g

and

Manuel Campagnolo

Forest Research Centre, Associate Laboratory TERRA, School of Agriculture, University of Lisbon,

Tapada da Ajuda, 1349-017 Lisboa, Portugal

NOVA Information Management School (NOVA IMS), Universidade NOVA de Lisboa,

Campus de Campolide, 1070-312 Lisboa, Portugal

Direção-Geral do Território, Rua Artilharia Um, 107, 1099-052 Lisboa, Portugal

Keywords: Land Cover Change Classification, Thematic Map, Spectral Composites, NDVI, CCDC, COSc, Earth Observation.

Abstract: Land use/land cover (LULC) change detection and classification in maps based on automated data processing

are becoming increasingly sophisticated in Earth Observation (EO). There is a growing number of annual

maps available, with diverse but related production structures consisting primarily of classification and post-

classification phases, the latter of which deals with inaccuracies of the first. The methodology production of

the “Carta de Ocupação do Solo conjuntural” (COSc), a thematic land cover map of continental Portugal

produced by the Directorate-General for Territory (DGT) mostly based on Sentinel-2 images classification,

includes a semi-automatic phase of correction that combines expert knowledge and ancillary data in if-then-

else rules validated by photointerpretation. Although this approach reduces misclassifications from an initial

Random Forest (RF) prediction map, improving consistency between years and compliance with ecological

succession, requires a lot of time-consuming semi-automatic procedures. This work evaluates the relevance

of exploring an additional set of variables for automatic classification over disturbance-prone areas. A

multitemporal dataset with 124 variables was analysed using data dimensionality reduction techniques,

resulting in the identification of 35 major explanatory indicators, which were then used as inputs for RF

classification with cross-validation. The estimated importance of the explanatory variables shows that

composites of spectral bands, which are already included in the current COSc workflow, in conjunction with

the inclusion of additional data namely, historical land cover information and change detection coefficients,

from the Continuous Change Detection and Classification (CCDC) algorithm, are relevant for predicting land

cover classes after disturbance. Since map updating is a more challenging task for disturbed pixels, we focused

our analysis on locations where COSc indicated potential land cover change. Nonetheless, the overall

classification accuracy for our experiments was 72.34 % which is similar to the accuracy of COSc for this

region of Portugal. The findings suggest new variables that could improve future COSc maps.

1 INTRODUCTION

Land use/land cover (LULC) products by remote

sensing and satellite image classification are

https://orcid.org/0000-0002-8979-8906

https://orcid.org/0000-0002-4568-8182

https://orcid.org/0000-0001-6298-041X

https://orcid.org/0000-0001-6207-8223

https://orcid.org/0000-0001-7213-5551

https://orcid.org/0000-0001-5858-6815

https://orcid.org/0000-0001-8913-7342

https://orcid.org/0000-0002-9634-3061

becoming increasingly accurate in land

representation. Earth observation (EO) has seen

significant practical and theoretical advances as data

and machine-learning tools have become more

Alves, A., Moraes, D., Barbosa, B., Costa, H., Moreira, F., Benevides, P., Caetano, M. and Campagnolo, M.

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal.

DOI: 10.5220/0011993100003473

In Proceedings of the 9th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2023), pages 87-97

ISBN: 978-989-758-649-1; ISSN: 2184-500X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

accessible (Wulder et al., 2018). In the age of big data,

these map products are becoming more refined in

terms of both thematic and spatial detail, with

increased class heterogeneity. Simultaneously, with

the growing number of multi-annual maps (Brown et

al., 2020; Buchhorn et al., 2020; Hermosilla et al.,

2018) and near-real-time products (Brown et al.,

2022), the temporal aspect has also received attention.

However, measuring vegetation recovery and

predicting post-disturbance classes in line with

ecological succession principles, as well as other

concerns of temporal consistency of land cover

change direction, remains a challenge in time series

of land cover (Bartels et al., 2016; White et al., 2022;

Wulder et al., 2018).

Land succession processes are complex since they

can be of high or low magnitude, abrupt or subtle, and

completed in months or take several years (Zhu et al.,

2022). Also, misclassifications can be numerous

because of spectral confusion, land cover

heterogeneity within the pixel resolution, different

phases of vegetation growth and the broad conceptual

definition of classes (e.g., bushland, scrubland,

moorland, shrub-steppe, etc., are normally classified

in the same class). Several examples of works

presenting methodological contributions to post-

classification error reduction, improving annual

consistency and better-predicting land cover change

trajectories, can be found in the literature. Annual

filters limiting LULC misclassifications (Franklin et

al., 2015), spatial-temporal joint classifications (Cai

et al., 2014) and time-series post-classification

(Hermosilla et al., 2018), are a few examples that

illustrate how diverse the proposed methodologies

are. Those examples used several dimensions of data

for consistency refinement. Ranging from transition

rules based on prior knowledge, spectral and change

detection information, historical land cover and class

membership data, to contextual information of

adjacent pixels, a wide range of potential features can

be contemplated to more accurately predict the land

cover class in a context of ecological succession.

