EVALUATING THE SEPARABILITY BETWEEN DRY TROPICAL FORESTS AND SAVANNA WOODLANDS IN THE BRAZILIAN SAVANNA USING LANDSAT DENSE IMAGE TIME SERIES AND ARTIFICIAL INTELLIGENCE

The Brazilian Savanna is the second largest biogeographical region in Brazil and present different vegetation types, consisting mostly of tropical savannas, grasslands, and forests. The forest types have different tree cover and floristic composition, which is associated to leaf deciduousness. Considering the importance of Cerrado to biodiversity conservation and the maintaining of environmental services, the development of methods to map the different forest types in Cerrado is important for conservation programmes, subsidize restauration plains, and to allow estimations of carbon sink and stock. Mapping heterogeneous tropical areas, such as the Brazilian Savanna, is very complex due to the natural factors and peculiarities of the vegetation types, and it's still particularly challenging to separate between different forest formations. In this study we tested machine learning approaches based on the use of dense image time series, in order to evaluate the separability Dry Tropical Forests and Savanna woodlands. We considered the Brazilian State of Tocantins as the study area, which is located in the Northern region of the country. RF classification of Landsat dense time series showed an overall accuracy of 0.85005, while the LSTM approach presented an overall accuracy of 0.88601, with the highest f1-score for the savanna woodlands class, suggesting the capability of the recurrent neural networks on handling complex long-term dependencies such as the EVI dense time series data. This study showed the potential for the development of a semiautomatic method for discriminating the different types of forest formations in the Brazilian Savanna, based on remote sensing.


