Multi-Decadal Annual Land Cover Dynamics and Forest Disturbance in the Brazilian Amazon Biome

Abstract

The Brazilian Amazon biome is undergoing a fast land cover change in the past 50 years. Most of these changes are associated with deforestation and forest disturbance caused by selective logging and fires, and recently to extreme climate change events. Deforestation monitoring from the INPE (National Institute for Space Research), based on Earth Observation (EO) sensors, is in place since the late 1980s. After 2000, new deforestation monitoring systems emerged driven by lower computational costs and free EO data, notably Landsat. The advent of the Google Earth Engine platform has broadened monitoring initiatives in the Amazon, allowing monitoring to expand beyond forest clearing by deforestation. Here, we present our efforts to develop and the results of annual multi-decadal land use and land cover (LULC) change for the Brazilian Amazon biome between 1985 to 2019. We processed 74,000 Landsat scenes available for this period with Cloud Cover less or equal to 50%, covering 201 path-rows. We then trained and validated a random forest classifier (RFC) using 35,000 independent random samples (10,000 for training and calibration and 25,000 accuracy assessment) for the entire Amazon biome generated by the LAPIGUFG research team. The input features were selected with the random forest package available in R Language because Google Earth Engine does not have specialized statistical libraries. The final feature space ended up with eight variables derived from a Spectral Mixture Analysis (SMA) model, including Green Vegetation (GV), Non-Photosynthetic Vegetation (NPV), Soil, Cloud, Green Vegetation Shade (GVS), Normalized Difference Fraction Index (NDFI), Shade and Canopy Shade Fraction (CSFI). The LULC classes included: Forest, Savanna, Grassland, Pasture, Agriculture, Water, and Non-Vegetated Area. Annual LULC maps were produced based on intra-annual post-classification rules applied to all available scenes in a given year. The SMA model fractional outputs were also used to map surface water and forest disturbances based on empirically defined thresholds, already tested and published in the scientific literature. Annual maps of surface water and forest disturbances were combined with the LULC yearly maps. Additionally, we estimated the extent of secondary vegetation (SV) based on pasture and agricultural transitions to forests. As the final step, we integrated annual LULC maps with forest disturbance, SV, and surface water maps to produce the most comprehensive land change information about the Amazon biome since 1985. The LULC final maps and intermediate map products are available in the MapBiomas platform, allowing multiple scientific and societal applications. As further steps, we will investigate our LULC maps' applications to improve deforestation and climate change risks and carbon emission modeling for the Amazon biome.