This study intends to explore an approach to

improve the “Carta de Ocupação do Solo

conjuntural” (COSc). The COSc is a 10-meter raster

thematic annual map of land cover for continental

Portugal with 15 classes. As the outcome of a

supervised classification of satellite images and

ancillary data, its production settles on a multi-stage

workflow with preliminary maps being produced

along the way (Costa et al., 2022b). The first phase is

the COScA, a step of automatic data processing that

consists of a supervised classification with the

Random Forest (RF) algorithm to classify land cover

classes based on stratified point samples.

Subsequently, a semi-automatic phase takes place,

COScR, minimizing misclassifications through

expert knowledge implemented by if-then-else rules.

It is estimated that COScR map is at least 13 % more

accurate than COScA (Costa et al., 2022b).

Subsequently, the intra-annual vegetation losses,

such as wildfires during summer, are assessed in the

COScP phase. Finally, a harmonization process over

landscape units concludes the COSc workflow

(COScH).

The inputs for the automatic steps of the COSc

workflow are partly derived from the time series of

Sentinel-2 imagery. COScA uses as inputs monthly

composites and COScP relies on detected temporal

breaks of a vegetation index to identify potential

intra-annual vegetation losses. The COSc map has

been made available to users since 2018 and the

release of successive yearly products follows a “map

updating” strategy, where only pixels with evidence

of change are possibly reclassified. More precisely,

successive updates focus only on pixels expected to

change (e.g., fire scars) or identified as disturbed

according to a temporal analysis of the spectral

profile and assume that the rest of the map remains

unchanged (Costa et al., 2022a). COSc has been

updated to 2020, 2021 and the 2022 version is under

production.

The current work explores the suitability of a

broad range of variables as new input data for the

COSc workflow. Towards that end, a dataset was

created with a set of variables for land cover

classification, which includes spectral data, historical

land cover information, Markov chain transition

probabilities, and change detection information from

the Continuous Change Detection and Classification

(CCDC) algorithm, which also models the temporal

pattern of the land cover signal between transitions.

At the time of writing, none of the existing versions

of the COSc map were created using CCDC or past

land cover and their transitions, thus their potential

for classification is unknown. Since our set of

variables is very large and exhibits some high

correlations, data was processed to identify

intercorrelated groups of variables, reduce the dataset

dimensionality, and minimise multicollinearity when

assessing for variables' importance.

Our experiments were made using a reference

dataset for a forest fire-prone region in the Center of

Portugal (tile 29TNE), where high-intensity large

fires occurred in 2017. Our reference data included

300 plots, distributed over fire-affected and non-fire-

affected areas, and was created through

photointerpretation of a temporal series of Sentinel-2

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

false colour (RGB 843) composites from September

2018 to September 2021 and orthophotos from the

summer of 2018 and 2021.

The objectives of the paper are twofold: (i) assess

the accuracy of fully automatic land cover

classification; (ii) identify the relevance of new

candidate variables to be included in the COSc

workflow. The remainder of the paper is organized as

follows: Section 2 – Data and Methods, Section 3 –

Results, Section 4 – Discussion and Section 5 –

Conclusions.

2 DATA AND METHODS

The methodological approach was supported in

Python environment. First, a hierarchical clustering

analysis identified groups of highly correlated

explanatory variables, that were subsequently

generalised by a principal component analysis (PCA)

to obtain a parsimonious dataset. The supervised

classification was performed using RF. Spectral,

thematic, transition probabilities and change

detection information were considered as inputs to

predict the 2021 class given by photointerpretation.

The accuracy of the model was evaluated using

traditional metrics and the calculation of Receiver

Operating Characteristic (ROC) curves. Feature

importance was assessed, and the results were

mapped.

2.1 Case Study Area

This research was conducted on the Sentinel-2 tile

29TNE, which covers a ~100 km wide swath along

the West coast of Central Portugal (Figure 1). This

area of continental Portugal landscape is

characterized by the coexistence of dense forest,

agricultural areas and urban land use bordering

wilderness (urban-rural interface areas). According to

Alves et al. (2022), there have been significant

changes in forest ecosystems in this area over the past

few decades. One of the most notable changes has

been the decline of maritime pine forests, as the result

of the expansion of eucalyptus plantations which now

dominate the region (Alves et al., 2022). Furthermore,

the study area was affected by large forest fires in

2017 (San-Miguel-Ayanz et al., 2020). As a result, it

is a dynamic area with the potential for land change

due to vegetation gains.

To run classification tests, the study area was

sampled with 300 circular plots with a 200-meter

radius. Those plots were randomly located over tile

29TNE, according to a stratification that separated the

sample into 150 plots that exhibited a spectral change

in the agricultural years from 2018 to 2021 and 150

that did not. The occurrence of change was identified

with the CCDC algorithm applied to the full

Sentinel-2 Level-2A time series, which initiates in

early 2017.

