1 Introduction

The Paraíba do Sul River Valley is a very affluent and prosperous region in Brazil, located between São Paulo and Rio de Janeiro states [1]. The region is the home of a huge piece of the Atlantic Forest, one of the major biodiversity hotspots in the world with many endemic and endangered species [2]. Despite its high biodiversity and ecological importance, the Atlantic Forest has experienced alarming levels of habitat destruction and fragmentation, mainly due to the disorderly growth of agricultural and industrial activities [3,4,5]. According to SOS Mata Atl\(\hat{a}\)ntica Foundation [6] – a Brazilian NGO dedicated to conserving this biome, the deforestation of Atlantic Forest is critical, as only \(12.4\%\) of the biome’s native vegetation still remains undisturbed. This devastation scenario has contributed to transform the forest landscape into anthropic spaces, as well as to promote the fragmentation of its green zones [7]. As a result, the biome formations are not very representative in the so-called Paraíba Valley, since they have been substantivally altered over the time so that only a reduced number of green areas remains totally preserved.

The region has undergone different changes over the past few centuries, from the coffee growing in 19th-20th centuries to the eucalyptus cultivation nowadays. Coffee crop had its ”gold century” between 1830-1930, promoting prosperity to the Paraíba Valley economy, but with the high price of Atlantic Forest deforestation. Due to the urban and industrial expansion, coffee culture was weakened, leading to the predominance of extensive pastures in the region. Moreover, the coffee productivity reduction trigged an intense rural exodus to large urban centers around the Paraíba Valley so that many of the abandoned lands were converted into small vegetation areas, thus partially contributing to the increase of forest coverage [8]. From 1966 to 1974, there was an exponential growth in eucalyptus culture, encouraged by tax incentives given by the Brazilian government for reforestation projects, in particular, eucalyptus crops. Nowadays, the Paraíba Valley has a dense concentration of eucalyptus plantations which is an important economic pilar for the region [7].

Among several technologies that may support Spatio-temporal analysis of Paraíba Valley, the use of change detection techniques arises as a convenient alternative. Such techniques allow the identification of differences in multitemporal imagery acquired by Remote Sensing. Moreover, if the identification process is unsupervised, no ground-truth data is required so that the output will be a mapping which highlights the regions of the image where changes have occurred. Unsupervised change detection techniques have been successfully applied for different purposes, such as crop mapping [9, 10], forestry monitoring [11, 12], natural disasters inspection [13, 14], and urban planning [15, 16].

Considering the highly changeable scenario as faced by Paraíba Valley during the last decades, a descriptive study focused on identifying spatial-temporal changes in land use/cover of Atlantic Forest is proposed in this paper. We take a remote sensing imagery to perform multi-temporal analysis of the region, allowing for capturing different representations of large aerial areas in the Atlantic Forest as the time varies.

More specifically, we tune and run three unsupervised change detection approaches so as to collect several data-driven evidences to support specialists and public agents in the task of monitoring the Atlantic Forest ecozone. Also, we provide pair-wise and multiple instant comparisons in order to properly determine the dynamics of land changes. In our study, we adopt aerial images collected from the Landsat-5 and -8 satellites by the Thematic Mapper (TM) and Operational Land Imager (OLI) sensors. In particular, we compare the years 1987, 1997, 2007 and 2017.

This paper is organized as follows: Sect. 2 brings an introductory review about the change detection methods we apply to generate the results, while Sect. 3 describes the study area, data sets, and methodological aspects of the study. Section 4 presents the results and their discussion. Finally, Sect. 5 summarizes our findings and conclusions.

2 Unsupervised Change Detection Techniques

Change detection accomplishes the identification of changes in land surfaces through images acquired during a given time period [17]. This task plays a fundamental role in environmental studies, for example, to track the ecosystem transitions, as well as to better comprehend the successive interaction between natural phenomena and human activities. Once these changes are identified and properly quantified, the land use and land cover analysis can be performed in a more effective and comprehensive way [18].

