Cropland expansion in Ecuador between 2000 and 2016.

We describe changes in the cropland distribution for physiographic and bioregions of continental Ecuador between 2000 and 2016 using Landsat satellite data and government statistics. The cloudy conditions in Ecuador are a major constraint to satellite data analysis. We developed a two-stage cloud filtering algorithm to create cloud-free multi-temporal Landsat composites that were used in a Random Forest model to identify cropland. The overall accuracy of the model was 78% for the Coast region, 86% for the Andes, and 98% for the Amazon region. Cropland density was highest in the coastal lowlands and in the Andes between 2500 and 4400 m. During this period, cropland expansion was most pronounced in the Páramo, Chocó Tropical Rainforests, and Western Montane bioregions. There was no cropland expansion detected in the Eastern Foothill forests bioregion. The satellite data analysis further showed a small contraction of cropland (4%) in the Coast physiographic region, and cropland expansion in the Andes region (15%), especially above 3500m, and in the Amazon region (57%) between 2000 and 2016. The government data showed a similar contraction for the Coast (7%) but, in contrast with the satellite data, they showed a large agricultural contraction in the Andes (39%) and Amazon (50%). While the satellite data may be better at estimating relative change (trends), the government data may provide more accurate absolute numbers in some regions, especially the Amazon because separating pasture and tree crops from forest with satellite data is challenging. These discrepancies illustrate the need for careful evaluation and comparison of data from different sources when analyzing land use change.

Biases and limitations of Global Forest Change and author-generated land cover maps in detecting deforestation in the Amazon.

Studying land use change in protected areas (PAs) located in tropical forests is a major conservation priority due to high conservation value (e.g., species richness and carbon storage) here, coupled with generally high deforestation rates. Land use change researchers use a variety of land cover products to track deforestation trends, including maps they produce themselves and readily available products, such as the Global Forest Change (GFC) dataset. However, all land cover maps should be critically assessed for limitations and biases to accurately communicate and interpret results. In this study, we assess deforestation in PA complexes located in agricultural frontiers in the Amazon Basin. We studied three specific sites: Amboró and Carrasco National Parks in Bolivia, Jamanxim National Forest in Brazil, and Tambopata National Reserve and Bahuaja-Sonene National Park in Peru. Within and in 20km buffer areas around each complex, we generated land cover maps using composites of Landsat imagery and supervised classification, and compared deforestation trends to data from the GFC dataset. We then performed a dissimilarity analysis to explore the discrepancies between the two remote sensing products. Both the GFC and our supervised classification showed that deforestation rates were higher in the 20km buffer than inside the PAs and that Jamanxim National Forest had the highest deforestation rate of the PAs we studied. However, GFC maps showed consistently higher rates of deforestation than our maps. Through a dissimilarity analysis, we found that many of the inconsistencies between these datasets arise from different treatment of mixed pixels or different parameters in map creation (for example, GFC does not detect reforestation after 2012). We found that our maps underestimated deforestation while GFC overestimated deforestation, and that true deforestation rates likely fall between our two estimates. We encourage users to consider limitations and biases when using or interpreting our maps, which we make publicly available, and GFC's maps.