After the 300 plots were chosen as described

above, a team of photointerpreters relied on a series

of monthly Sentinel-2 temporal composites from

September 2018 to September 2021 as well as 25 cm

orthophotos from the summer of 2018 and 2021 to

segment those plots based on land cover and actual

occurrence of change. When a change was identified

by the photointerpreter, the thematic class for the

polygon pre-change and post-change was registered

according to the COSc legend. Therefore, the

reference dataset includes the class (minimum

mapping unit of 0.5 ha) for polygons that registered

changes from September 2018 to September 2021.

From these polygons originally 380,951 pixels were

derived (10-meter spatial resolution). The official

COSc2021 was used to label the polygons identified

Figure 1: Definition of the study area: a) Plots in tile

29TNE; b) Change/No-change polygons; c) Disturbed

pixels under study; d) COSc map land cover classes

considered.

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal

as “no change” when the dominant class remained

equal to 2020 and covered more than 80 % of the

polygon area. “No change” polygons not meeting that

threshold were not included in our study area. In

addition, only pixels with potential vegetation gains

were kept (COScA i.e., areas where a disturbance

occurred in the last years (2015-2020) and COScP

i.e., losses (clear-cuts and fires) in the 2021

agricultural year). From the 15 classes of the COSc

legend, only 6 classes were relevant over the study

area: bare soil, natural grassland, shrubland,

eucalyptus, other broadleaves and maritime pine. As

a result, the analysis focused on those 6 land cover

classes (see Figure 1).

After these conditions were imposed, 147,060

pixels remained for the classification experiments

which covered about 14,7 km

. From 2020 to 2021,

according to the photointerpreters, 73945 pixels

changed land cover class, of which 20688 registered

a loss of vegetation while 53257 had some type of

growth.

2.2 Data

The multidimensional dataset (Table 1) was based on

a literature review of studies that address

multitemporal map production, annual consistency

refinement and ecological succession (Abercrombie

& Friedl, 2016; Franklin et al., 2015; Hermosilla et

al., 2018; Liu et al., 2021; Xian et al., 2022; Xie et al.,

2022). The 124 selected variables arose from a

compromise between information that is not

considered for the current COSc classification and

that could be easily generated without disrupting the

existing workflow. Furthermore, the new variables

were supposed to inform about the ecological

succession trajectory and the type of land cover

change. All information was resampled to the 10-m

spatial resolution of Sentinel-2 visible and near-

infrared (NIR) bands.

Table 1: Per-pixel raw data.

Dimension Variable

Historical land

cove

COSc2018 and 2020 land cover

classes

Class transition

likelihoo

Markovian Conditional

Probabilities 2018-2020

Spectral images

B2, B3, B4, B8, B11, B12 and

Normalized Difference

Vegetation Index (NDVI)

monthly composites

Change detection

information

(CCDC)

Annual Cosine term of B2, B3,

B4, B8, B11, B12 and NDVI

Annual Sine term of mentioned

ands and NDVI

Magnitude change (mentioned

ands and NDVI

)

Time trend slope (mentioned

ands and NDVI)

Time trend intercept (mentioned

ands and NDVI

)

Time duration of the last

ment

Number of detected breaks

Chan

e in the

ear 2021

2.2.1 Historical Land Cover Information

The incorporation of prior land cover classes

constitutes information relevant for predicting the

next class and its inclusion in classification and post-

classification processes is not atypical (Cai et al.,

2014; Reis et al., 2020). The inclusion of previous

classes can be informative for two main reasons. On

the one hand because if a specific tree species class

has already occurred in a determined location, it is

more likely to occur in the future than other tree

species. On the other hand, the model can learn what

were the most common transitions to occur in the

past, giving it greater confidence in predicting the

future class. In this sense, the COSc2018 and 2020

classes were used as inputs for classification.

2.2.2 Class Transition Likelihood

In research aimed at enhancing the consistency of

annual maps, transition probabilities are typically

used as input variables in post-classification schemes

(Gong et al., 2017). In our approach transition

probabilities were derived using Markov Chains

based on the Markovian transition estimator in

IDRISI Selva. The reference images used were the

2018 and 2020 COSc maps of the study area, but

since our objective was to predict the land cover class

in 2021 it was defined that the period to project

forward from the second image was only one year.

The probability calculation assumed that the existing

thematic error in COSc did not propagate over time.

2.2.3 Spectral Images

The spectral variables used correspond to the

Sentinel-2 bands 2 (blue), 3 (green), 4 (red), 8 (near-

infrared), 11 and 12 (short-wave infrared). In

addition, the Normalized Difference Vegetation

Index (NDVI), with a known strong capacity for

discriminating phenological profiles of different land

covers (Balata et al., 2022; García et al., 2019), was

calculated. These spectral data consisted of monthly

composites computed using the median value of

observations covering the 2021 agricultural year

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

(October 2020 until September 2021), with the

Sentinel-2 Surface Reflectance image collection from

Earth Engine Datasets. The s2cloudless was used as

the cloud filter and gaps due to missing data were

filled based on interpolation using a harmonic model.