INTRODUCTION
Considered to be one of the richest and diverse savannas in the world, the Brazilian Savanna, also known as Cerrado, is the second largest biogeographical region in Brazil. It has different vegetation types consisting mostly of tropical savannas, grasslands and forests (Ribeiro and Walter, 2008), and it covers approximately 24% of the country (MMA, 2021). Furthermore, the Brazilian Savanna is considered a global biodiversity hotspot for conservation (Myers et al. 2000), and houses many endemic and threatened species (Colli et al. 2020). The Cerrado provide important environmental services such as climate regulation and water supply to different regions in Brazil (Oliveira et al. 2015;Strassburg et al. 2017). It also contributes to 43% of surface water outside the Amazon in Brazil (Strassburg et al. 2017).
Despite that, the Cerrado has lost around 88 Mha (46%) of its native vegetation with a projection that 31-34% of the remaining biome is likely to be cleared by 2050 (Strassburg et al., 2017). While much attention has been placed on Amazon, the rate of conversion of native Cerrado vegetation can be up to 2.5 times the conversion observed in the Amazon (Rocha et al., 2011;Strassburg et al. 2017).
Most of the native vegetation conversion tends to occur in areas with dense vegetation that have favourable climate and soil conditions and in flat terrains that are suitable for mechanized 1 More information in: http://fip.mma.gov.br/projeto-fm * Corresponding author: hnbendini@gmail.com farming (Alencar et al., 2020). Therefore, the conversion of natural vegetation into agricultural land (e.g., soybean) and pasture is leading to major carbon emissions (Noojipady et al., 2017) and biodiversity loss (Ratter et al., 1997) through the forest clear-cutting, stressing the importance of frequent mapping approaches that enable monitoring and assessing ongoing change processes.
The Brazil Investment Plan (BIP) under the Forest Investment Program (FIP) seeks to promote sustainable land use and forest management improvement in the Cerrado Biome in order to reduce pressure on remaining forests, reduce greenhouse gas (GHG) emissions and increase carbon dioxide sequestration (Tuchschneider, 2013). As part of BIP, the project "Development of systems to prevent forest fires and monitor vegetation cover in the Brazilian Cerrado" aims to improve Brazil's capacity to monitor deforestation, prevent the risk of forest fires and improve models for estimating greenhouse gas (GHG) emissions, making tools and data available to environmental agencies 1 . The project will provide the basis for improving the management of water, forest and soil resources in the Brazilian Cerrado, which, together with other projects financed by the FIP in Brazil, should promote the sustainable management of forests. In the context of this project, one of the activities it to modify the existing land cover classification system for the Cerrado developed by IBGE (Brazilian Institute of Geography and Statistics) on the basis of the Food and Agriculture Organization of the United Nations (FAO) Land Cover Classification System framework. Therefore, it will be possible to discriminate forest from non-forest vegetation taking into account the spectrum of structural vegetation complexity in the Cerrado. However, mapping heterogeneous tropical areas, such as the Cerrado, is challenging due to the natural, climatic and topographic factors and the peculiarities of the characteristic physiognomies .
The woody-dominated stratum of Cerrado can be divided in riparian forests, dry tropical forests and savanna woodlands. These forest types have different tree cover and floristic composition, which is associated to leaf deciduousness. The loss of leaf area corresponds to a physiological response to water shortages in dry seasons (Zalamea and González, 2008).
Dry Tropical Forests are characterised by different levels of deciduous trees, depending on climatic conditions and mainly on the depth of the soil (Ribeiro and Walter, 2008). In the Cerrado, the Dry Tropical Forests can be divided according to the tree cover in different seasons. Both Semideciduous Forest and Deciduous Forest have a tree cover ranging from 70 to 95% in the wet season, however the tree cover of Semideciduous Forest ranges from 50 to 60% in the dry season, while the tree cover of Deciduous Forest ranges from 30 to 50%. Furthermore, the canopy height of Semideciduous Forest is about 16 meters, while in Deciduous Forest is about 10 meters (Ribeiro and Walter 2008). In addition, Dry Tropical Forests were considered one of the most threatened tropical ecosystems (Miles et al. 2006). The savanna woodlands are forests with similarities to savannas due to species composition. The tree cover can range from 50 to 90%. Although it may contain evergreen plants, many species show deciduousness (Haidar, 2017).
For these reasons, it's still a challenge to separate between these different forest formations (Ferreira et al., 2004;Sano et al., 2010;Grecchi et al., 2013;Schwieder et al., 2016;Girolamo-Neto et al., 2017;Neves et al., 2021;Bendini et al., 2020). Previous studies highlighted the benefits of dense remote sensing time series, derived land surface phenological metrics (here, also called as phenometrics) and analysed their relationship to the grassland-savanna-forest gradient of the Cerrado (Schwieder et al., 2016;Bendini et al., 2019b). They tried to separate between the different vegetation types using traditional machine learning approaches such as the Support Vector Machine (SVM) and Random Forests (RF), but accurately differentiating between dry forests and savanna woodlands still require more efforts. The Deep Learning approaches, such as the Convolutional Neural Networks (CNN) and Recurrent Neural Networks approaches were already proved to be capable on learning patterns of different land-use and land-covers, for instance on the detection of deforestation on the Brazilian Amazon (Maretto et al., 2020) and Cerrado (Taquary et al., 2021;Matosak et al., 2022), and for vegetation mapping in the Cerrado Bendini et al., 2021b).
Considering the importance of Cerrado to biodiversity conservation and the maintaining of environmental services, the development of methods to map the different forest types inserted in Cerrado is important to create bases for conservation programmes, subsidize restauration plains, and to allow estimations of carbon sink and stock. In this study we tested two different machine learning approaches based on the use of dense image time series, RF and LSTM, in order to evaluate if it's possible to separate Dry Tropical Forests and Savanna woodlands.