Several change detection techniques have been proposed in the literature. We can roughly group these techniques between supervised and unsupervised ones. Due to specific constraints and limitations, which include high financial costs and the annalistic efforts to get a large amount of data to validate a supervised method, recently the scientific community has focused on unsupervised learning-based strategies [19]. Indeed, the ability of identifying temporal changes without any additional data resource makes unsupervised methods more attractive than the supervised-based ones [20, 21]. On the other hand, the unsupervised change detection methods are not able to discriminate land cover types behind the change or non-change events.

Usually, these techniques embrace post-classification analysis, simple arithmetic operations, Principal Component Analysis (PCA) and morphological filters [22]. In the following, we discuss each one of these change detection strategies.

2.1 Change Vector Analysis

The so-called Change Vector Analysis (CVA) [23] is an unsupervised learning-based model which comprises three successive steps. First, geometric and radiometric corrections are performed on an input pair of images. Thereby, let us denote by \(\mathcal {I}^{(1)}\) and \(\mathcal {I}^{(2)}\) the pair of the calibrated images, defined on a support \(\mathcal {S} \subset \mathcal {N}^{2}\) of a given geographic region whose the position \(s \in \mathcal {S}\) (i.e., a pixel) in \(\mathcal {I}^{(1)}\) and \(\mathcal {I}^{(2)}\) defines attribute vectors \(\mathbf {x}_{s}^{(1)}\) and \(\mathbf {x}_{s}^{(2)}\), respectively. These vectors embed a particular measures computed from a specific target regarding different wavelength intervals.

Next, \(\mathcal {I}^{(1)}\) and \(\mathcal {I}^{(2)}\) are compared through the vectors \(\mathbf {x}_{s}^{(1)}\) and \(\mathbf {x}_{s}^{(2)}\) for all \(s \in \mathcal {S}\). The amplitude difference \(\Vert \mathbf {x}_{s}^{(1)} - \mathbf {x}_{s}^{(2)} \Vert\) gauges how similar \(\mathbf {x}_{s}^{(1)}\) and \(\mathbf {x}_{s}^{(2)}\) are, producing the resulting difference image \(\mathcal {I}^{(1-2)}\). Finally, the discrimination between the relevant and irrelevant features of the images is performed, by exploiting the difference image \(\mathcal {I}^{(1-2)}\). Relevant differences, usually with high magnitudes, stands for changes, while low magnitude differences correspond to non-change events. We denote changes and non-changes by the classes \(\omega _c\) and \(\omega _n\), respectively. The use of thresholding techniques, for example, the Otsu [24] or Kittler-Illinghworth [25] algorithms, allows learning-based strategies to distinguish between the low and high magnitude differences.

2.2 PCA and K-Means framework

The well-known mathematical model Principal Component Analysis (PCA) [26] has been the basis of many unsupervised change detection methods [27,28,29]. A good representative of PCA-based method is the one proposed in [30], which combines the versatility of PCA with the effectiveness of K-Means (KM) clustering [31] into a unique unsupervised change detection approach. Despite its simplicity in terms of algorithmic architecture, the method does a very good job in identifying changes while still keeping robust to noisy data.

Similar to CVA, typical PCA-based approaches rely on an amplitude difference image \(\mathcal {I}^{(1-2)}\) built from a pair of images \(\mathcal {I}^{(1)}\) and \(\mathcal {I}^{(2)}\) with support \(\mathcal {S}\). First, the domain \(\mathcal {S}\) is decomposed into non-overlapping squares of dimensions \(h \times h\), where \(h = 2k+1; k \ge 2\). Next, a vectorial representation for each image region in \(\mathcal {I}^{(1)}\) and \(\mathcal {I}^{(2)}\) is derived so that the PCA is computed on this set of vectors. This is also performed for each pixel s in the composite difference image \(\mathcal {S}\) of \(\mathcal {I}^{(1-2)}\). The extracted vectors are then projected onto a different feature space by the PCA technique, wherein the first p components are taken. Finally, the KM algorithm is applied on the projected data so that two clusters are derived. The cluster of the centroid closer to the origin of the projection space determines the set of non-change pixels (\(\omega _n\)), while the remaining cluster stands for the pixels that have been changed (\(\omega _c\)).

Here, we refer to the combination of PCA and K-Means as PCA-KM.