2.2.4 Change Detection Information

The inclusion of change information in land cover

classification can enable a result with a lower degree

of misclassifications when considering areas

experiencing changes since it is informative about the

trajectory of each pixel. Change information was

calculated using the CCDC algorithm (Zhu &

Woodcock, 2014) in Google Earth Engine. CCDC

uses linear harmonic models to detect breaks in time

series based on EO data. In particular, the harmonic

terms are modelled by periodic functions of sine and

cosine for varying periods, although we only use the

annual terms (see Table 1). The algorithm was

executed on the Sentinel-2 Surface Reflectance image

collection and its parameterization is exhibited in

Table 2.

The CCDC algorithm was not used to detect new

areas with disturbance, since the study area mask was

already defined based on the current COSc

framework, but for deriving relevant information for

classification. Although CCDC is designed to detect

more than one change per pixel, and model all

temporal segments between detected breaks in the

time series, we restricted our analysis just to the most

recent segment.

Table 2: Parameters used to run the CCDC algorithm.

Paramete

Value

Lambda 50

Chi-s

uare 0.995

Minimum number of

ears facto

Minimum observations 6

Maximum iterations 25000

Breakpoint bands B3, B12, NDVI

TMask bands B3, B12

The CCDC estimated coefficients (intercept,

trend, and annual periodicity fitted with sine and

cosine terms) model the pattern of the time series after

the most recent disturbance and are suitable to land

cover classification processes (Xian et al., 2022). The

coefficients represent a temporal segment between

two breaks. If no disturbance is detected by the

algorithm since the beginning of the time series

(which is early 2017 for Sentinel-2 surface

reflectance data) the full series corresponds to the

most recent segment. However, we used the total

number of detected breaks for the whole time series

as a proxy for the overall frequency of disturbance at

the pixel location.

2.3 Data Dimensionality Reduction

To obtain a parsimonious model, avoiding data

redundancy and multicollinearity when measuring

the feature importance, a double-step dimensionality

reduction approach was used. First, a hierarchical

cluster analysis was performed using all variables'

correlation as the distance matrix, applying Ward’s

aggregation rule. This allowed us to identify groups

of highly intercorrelated variables. The variable that

represents each group was defined as its principal

component according to a PCA. Due to their scales,

the prior land cover class (2 variables), Markovian

probabilities (6 variables, 1 for each class), and the

binary variable of change in 2021 were set aside and

added later to the representative variables determined

as described above. This approach resulted in a total

of 26 new variables (derived from clustering and

PCA) plus the 9 ones that had been set aside (see

Table 1).

2.4 Random Forest Classification and

Accuracy Assessment

We used the sklearn library in Python (Pedregosa et

al., 2011) to run RF classifications with stratified 10-

fold cross-validation. Data were partitioned into

training and testing polygons stratified by the 2021

class, ensuring that polygons used for training would

not be used for testing, and vice versa.

Parametrization ensured a maximum number of 300

trees and

√

𝑛

as the number of features available at

each split.

The measurement of variable importance was

obtained by the feature permutation algorithm,

determining the mean decrease in accuracy for each

variable. The permutation method to estimate

variable importance takes an explanatory variable x

and randomly shuffles its values in the dataset before

re-fitting the model. Then, it measures the increase in

prediction error regarding the model fitted to the

original dataset. Repeating this procedure for each

variable 𝑥 at a time estimates its importance and

compares it to the remaining variables.

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal

3 RESULTS

Results highlight that some variables, due to their

significance, have the potential for inclusion in future

COSc map production, specifically, by aiding in the

integration of land cover trajectory-related

parameters into the classification.

3.1 Accuracy

Table 3 shows the cross-validation accuracy

assessment with the RF classifier. The parsimonious

model achieved an accuracy of 72.34 % (± 1.86 % at

the 95 % confidence level). Pixels that were identified

by the photointerpreters as “change” exhibited a

slightly lower accuracy, particularly in the case of

vegetation gains.

Table 3: Model accuracy.

Type of pixels

Correctly classified

ixels (%)

Global 72.34

No chan

e in 2018-2021 77.30

Chan

e in 2018-2021 67.61

etation

ains 2020-2021 65.31

Vegetation losses 2020-2021 73.52

Land cover changes leading to loss of vegetation

(e.g., eucalyptus to bare soil; shrubland to natural

grassland, etc.) had higher accuracy than vegetation

gains. However, the model’s capacity to correctly

classify vegetation gains seems to have been

influenced by the type of occurrence that caused the

disturbance. According to Table 4, a greater

percentage of incorrectly classified pixels was related

to wildfires (60 %), whereas polygons with clear-cuts

had lower misclassification occurrences. Even though

these proportions apply to all pixels and not only

those with vegetation gains, it appears that the

classification outcome was influenced by the fact that

burned scars recover more slowly than areas that had

tree cutting.