Study Area and Reference
We considered the Brazilian State of Tocantins as the study area, which is located in the Northern region of the country and has a large area (277,423.63 Km² according to IBGE 2020), harbouring different ecosystems types. This state is covered by two biogeographic regions, the Cerrado (predominance of savannas) and the Amazon Forest (predominance of forests).
According to Sano et al (2009), which divided the Cerrado into 19 ecoregions that are unique in terms of landscape characteristics, the Cerrado in Tocantins is composed by eight ecoregions: Alto Parnaíba, Araguaia Tocantins, Bananal, Bico do Papagaio, Chapadão do São Francisco, Parnaguá, Planalto Central e Vão do Paranã.
The State of Tocantins is situated in a climate transition area, and the climatic classification of Thornthwaite applied to this State pointed out the presence of three types: i) C1A'w2a' (dry subhumid, megathermal, with large summer water surplus, and a temperature efficiency regime normal to megathermal) in the eastern end of the State; ii) C2A'wa' (moist subhumid, megathermal, with moderate winter water deficiency, and a temperature efficiency regime normal to megathermal) in the central range; and iii) B1A'wa' (humid, megathermal, with moderate winter water deficiency, and a temperature efficiency regime normal to megathermal), in the southwest of Tocantins (Souza et al. 2019).
In order to build a relevant dataset for training and validation, we used reference data provided by the Secretariat of Planning and Budget of the Tocantins state (SEPLAN, 2013), which consists on a detailed map in a scale of 1:100,000, produced with Landsat imagery, supported by visual interpretation of high spatial resolution imagery (CBERS 2B), the SRTM (Shuttle Radar Topography Mission) elevation model geology, terrain and soil maps.

Data and Preprocessing
We extracted a set of 40,041 EVI (Enhanced Vegetation Index) Landsat 7 and 8 point-based time series, using the GEE Time Series Explorer QGIS Plugin (Rufin et al., 2021), during the period of August-2014 to October-2015, involving the 3 classes obtained from the reference database, and considering the Ribeiro & Walter (2008) definition. The classes were: 1) Savanna Woodlands, 2) Deciduous and 3) Semideciduous Dry Forests. We made it compatible over time, removing the areas deforested after 2014 with the PRODES Cerrado deforestation mask (INPE, 2019) 2 . The EVI was chosen because it is known to increase sensitivity for biomass estimation through a de-coupling of the canopy background from the signal and a reduction in atmospheric and soil reflectance influence (Huete et al., 2002).
Then, we applied a weighted ensemble of Radial Basis Function (RBF) convolution filters as a kernel smoother to fill data gaps such as cloud cover and Scan Line Corrector (SLC)-off data (Schwieder et al., 2016, Bendini et al., 2019a. A total of 11 phenometrics were derived using TIMESAT (Jönsson, Eklundh, 2004), extracted for the seasonal cycle observed in the EVI time series. Phenometrics included day-of-the-year (DOY) of start, mid, end, length of crop seasons and phenological proxies like peak, base value, seasonal amplitude or rate of increase, decrease (Jönsson & Eklundh, 2004). and the polar features, which are based on the representation of the time series by projecting the values onto angles in the interval [0,2π] (Körting et al., 2013).

Random Forest and LSTM classification
The classification was done using the RF algorithm (Breiman, 2001), with different sets of metrics (RBF fits, phenometrics, polar metrics and the combination of them). We randomly selected 70% of the samples to train and 30% for validation. RF is a non-parametric machine learning algorithm that is based on decision trees. As individual decision trees are prone to errors, RF uses an ensemble of many decision trees that were independently trained with random subsets of the input data to overcome this limitation (Breiman, 2001). The RF classifier works based on creating decision trees that are used to predict over data. The class with the majority of votes among all trees is chosen as the final prediction result. RF needs two parameters to be tuned including the number of trees to grow (ntree), and the number of variables randomly sampled as candidates at each split (mtry). Figure 2 shows a generalisation example for how a classification is made using the RF algorithm. The results were evaluated by the confusion matrix, which was used to derive the overall accuracy (Chinchor & Sundheim, 1993) and the class f1scores (Shapiro, 1999). The "randomForest" package in R was used for our classification tasks (R Core Development Team, 2019).  The algorithm implementation in R further allows to assess the variable importance of each input variable based on the Gini coefficient (Liaw et al., 2002).
We also tested recurrent neural network models with Long-Short Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) layers, which is a Recurrent Neural Network (RNN). The LSTM was proposed in order to solve problems related with loss of longterm dependency and gradient vanishing present in previous RNN architectures. The LSTM cell is composed of a cell state and input, output, and forget gates, as it is shown in Figure 3. For each input element, ht is computed with Equations 1 to 6: The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2022 XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France ̃= ℎ (̃+̃ℎ −1 +̃) where ℎ : is the hidden state at time t; : is the cell state at time t; : is the input at time t; ℎ −1 : is the hidden state at time t-1 or the initial hidden state; : is the input gate; : is the forget gate; ̃: is the cell gate; : is the output gate; : is the sigmoid function.
The same samples used in the RF classifier were also used with the LSTM. In this case, five distinct sequential models were used.
Regarding their composition, model number 5 had 5 LSTM layers, with 96, 48, 24, 12, and 6 units, respectively. Each LSTM was followed by a batch normalisation layer, and at the end there was a fully connected layer (Dense) with 3 units and SoftMax activation function. Model number 4 was similar to model number 5, but without the first LSTM and batch normalisation layers, and so on until model 1, which had only one LSTM layer with 6 units, the batch normalisation and fully connected layers. Table 1 shows details about the layers of each model. The models were trained with the Categorical Cross Entropy loss function, Adam optimizer with a learning rate of 0.00001, batch size of 256 samples, and 1000 epochs.

