Surface albedo as a proxy for landcover clearing in seasonally dry 1 forests : Evidence from the Brazilian Caatinga

Surface albedo as a proxy for land-cover clearing in seasonally dry 1 forests: Evidence from the Brazilian Caatinga 2 3 [Updated version: 11 Mar 2019] 4 5 John Cunha , Rodolfo L. B. Nóbrega , Iana Rufino, Stefan Erasmi, Carlos 6 Galvao Fernanda Valente 7 8 aFederal University of Campina Grande, Center for Natural Resources and Technology, Campina 9 Grande, Brazil; 10 bUniversity of Reading, School of Archaeology, Geography and Environmental Science, Reading, 11 United Kingdom; 12 cImperial College London, Faculty of Natural Sciences, Department of Life Sciences, Ascot, United 13 Kingdom; 14 dUniversity of Gottingen, Institute of Geography, Cartography GIS & Remote Sensing Section, 15 Goettingen, Germany; 16 eGriffith University, Cities Research Institute, Nathan Campus, Queensland 4111, Australia; 17 fUniversity of Lisbon, School of Agriculture, Forest Research Centre (CEF), Tapada da Ajuda, 134918 017 Lisbon, Portugal. 19 *Corresponding author: john.brito@ufcg.edu.br 20 21


46
The identification of land-cover alteration driven by human action is one of the alone. The TSS-RESTREND method can be divided into two main components, one 120 for a structural change (breakpoint) detection and the other for an overall trend 121 estimation. While the first one is feasible to detect changes that occur abruptly, such 122 as LCC, the latter is suitable to identify trends that happen over a longer period of time. 123 In our study, we focus on the use of the structural change detection component years) remote sensing time series use vegetation indices at low spatial resolution, i.e., 135 1 to 8 km (Leroux et al., 2017). 136 Our hypothesis is that SA is a better indicator for LCC detection in seasonally 137 dry forests, such as the Caatinga, compared to VIS-NIR vegetation indices, here 138 represented by EVI and NDVI. Although SA is known to exhibit different responses 139 between vegetated and bare soil surfaces, its use to identify or monitor LCC in dry 140 forests has been poorly documented. We ascribe this scientific gap to the lack of global  The study area is located in the Brazilian Caatinga, a seasonally tropical dry 150 forest that lies in Northeastern Brazil (Fig. 1A (Fig. 2). SPEI is a drought index based on the difference between precipitation 196 and evapotranspiration that is often used to detect and monitor drought periods.    (2) 222 where and b are the surface bidirectional reflectance values and their corresponding 225 conversion coefficients for the six non-thermal Landsat bands, i.e., blue, green, red, 226 NIR and the two shortwave infrared bands (SWIR1 and SWIR2). Table 1    The precipitation data used in this work were obtained from the Climate Hazards    The validation dataset used in this work was built using a two-step procedure.

348
Our analyses show that the two main differences between SA, EVI and NDVI 349 are the range of values (Fig. 4), and the average number of breakpoints detected by is greater by using EVI and NDVI than SA (Fig. 4). Most of the breakpoints occurred 354 during a drought period (SPEI < -1, Fig. 2), especially for EVI and NDVI. to that of EVI and NDVI (Fig. 5,  2012, which is shown by nine patches (Fig. 6). Each patch is identified by the actual

377
LCC year that is dominant among its pixels. The analysis of these patches revealed 378 that when EVI and NDVI were used, a substantial number of pixels (sometimes > 40%) 379 were categorized as false negative (Fig. 7A). This situation was particularly relevant in  The best performance of EVI and NDVI was observed for the patches where the    Table 2, and Figs. 6 and 7A), and the matching between actual 436 and detected LCC years were less than 25%, which is far from an acceptable standard was 50% and 43%, respectively. In comparison, SA's accuracy was 79% (Fig. 5). We 451 interpret this as an effect of adverse climate period or degradation on the vegetation.

452
Although the NDVI and EVI might show ability to detect intra-annual phenological bare soil SA response, which in turn will cause a "time wrong" score for the detected 503 LCC year that is after the actual one. Another aspect to consider is that some LCC  Fig. S6). In addition, the conclusions that 548 are drawn from statistical significance tests (e.g., the Chow test) based on a small 549 sample can be unreliable because the null hypothesis that corresponds to a non-550 28 significant breakpoint will hardly be rejected at the standard significance levels.

551
Therefore, the use of long time series is essential to reduce this uncertainty. In our 552 study, most of the LCC occurred in the 1990s after the first five years of the time series.

553
The Landsat dataset was a valuable source of information by providing long time series 554 where these LCC processes could be evaluated with a low impact from these edge 555 effects.

556
The limitations in LCC detection for the indices that use only the VIS-NIR  Table S1 and Figs. S1 to S5). The use of 567 sensors with higher temporal resolution that aggregate rather than gap fill information 568 is an alternative approach that can be used to assess the performance of EVI and 569 NDVI to detect LCC.