Table 4: Error distribution by disturbance processes.

Type of disturbance

Incorrectly classified

ixels

(

)

No disturbance detected in

2018-2021

2.16

Fire in 2018-2021 59.89

Clea

-cut in 2018-2021 37.96

The land cover classes with the greatest accuracy

were shrubland and eucalyptus (Figure 2). The

classification of maritime pine and other broadleaves,

which are not abundant in the study area, were more

prone to errors.

Figure 2: Classification ROC curves.

3.2 Dimensionality Reduction and

Feature Importance

The data dimensionality reduction approach led to the

reduction of the original dataset of 124 variables to

meaningful groups of just 35 variables, which are

listed in Figure 3. For instance, the first group

includes monthly vegetation indices (NDVI) for all

spring and summer months, while the "Visible

summer” group contains all Sentinel-2 visible bands

for the summer months. Considering the variable

selection approach, our results show that NDVI

during spring and summer months, prior land cover

classes (particularly COSc2018), the length of the

CCDC's last segment and the number of valid breaks

had the larger importance for classification (Figure

3). Indicators from the CCDC algorithm and spectral

information stand out as those with the greatest

contribution to land cover monitoring.

Generally, the variables corresponding to winter

and fall months were of less importance. Transition

probabilities based on Markov Chains appeared with

low importance and those that seem to have been the

most informative for the model correspond to class

change to other broadleaves and eucalyptus, two

classes with spectral similarity issues.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

Figure 3: Permutation importance given by the mean

decrease in cross-validation estimated accuracy.

3.3 Spatial Error Distribution

The spatial distribution of classification errors

highlights the predominance of misclassifications

related to smaller polygons (Figure 4) and pixels near

polygon boundaries (Figure 5). In other words, the

model performed better on larger polygons and in

pixels distant from borders (i.e., less heterogeneous).

Despite the error rates being lower in non-forest

classes (Figure 2), their higher prevalence in the study

area caused the most misclassifications. More

specifically, the absolute number of mis-

classifications occurred predominantly in the classes

of natural grassland (23.69 %), shrubland (34.90 %),

and bare soil (29.78 %).

Polygons with high land cover heterogeneity and

diverse types of forest species at distinct stages of

evolution appear to have penalized the model since

several plots with error patches have been observed

as in Figure 6. This figure includes several polygons

that were misclassified as shrubland, suggesting that

our model tended to select that class over other

classes with a similar spectral signal. It also illustrates

an error type that may arise from the fact that

photointerpretation sampling had a minimum

mapping unit, which resulted in the inclusion of

multiple land covers inside the same polygon. In the

example in Figure 6, a eucalyptus stand is dominant

in a polygon but it is mixed with shrubland leading to

incorrect classification. The same explanation is valid

for other classes that exhibit a large enough degree of

heterogeneity.

Figure 4: Small polygons error pattern (green – correctly

classified; brown – classification error). The base map is a

DGT Orthophoto for 2021.

We stress that the validation partitioning approach

that was followed to obtain the results described

above ensured that pixels in the same polygon could

not be used for both training and testing, mitigating

spatial autocorrelation and potential overestimation

of classification accuracy.

Figure 5: Boundary error pattern (green – correctly

classified; brown – classification error). The base map is a

DGT Orthophoto for 2021.

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal

Figure 6: Mixed polygons error pattern (green – correctly

classified; brown – classification error; S – sample; P -

prediction). The base map is a DGT Orthophoto for 2021.

4 DISCUSSION

In this research, we have explored a set of variables

that can potentially improve the COSc annual map

classification and also reduce the manual work in the

post-classification stage. A multidimensional dataset

including time-series spectral and change detection

information, historical land cover, and transition

probabilities was created for this purpose. We

demonstrated how this methodological approach

yielded insights about information useful for a better

land cover change prediction.

4.1 Major Findings

Costa et al. (2022b) concluded that the accuracy of

the semi-automatic COSc thematic map for 2018 for

our study area (tile 29TNE) is close to 75 %. Even if

the results are not fully comparable (Costa et al.,

2022a), our findings suggest that it is possible to

obtain a similar accuracy using an automatic

approach by adding new variables to the COSc

workflow.

We showed (see Table 3) that for pixels where no

recent changes had been identified the accuracy is

highest (77 %) even if we restricted our analysis to

areas where evidence of change exists for the period

2015-2021 according to the COSc workflow. This

points out that automatic classification tends to have

higher accuracy when no recent change occurs, which

is expectable for change detection methods like

CCDC that perform better with a sufficiently long

series of observations to model the land cover signal

after a change. Since classification tends to be more

accurate for stable land covers, the obtained accuracy

results are likely to underestimate overall accuracy

since our reference dataset has a proportion of

changed land cover (50.3 %, according to the

photointerpreters) much higher than the average

proportion of changed pixels for Portugal over 2018-

2020 which is 5.4 % according to Costa et al. (2022a).