2.3 Alternating Sequential Filters for Change Detection

The use of morphological filters, especially the class of Alternating Sequential Filters (ASF), arises as a valid alternative to perform change detection tasks. ASF consists in a sequence of opening and closing morphologic operations so that the images are filtered several times in order to capture their objects of interests [32, 33]. By applying ASF, one can highlight the differences between images while preserving the geometric features of the targets in the image [19].

For instance, in Mura et al. [19], the authors proposed an unsupervised change detection method inspired on the ASF concepts. Analogously to CVA and PCA-KM, their method computes the amplitude difference image \(\mathcal {I}^{(1-2)}\), and then the ASF is applied to reduce the high-intensity values of \(\mathcal {I}^{(1-2)}\), as well as to preserve the geometric features of the difference image. The ASF filtering operations are successively applied based on gradual increases of the kernel size. Notice that ASF takes the number of iterations as a free parameter, which is used to regulate the change detection capability of the method.

Lastly, the Kittler-Illinghworth thresholding then applied to the ASF filtering result from \(\mathcal {I}^{(1-2)}\). While filtered pixel values inferior to the defined threshold are classified as non-changes (\(\omega _n\)), pixels whose values are greater than such threshold are assigned to change locations (\(\omega _c\)).

For the sake of simplicity, we denote this method as ASF.

3 Data and Experiment Design

In this section, we detail the data sets and the methodological steps we have conducted in this study. Specifically, Subsection 3.1 describes the study area and the examined data, while Subsection 3.2 presents the experiment design.

3.1 Study Area Description

The study area covers the Metropolitan Region of Paraíba do Sul River Valley, in Brazil, a region located between São Paulo and Rio de Janeiro cities which encompasses an area of 14,236 \(\text {km}^\text {2}\), with a population of approximately two million people [34]. Figure 1 depicts the study area location. According to Silva et al. [8], the study area has been undergoing a forest transition process, migrating from a period of constant reduction of native vegetation to the natural expansion of genuine forest territories, which is related to the abandonment of deep slope areas that are not compatible with mechanized agriculture, demographic and economic aspects.

As previously discussed in Sect. 1, Paraíba do Sul River Valley is the home of the Atlantic Forest, but nowadays such a green coverage of this biome is highly fragmented [6]. The current scenario is the result of intense deforestation caused by anthropic activities over the years [4].

Fig. 1
figure 1

The study area location

Full size image

We take images acquired by the TM and OLI sensors aboard the Landsat-5 and -8 satellites, respectively. These images refers to the study area in the years of 1987, 1997, 2007 and 2017. The aerial scenes from 1987 to 2007 were acquired by the TM sensor (Landsat-5), while the 2017 image was acquired by the OLI sensor (Landsat-8). The images have a spatial resolution of 30m and spectral bands from blue and short-wave infrared wavelengths. Additionally, the full data was obtained from the EarthExplorer platform [35] and express the surface reflectance after the atmospheric correction has been performed by LEDAPS [36] and LaSRC [37] algorithms. Finally, the output data produced by such algorithms were used to properly detect clouds and shadows masks in the images for each time instant.

Because of the large-size dimensions of the study area, we generate composite image mosaics to carefully inspect all the region. Four scenes were taken to produce each mosaic. These scenes were chosen in the months of August and September in order to ensure low occurrence of clouds and similar solar incidence. Figure 2 illustrates the resulting mosaics for the analyzed years. Each mosaic is defined on a wide support of \(7800 \times 4650\) pixels. The studied region has around 17.27 millions of pixels.

Assuming the interest in conducting quantitative assessments of the change detection results, it makes necessary the availability of ground-truth samples. Under this premise, a detailed inspection and comparison of the study area mosaics in different pairs of years were carried out, by selecting samples over locations where occur and do not occur changes. Also, the land use and land cover maps provide by MapBiomas project [38] supported the process of comparing pairs of instants and identifying candidate regions for samples collection. Table 1 summarizes the selected samples in the four pairs of periods. Notice that, once the sample dimensions are much smaller than the entire region, we can not illustrate in details the spatial distribution of these samples.

