Using remote sensing and machine learning to evaluate the interaction between agricultural expansion and the environment: a study of the Brazilian Cerrado

Abstract

Global population growth has increased the demand for food, and many countries have answered this problem by expanding agricultural lands. Brazil stands out as one of the world’s fastest growing agricultural development zones, especially in the Brazilian savanna, which has been transformed into an important world breadbasket. Meanwhile, the region is also one of the world’s biodiversity hotspots. Continuous agricultural expansion including the new agricultural frontier (Matopiba region, which is in the northern part of the Cerrado) has affected the natural environment and ecosystems in the region. Although many studies have used different methods to estimate the interaction between agricultural expansion and the environment, the performance of combining remote sensing and machine learning is still unclear. The main goal of this dissertation is to examine the interaction between agricultural expansion and the environment using remote sensing and machine learning from aspects of pollinator, crops, vulnerability, and fire activity. In the following chapters, the interaction between agricultural expansion and the environment will be investigated using a combination of model approaches, remote sensing, GIScience, machine learning, deep learning, and data mining. Chapter 2 presents a spatial distribution of selected bee species richness and soybean production at a regional scale. The findings indicate that higher bee species richness and higher soybean production are in the southern Cerrado, and the environment has a stronger impact on bee species richness than soybean production. Additionally, the analysis of the interaction of bee species richness and soybean production reveals that their relationship is not a linear one. Chapter 3 develops an indicator system to estimate environmental vulnerability in the entire Cerrado. The main finding is that areas of high environmental vulnerability are in the southern Cerrado. Additionally, mined historical Twitter results reveal that social media data is a promising data set for environmental vulnerability assessment. Chapter 4 creates a novel deep learning model (Conv-LSTM) to classify two agricultural expansion sites in the Matopiba region over time and estimates the correlation between land use types and burned areas in September (the last month of the dry season) using classification results and the MODIS products. The findings determine that the proposed model can classify different land structure areas at coarse spatial resolution. Additionally, the overlay analysis with burned areas indicates that fire activities easily occurred in the grasslands in Site A and the forestlands in Site B. The results also claim that fire activities more readily occurred at the edge of cropland areas, which suggest that fire activities are still a common way to expand agriculture in this region.