Change information, which Wulder et al. (2018)

stated as an essential component of modern land

cover monitoring, was especially pertinent to

discriminate the classes under study. CCDC

variables’ high importance confirms the potential of

this change detection technique to generate pertinent

data for land cover classification. The relevance of

CCDC outputs for classification has already been

demonstrated in LCMAP production (Xian et al.,

2022). However, the major novelty of our work in this

domain is the use of a short time series of Sentinel-2

images, while most work implementing CCDC relied

on long series of Landsat data (Franklin et al., 2015;

Xian et al., 2022; Xie et al., 2022). The importance of

the temporal component in the classification process,

considered by Gómez et al. (2016) as indispensable in

the current state of EO, was confirmed by the CCDC

because the coefficients used were of the last

segment, i.e., information after the last disturbance.

Other variables informed about trajectory parameters

and ecological succession. The historical information

revealed itself to be especially useful for the model to

learn the class that follows, both in terms of

vegetation losses and gains. In terms of spectral

information, the months that mark the biggest

difference in vegetation greenness of the land cover

classes (spring and summer) were the most important.

This outcome is not difficult to comprehend for the

Portuguese Mediterranean climate since the

interannual greenness variability of some classes is

more pronounced at the end of summer, when the

maximum dryness is reached, and at the beginning of

fall, when the growth in greenness resumes.

Transition probabilities were the least important data

dimension. Its limited contribution may be attributed

to its exclusive consideration of the study area's

overall transition probabilities, thereby failing to

distinguish contextual characteristics of pixel sets.

The spatial distribution of misclassifications

reflects essentially three situations: boundary pattern,

small polygon, and heterogeneous polygon.

Boundaries between different geographic features

can cause diverse spatial patterns in the occurrence of

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

errors. Since the study samples comprised polygons

with different shapes and sizes, the characteristics of

these polygons (such as their size, shape, and spatial

arrangement) influence the error patterns (Corcoran

et al., 2015; Moraes et al., 2021). In addition, due to

the minimum mapping unit of the polygons, a certain

level of spectral diversity was inevitable, which made

it more difficult to classify polygons with more

intricate land cover patterns.

Regarding the classification errors by land cover

types, the phenological profiles of natural grassland

and shrublands have common behaviours and spectral

similarities. The model overestimated the number of

shrubland pixels however, this class had high

accuracy. Because we are dealing with a dynamic

environment, the transition from herbaceous to

shrubland can be gradual and undetectable. As a

result, the model underperformed in classifying

natural grassland, with more than 40 % of pixels

misclassified as shrubland or bare soil. Since abrupt

changes (losses) were better classified than gains

(slower transitions), there is still a need to investigate

a wider collection of metrics that can assist in

monitoring more subtle changes and longer

succession processes.

4.2 Limitations and Future

Development

The ultimate goal of this research is to improve the

COSc workflow for the whole Portuguese territory.

Our study area was limited to a disturbance-prone

region previously identified by the COSc processing

steps. In future work, we will explore the suitability

of the new candidate variables not only for

classification but also for identifying pixels for map

updating.

Also, our approach was limited in describing

long-term ecological succession. The Markovian

transition probabilities, which we had anticipated to

be highly informative due to their importance in

previously cited studies, were found to have relatively

low levels of significance. This is partly because the

Markov Chain processes are memoryless, meaning

that the land cover transition probabilities to 2021

were independent of the previous classes in instances

before the period 2018-2020. Although with residual

importance, Figure 3 shows that the two most

important probabilities from the six variables were

transitions to other broadleaves and eucalyptus. These

two classes are often difficult to distinguish in the

COSc mapping due to their spectral similarities and

the presence of mixed forest patches in the study area,

which suggests that this type of stochastic transition

element could be explored to discriminate forest

species. If these probabilities are complemented by

additional information, such as per-pixel vegetation

growth suitability, their relevance may increase for

dealing with the unique context of some pixel sets.

Future work should focus on extending these

experiments to a larger area and with a higher number

of land cover classes. Additional spectral bands

relevant to land cover monitoring, such as the red-

edge, should be explored. It may also be crucial to

assess potential accuracy gains from the use of a

classifier with spatial dependency or the

incorporation of variables with contextual

information from neighbouring pixels.

5 CONCLUSIONS

The main goal of this work was to identify new

variables to improve the automatic steps of the COSc

workflow. This was accomplished by testing a set of

variables across multiple dimensions, including

previous land cover data, transition probabilities,

spectral and change detection information. The

analysis applied to a fire-prone area showed that some

variables not yet included in the COSc workflow,

specifically land cover classes in previous years and

change detection information produced by the CCDC,

have a high potential for improving the classification.

It was discovered that, when combined with Sentinel-

2 temporal composites, CCDC coefficients have high

importance, particularly the duration of the last

segment post-disturbance and the number of breaks in

the study period. Variable selection combining

hierarchical clustering and PCA effectively resulted

in a more parsimonious model without compromising

classification performance.