Fig. 2
figure 2

The study area time-serie mosaics

Full size image
Table 1 Summary of change and non-change samples in different periods
Full size table

3.2 Experiment Design

Figure 3 illustrates the experiment design of this study. From the mosaics in Fig. 2, and by applying the change detection methods already discussed in Sections 2.1-2.3, we generated temporal changes/non-changes maps for different pairs of instants. In particular, the following pair-wise years were considered: 1987–1997; 1997–2007; 2007–2017; and 1987–2017.

Fig. 3
figure 3

Experiment design flowchart

Full size image

Different parameter settings were tested for all the change detection methods. For the CVA approach, the parameter refers to the thresholding option in either Otsu or Kittler-Illinghworth algorithms. The values of h (squared region length) and p (number of principal components) for the PCA-KM were the following: \((h,p) \in \{(3,3),(3,5),(5,5)\}\). Concerning the ASF approach, the number of iterations ranged from 1 to 7. Notice that, for PCA-KM method, h has been limited to 5, as greater values for h may lead to feature spaces with very high dimensions, and consequently, computational prohibitive in face of the large image supports taken as input. Additionally, the values for p are limited by h, which have been empirically chosen after running successive experimental tests. Regarding the ASF method, it was observed from a preliminary battery of tests that iterations above 7 provided over-smoothed and unrealistic change/non-change mappings.

Once the change maps were obtained, and the parameters properly defined, we computed the accuracy assessment of the results, by comparing the outputs produced by the change detection methods with the ground-truth samples (see Table 1). This was quantitatively assessed via the F1-Score measure, a statistical-based metric that balances the true/false positive/negative rates [39]. The higher the F1-Score, the better the results. Finally, the best results for each period was taken to conduct a more detailed investigation about the spatio-temporal dynamics of the study area.

The change detection methods were coded in IDL (Interactive Data Language), as well as the assessments presented in this study. Data and codes are freely available at https://github.com/rogerionegri/UCD.

4 Results and Discussions

Following the experiment design (Subsection 3.2) and the multi-temporal data analysis (Subsection 3.1), a total of 48 change/non-change maps were generated (i.e., 4 periods \(\times\) 12 methods/configurations). Table 2 shows the accuracy values, in terms of F1-Score, computed from these change detection maps.

Table 2 F1-Score computed for the change/non-change maps under different periods, methods and parameter settings. The highest scores are highlighted in italic
Full size table

From the tabulated scores in Table 2, one can verify that the PCA-KM approach produced the lowest values in comparison with the other change detection strategies. Although PCA-KM has been reported in the literature as a very robust approach, we have empirically observed that the low performance of the technique may be due to the high contrast elements (e.g., clouds occurrence), which impair the computation of suitable clusters of change and non-change events.

Regarding the ASF approach, there was an inverse relation between accuracy and number of iterations: the accuracy increases as the number of iteration decreases. Indeed, this behavior is the result of successive applications of morphological operations. Also, notice that if one takes a single iteration, the ASF produces similar scores to the CVA method.

At last, one can see that the CVA approach (equipped with the Kittler-Illinghworth algorithm) achieves the highest accuracies in the study area. On the other hand, by applying the Otsu algorithm, the CVA technique delivered poor results. These are similar to the ones produced by PCA-KM for 1997–2007, 2007–2017 and 1987–2017 periods. Such a behavior may be explained by the occurrence of significant contrast differences, including pixel interferences after computing the best cutting-point by the Otsu algorithm.

Despite its simplistic architecture, the CVA turns out to be a very effective method for detecting changes in spatial-temporal images. Furthermore, the presence of clouds does not seem to influence the results achieved by “CVA + Kittler-Illinghworth” approach. Another interesting aspect to be observed is that from 1987 to 2017, the CVA delivered the lowest accuracy among all the examined periods. This performance reduction comes from the use of images acquired by different sensors (i.e., TM and OLI).

Once the CVA has been numerically attested as the most effective approach to cope with the study area, we perform visual inspections by taking the change/non-change maps as obtained by such method. Figure 4 presents the CVA results for all the evaluation periods. One can observe for each pair-wise comparison that the percentage of non-change is dominant. Also, notice that the changes are represented by small regions that are spread over all the study area, especially along the southeastern-northwest track. This portion contains the main highway that crosses the study area, which has led to a dense occupation, urban expansion and industrial activities. In fact, as pointed out by Andrade et al. [40], it was observed a great urban expansion between 1995 and 2015, when urbanized areas reached an increase of 133%. According to Silva et al. [8], demographic and market shifts have resulted in rural land abandonment over the last years and some lands transitioned to industrial production in the Paraíba do Sul River Valley, especially in areas near to cellulose and paper industries. Finally, Silva et al. [41] reinforce that urbanization, international market demands, industrialization and agricultural modernization are the main factors related to landscape changes in Atlantic Forest over the last decades.

