Using Spatial Features to Reduce the Impact of Seasonality for Detecting Tropical Forest Changes from Landsat Time Series

Eduarda Silveira, Inácio Bueno, Fausto Acerbi-Junior, José Mello, José Scolforo, Michael Wulder
2018 Remote Sensing  
In forested areas that experience strong seasonality and are undergoing rapid land cover conversion (e.g., Brazilian savannas), the accuracy of remote sensing change detection is affected by seasonal changes that are erroneously classified as having changed. To improve the quality and consistency of regionally important forest change maps, we aim to separate process related change (for example, spectral variability due to phenology) from changes related to deforestations or fires. Seasonal
more » ... s are typically used to account for seasonality, but fitting a model is difficult when there are insufficient data points in the time series. In this research, we utilize remotely sensed data and related spectral trends and the spatial context at the object level to evaluate the performance of geostatistical features to reduce the impact of seasonality from the NDVI (Normalized Difference Vegetation Index) of Landsat time series. The study area is the São Romão municipality, totaling 2440 km 2 , and is part of the Brazilian savannas biome. We first create image objects via multiresolution segmentation, basing the objects on the characteristics found in the first image (2003) of the 13-year time series. We intersected the objects with the NDVI images in order to extract semivariogram indices, the RVF (Ratio Variance-First lag) and AFM (Area First lag-First Maximum), and spectral information (average and standard deviation of NDVI values) to generate the time series from these features and to derive Spatio-Temporal Metrics (change and trend) to train a Random Forest (RF) algorithm. The NDVI spatial variability, captured by the AFM semivariogram index time series produced the best result, reaching 96.53% of the overall accuracy (OA) to separate no-change from forest change, while the greatest inter-class confusion occurred using the average of the NDVI values time series (OA = 63.72%). The spatial context approach we presented is a novel approach for the detection of forest change events that are subject to seasonality (and possible miss-classification of change) and mitigating the effects of forest phenology without the need for specific de-seasoning models. pressures with the most rapid land conversion in Brazil, exceeding that of the tropical forests [2] . Given the high rates of deforestation and few formally protected areas, the savanna biome is among the most endangered ecoregions on Earth [3] . Accurate mapping and monitoring of areas undergoing conversion are needed to indicate the priority areas for conservation and to inform sustainable land use management, as well as to improve the understanding of the dynamics and related impacts on carbon balance, nutrient cycling, and water resources [1] . At present, much of the effort estimating forest changes in Brazil has been focused on the tropical rainforests with less attention dedicated to the less humid seasonal regions [4] . Many different models can be used to estimate deforestation trends in tropical regions [5] , however, in Brazil, most of these models were developed for the Amazon region with a few studies representing the Savannas biome [4, 6] . Although human-induced disturbances are threatening these ecosystems, systematic investigations of land cover change are lacking [7] . Remotely sensed data have been widely recognized as essential data sources for comprehensive mapping and quantification of land cover changes [8] [9] [10] [11] [12] [13] [14] . Studies have used the NDVI (Normalized Difference Vegetation Index) time series from moderate resolution sensors (that is, MODIS) for land cover change detection benefitting from their frequent revisit time [15, 16] . However, images from medium spatial resolution sources such as Landsat offer more detailed spatial information, providing insights at smaller spatial scales over a longer period of time. Initiated in 1972, Landsat is the longest running cross-calibrated globally consistent record of the Earth's surface at a medium resolution [13] . The ability to utilize the dense time series of imagery from Landsat has been demonstrated for mapping land cover changes, such as deforestation, forest degradation, and impacts of fire [17] [18] [19] [20] [21] . Given the long time series of Landsat data at the global scale [22] , the production of forest change maps is common as first steps towards estimating biomass and related CO 2 emissions [23] and biodiversity loss. Based on investigations using Landsat data, the greatest total forest change due to deforestation from 2000 to 2012 of the four climate domains (tropical, subtropical, temperate, and boreal) is the tropical domain [24] . However, estimating forest change by remote sensing is not a trivial task. Satellite images can contain a combination of natural forest phenology, cloud cover, atmospheric scattering, and geometric errors [25] , impairing the precision of change estimates, especially in tropical areas that exhibit strong seasonality in photosynthetic activity [12] . When images from different seasons are acquired, changes caused by phenological differences are present and can result in the mapping of a change in condition, rather than a change in land cover or land use [26] . To reduce the seasonal variations captured by remotely sensed images, techniques for de-seasoning time series have been used [27] on the assumption that seasonal patterns can be identified and removed [15, 16, 26, 28, 29] . However, this assumption may not always hold if the time series is for satellite images which are not acquired at regular intervals or have gaps due to the presence of clouds [12] . To minimize the impacts of seasonality, methods combining different satellites [30, 31] and others that explore the spatial context [12] have been developed. Previous studies have shown that pixel-based change detection can suffer from not including additional information regarding the spatial context [32] . The inclusion of spatial information can reduce the small or spurious changes [33] and also allow for the incorporation of supplemental spatial information, such as geostatistical features [34] [35] [36] [37] . Although pixel-based approaches also generate spatial information, the neighborhood used to retrieve the information is often defined by a square window that is easy to implement. However, they are computational expensive [27], biased along their diagonals, and can straddle the boundary between two landscape features, especially when a large window size is used [38] . Here, we propose a new approach to reduce the impacts of the seasonal changes introduced by forest phenology on forest change detection using the spatial variability of Landsat NDVI time series at the object level. We hypothesize that variability measures are not affected by vegetation seasonality as the NDVI variability remains constant in the presence of seasonal changes and substantially increases in the presence of land cover changes. Thus, we investigated if the geostatistical features can be used to eliminate seasonal variations in satellite image time series and improve the change detection results. Remote Sens. 2018, 10, 808 3 of 21 We used an object-based approach associated with a set of Spatio-Temporal Metrics by computing geostatistical features, specifically the semivariogram indices developed by Reference [35]. We also compared the change detection results to those provided by traditional spectral features (average and standard deviation of NDVI values). Implementation of this approach enables us to enrich object-based remote sensing methods using the spatial information provided by semivariograms that capture the spatial variability of images which are directly related to land cover changes, improving the change detection classification performance in highly seasonal and ecologically important environments. Study Area The study area is the São Romão municipality in the state of Minas Gerais (MG), Brazil (Figure 1a ). Totaling 2440 km 2 , approximately 84% of this area is occupied by land cover classes comprising Brazilian savannas [39] (Figure 1b) . The Brazilian savanna's natural vegetation consists of mixtures of grasslands, shrublands, and woodlands which are found over a gradient of vegetation density, growth form, vertical structure, and biomass [40] (Figure 1c ). Riparian forests are found along the rivers and streams of the study area. As vegetation progresses from grasslands to woodlands and forests, there is an increase in cover, height, density, and the basal area of the trees present [41] . Our analysis focused on the main vegetation types found in the study area (Table 1) : open grassland, open grassland with sparse shrubs, mixed grassland, shrublands, trees up to seven meters, and riparian vegetation (palm swamps and semideciduous forests) [42, 43] . The climate is typically tropical (Aw) with rains concentrated in the October-May period, which characterizes the high seasonality in the region [44] .
doi:10.3390/rs10060808 fatcat:gqdjku5ecfbpfgbe4s2xwlpmsi