ACKNOWLEDGEMENTS

We thank the Directorate General for Territory's

young research fellows for their effort in drawing and

identifying the different types of land cover in

polygons.

FUNDING

This research was conducted under the collaboration

contract DGT-ISA 261/2021 with funding from

Compete2020 (POCI-05-5762-FSE-000368),

supported by the European Social Fund, and Centro

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal

de Investigação em Gestão de Informação (MagIC),

Project UIDB/00239/2020 (Forest Research Centre),

both supported by the Portuguese Foundation for

Science and Technology (FCT).

REFERENCES

Abercrombie, S. P., & Friedl, M. A. (2016). Improving the

Consistency of Multitemporal Land Cover Maps Using

a Hidden Markov Model. IEEE Transactions on

Geoscience and Remote Sensing, 54(2), 703–713.

https://doi.org/10.1109/TGRS.2015.2463689

Alves, A., Marcelino, F., Gomes, E., Rocha, J., & Caetano,

M. (2022). Spatiotemporal Land-Use Dynamics in

Continental Portugal 1995–2018. Sustainability,

14(23), 30. https://doi.org/10.3390/su142315540

Balata, D., Gama, I., Domingos, T., & Proença, V. (2022).

Using Satellite NDVI Time-Series to Monitor Grazing

Effects on Vegetation Productivity and Phenology in

Heterogeneous Mediterranean Forests. Remote

Sensing, 14(10), 2322. https://doi.org/10.3390/

rs14102322

Bartels, S. F., Chen, H. Y. H., Wulder, M. A., & White, J.

C. (2016). Trends in post-disturbance recovery rates of

Canada’s forests following wildfire and harvest. Forest

Ecology and Management, 361, 194–207.

https://doi.org/10.1016/j.foreco.2015.11.015

Brown, C. F., Brumby, S. P., Guzder-Williams, B., Birch,

T., Hyde, S. B., Mazzariello, J., Czerwinski, W.,

Pasquarella, V. J., Haertel, R., Ilyushchenko, S.,

Schwehr, K., Weisse, M., Stolle, F., Hanson, C.,

Guinan, O., Moore, R., & Tait, A. M. (2022). Dynamic

World, Near real-time global 10 m land use land cover

mapping. Scientific Data, 9(1), 251. https://doi.org/10.

1038/s41597-022-01307-4

Brown, J. F., Tollerud, H. J., Barber, C. P., Zhou, Q.,

Dwyer, J. L., Vogelmann, J. E., Loveland, T. R.,

Woodcock, C. E., Stehman, S. V., Zhu, Z., Pengra, B.

W., Smith, K., Horton, J. A., Xian, G., Auch, R. F.,

Sohl, T. L., Sayler, K. L., Gallant, A. L., Zelenak, D.,

… Rover, J. (2020). Lessons learned implementing an

operational continuous United States national land

change monitoring capability: The Land Change

Monitoring, Assessment, and Projection (LCMAP)

approach. Remote Sensing of Environment, 238,

111356. https://doi.org/10.1016/j.rse.2019.111356

Cai, S., Liu, D., Sulla-Menashe, D., & Friedl, M. A. (2014).

Enhancing MODIS land cover product with a spatial–

temporal modeling algorithm. Remote Sensing of

Environment, 147, 243–255. https://doi.org/10.1016/j.

rse.2014.03.012

Corcoran, J., Knight, J., Pelletier, K., Rampi, L., & Wang,

Y. (2015). The Effects of Point or Polygon Based

Training Data on RandomForest Classification

Accuracy of Wetlands. Remote Sensing, 7(4), 4002–

4025. https://doi.org/10.3390/rs70404002

Costa, H., Benevides, P. J., Moreira, F. D., & Caetano, M.

R. (2022a). Detection and classification of changes in

agriculture, forest, and shrublands for land cover map

updating in Portugal. Em C. M. Neale & A. Maltese

(Eds.), Remote Sensing for Agriculture, Ecosystems,

and Hydrology XXIV (p. 19). SPIE. https://doi.org/

10.1117/12.2636127

Costa, H., Benevides, P., Moreira, F. D., Moraes, D., &

Caetano, M. (2022b). Spatially Stratified and Multi-

Stage Approach for National Land Cover Mapping

Based on Sentinel-2 Data and Expert Knowledge.

Remote Sensing, 14(8), 1865. https://doi.org/10.

3390/rs14081865

Franklin, S. E., Ahmed, O. S., Wulder, M. A., White, J. C.,

Hermosilla, T., & Coops, N. C. (2015). Large Area

Mapping of Annual Land Cover Dynamics Using

Multitemporal Change Detection and Classification of

Landsat Time Series Data. Canadian Journal of Remote

Sensing, 41(4), 293–314. https://doi.org/10.