Additionally, historical events reveled that this spatial behavior comes from the forest transition stage over the years, as it comprises the period from the permanent reduction of native vegetation to the gradual revitalization of the forest. Such an environment transformation is mainly related to the abandonment of deep slope areas that were not compatible with mechanized agriculture, as well as local demographic and economic issues [8]. Another critical aspect is related to the increase in eucalyptus crop areas, as a consequence of the cellulose industries hosted in the region [7].

Fig. 4
figure 4

Change and non-change maps generated by the combination CVA + Kittler-Illinghworth algorithm for the periods of 1987–1997, 1997–2007, 2007–2017 and 1987–2017

Full size image

The scores in Fig. 5 brings a comparison – in terms of area percentage – between change and non-change regions, as well as the parcel compromised by cloud occurrence. In general, the change/non-change percentages are similar in all periods. If we disregard the cloud locations, it is revealed that on average a \(92\%\) of study area does not change from 1987 to 2017.

Fig. 5
figure 5

Change and non-change percentage occurrences for each period

Full size image

Although the findings about the change percentage in 1987-2017, such analysis does not take into account the land cover iterations occurred in particular years of this period. In order to address this issue, a step-by-step multi-temporal analysis can lead to a better understanding of the study area dynamics. In order to do so, we first compute the map of multi-temporal land cover transitions, which is depicted in Fig. 6, and next we quantify these transition occurrences, as listed in Table 3.

Table 3 Multi-temporal transitions summary and color legend key
Full size table
Fig. 6
figure 6

Multi-temporal land cover transitions, in terms of a change/non-change sequence, over the period from 1987 to 2017

Full size image

By grouping the eight transitions in Fig. 6 (and Table 3) into four general cases, which comprises cloud cover, non-changes, changes in one/two periods, and changes in three periods, one may verify that \(73.42\%\) of the clusters refers to non-change areas. Figure 7 shows the frequency distribution of the established classes.

Furthermore, if cloud-covered areas are unconsidered, the non-change percentage drops to \(66.68\%\). This value represents a difference of \(25.32\%\) when the dynamics of the changes are examined in a pair-wise approach, thus revealing that the Paraíba do Sul River Valley has faced intermittent changes during the last decades.

Fig. 7
figure 7

General cases of multi-temporal transitions

Full size image

5 Conclusions

Unsupervised learning for change detection is capable of providing significant visual insights of the spatial-temporal dynamics between multi-temporal images. Moreover, these mappings can be handled and fused to obtain intuitive, multi-temporal representations of changes in the target region.

By taking such a remote sensing apparatus, we tracked the spatial-temporal changes in the Paraíba Valley along four decades, a southwestern Brazilian region which houses the Atlantic Forest biome. The changes were captured by applying different unsupervised change detection approaches, in an effort to find out the best strategy to identify disturbed areas in Atlantic Forest. We verified that the classic and well-established Change Vector Analysis method, when equipped with Kittler-Illinghworth thresholding approach, produced the best results in comparison with other analyzed methods.

The obtained results have also reveled a regular land-cover change rate about 8% under similar periods, i.e., 1987–1997, 1997–2007, 2007–2017 and 1987–2017. However, after the multi-temporal analysis with the integrated periods and disregarding cloud-covered areas, it was observed an increase of 33% w.r.t. changes over the three analyzed decades.

As future work, we plan to: (i) investigate the use of other thresholding strategies on the CVA algorithm, similar to the analysis conducted in [42], as we have found from our battery of tests that CVA provides accurate results when applied on the analyzed areas so that the thresholding value seems to play an important role in the the generated outputs; and (ii) decrease the temporal scale to five or less years, in an attempt to detect particular changes that may not have been captured when using multi-temporal dataset of ten-years steps, as properly discussed in the survery [43].