RESULTS AND DISCUSSIONS
Our best RF model was built with the dataset composed by RBF fits, in which we empirically determined the parameters mtry of 5 and ntrees of 500, and reached an average overall accuracy of 0.85005, while the F1-Scores for the classes Savanna Woodlands, Deciduous Dry Forest, and Semideciduous Dry Forest were 0.68261, 0.70348, and 0.90273, respectively. Figure  4 shows the confusion matrices.  Table 2 present the variable importance analysis results based on the Mean Decrease Gini (MDG). The mean decrease in Gini coefficient is a measure of how each variable contributes to the homogeneity of the nodes and leaves in the resulting RF. The higher the value of mean decrease Gini, the higher the importance of the variable in the model (Han et al., 2016).    We found that the EVI values during the end of the dry season were the most informative variables for separating among the different vegetation classes, suggesting that there are significant differences between their seasonal dynamics during this period ( Figure 6). This result suggests that the leaf cover is lowest in the end of the dry season. Therefore, it is the best period to acquire satellite images to separate the different forest types of Cerrado.

Variable Mean Decrease Gini
Haidar (2017) analysed the leaf phenological cycles of the remnants of dry forests, using vegetation indexes time series and phenological metrics. The author determined four phenological groups with trends that vary from deciduous to evergreen canopy species, as well as meaningful differences in the cycles of the remnants. The existence of these different phenological groups can explain the high variance within the semideciduous forests class, which may present a gradient varying between deciduous and evergreen forests. This high variance led to confusions between the semideciduous forests and savanna woodlands, in which pixels of savanna woodlands were misclassified as semideciduous forests. The variation of the phenology is related to the differences in the soil physical and chemical properties and the average temperature, which can vary across the Tocantins state. Bendini et al. (2020) used Landsat Analysis Ready Data (ARD) in combination with different environmental data for a modelling exercise for classifying the vegetation in the Cerrado in two different hierarchical levels. In the second level, the authors also reported the complexity of classifying in details the forest classes. Bendini et al. (2021a)

Final Considerations
In this work, we tested two different machine learning approaches, RF and LSTM, based on the use of dense image time series, in order to evaluate if it's possible to separate Dry Tropical Forests and Savanna woodlands. LSTM approach presented a higher accuracy for the savanna woodlands class, in comparison to the RF approach, reaching the highest overall accuracy. In general, it was possible to separate between the different forest classes. Although the results are motivating, there´s still the potential for improving the accuracy with other deep learning approaches, such as the convolutional Long-Short-Term-Memory (ConvLSTM), and also, testing different spectral bands, other spectral indices and data from Synthetic Aperture Radar (SAR), such as the Sentinel 1. The combination of these image time series with auxiliary data, such as terrain and soil type, can also add value to this classification task. This study showed the potential for the development of a semi-automatic method for discriminating the different types of forest formations in the Brazilian Savanna, based on remote sensing.