1080/07038992.2015.1089401

García, M., Moutahir, H., Casady, G., Bautista, S., &

Rodríguez, F. (2019). Using Hidden Markov Models

for Land Surface Phenology: An Evaluation Across a

Range of Land Cover Types in Southeast Spain. Remote

Sensing, 11(5), 507. https://doi.org/10.3390/rs

11050507

Gómez, C., White, J. C., & Wulder, M. A. (2016). Optical

remotely sensed time series data for land cover

classification: A review. ISPRS Journal of

Photogrammetry and Remote Sensing, 116, 55–72.

https://doi.org/10.1016/j.isprsjprs.2016.03.008

Gong, W., Fang, S., Yang, G., & Ge, M. (2017). Using a

Hidden Markov Model for Improving the Spatial-

Temporal Consistency of Time Series Land Cover

Classification. ISPRS International Journal of Geo-

Information, 6(10), 292. https://doi.org/10.3390/ijg

i6100292

Hermosilla, T., Wulder, M. A., White, J. C., Coops, N. C.,

& Hobart, G. W. (2018). Disturbance-Informed Annual

Land Cover Classification Maps of Canada’s Forested

Ecosystems for a 29-Year Landsat Time Series.

Canadian Journal of Remote Sensing, 44(1), 67–87.

https://doi.org/10.1080/07038992.2018.1437719

Liu, C., Song, W., Lu, C., & Xia, J. (2021). Spatial-

Temporal Hidden Markov Model for Land Cover

Classification Using Multitemporal Satellite Images.

IEEE Access, 9, 76493–76502. https://doi.org/1

0.1109/ACCESS.2021.3080926

Marcel Buchhorn, Lesiv, M., Tsendbazar, N.-E., Herold,

M., Bertels, L., & Smets, B. (2020). Copernicus Global

Land Cover Layers—Collection 2. Remote Sensing,

12(6), 1044. https://doi.org/10.3390/rs12061044

Moraes, D., Benevides, P., Costa, H., Moreira, F. D., &

Caetano, M. (2021). Influence of Sample Size in Land

Cover Classification Accuracy Using Random Forest and

Sentinel-2 Data in Portugal. 2021 IEEE International

Geoscience and Remote Sensing Symposium IGARSS,

4232–4235. https://doi.org/10.1109/IGARSS47720

.2021.9553924

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,

Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay,

É. (2011). Scikit-Learn: Machine Learning in Python. J.

Mach. Learn. Res., 12, 2825–2830.

GISTAM 2023 - 9th International Conference on Geographical Information Systems Theory, Applications and Management

Reis, M. S., Dutra, L. V., Escada, M. I. S., & Sant’anna, S. J.

S. (2020). Avoiding Invalid Transitions in Land Cover

Trajectory Classification With a Compound Maximum a

Posteriori Approach. IEEE Access, 8, 98787–98799.

https://doi.org/10.1109/ACCESS.2020.2997019

San-Miguel-Ayanz, J., Oom, D., Artes, T., Viegas, D.,

Fernandes, P., Faivre, N., Freire, S., Moore, P., Rego, F.,

& Castellnou, M. (2020). Forest fires in Portugal in 2017.

In Science for Disaster Risk Management 2020: Acting

today, protecting tomorrow. Publications Office of the

European Union. https://doi.org/10.2760/571085

White, J. C., Hermosilla, T., Wulder, M. A., & Coops, N. C.

(2022). Mapping, validating, and interpreting spatio-

temporal trends in post-disturbance forest recovery.

Remote Sensing of Environment, 271, 112904.

https://doi.org/10.1016/j.rse.2022.112904

Wulder, M. A., Coops, N. C., Roy, D. P., White, J. C., &

Hermosilla, T. (2018). Land cover 2.0. International

Journal of Remote Sensing, 39(12), 4254–4284.

https://doi.org/10.1080/01431161.2018.1452075

Xian, G. Z., Smith, K., Wellington, D., Horton, J., Zhou, Q.,

Li, C., Auch, R., Brown, J. F., Zhu, Z., & Reker, R. R.

(2022). Implementation of the CCDC algorithm to

produce the LCMAP Collection 1.0 annual land surface

change product. Earth System Science Data, 14(1), 143–

162. https://doi.org/10.5194/essd-14-143-2022

Xie, S., Liu, L., Zhang, X., & Yang, J. (2022). Mapping the

annual dynamics of land cover in Beijing from 2001 to

2020 using Landsat dense time series stack. ISPRS

Journal of Photogrammetry and Remote Sensing, 185,

201–218. https://doi.org/10.1016/j.isprsjprs.2022.01.014

Zhu, Z., Qiu, S., & Ye, S. (2022). Remote sensing of land

change: A multifaceted perspective. Remote Sensing of

Environment, 282, 113266. https://doi.org/10.1016/

j.rse.2022.113266

Zhu, Z., & Woodcock, C. E. (2014). Continuous change

detection and classification of land cover using all

available Landsat data. Remote Sensing of Environment,

144, 152–171. https://doi.org/10.1016 /j.rse.2014.01.011

Exploring Spectral Data, Change Detection Information and Trajectories for Land Cover Monitoring over a Fire-Prone Area of Portugal