Monitoring São Paulo State restoration forests: application of remote sensing and deep learning to classify successional stages and subsidize public policies

• Grant number: 22/14605-0
• Support Opportunities: Scholarships abroad - Research Internship - Post-doctor
• Start date: May 01, 2023
• End date: April 30, 2024
• Field of knowledge: Agronomical Sciences - Forestry Resources and Forestry Engineering - Nature Conservation
• Principal Investigator: Pedro Henrique Santin Brancalion
• Grantee: Angelica Faria de Resende
• Supervisor: Thiago Sanna Freire Silva
• Host Institution: Escola Superior de Agricultura Luiz de Queiroz (ESALQ). Universidade de São Paulo (USP). Piracicaba , SP, Brazil
• Institution abroad: University of Stirling, Scotland
• Associated to the scholarship: 19/24049-5 - Monitoring São Paulo State restoration forests: application of new remote sensing tools and subsidies for public policies, BP.PD

Abstract

The Atlantic Forest Biome overlaps the territories of three South American countries, with the majority (92%) falling under the jurisdiction and legislation of Brazil. Only ~22% of native Brazilian Atlantic Forest native cover remains. Since 2006 the Atlantic Forest biome has been protected by federal law 11,428/2006, which implements forest use restrictions based on successional stages, with more advanced stages receiving more protection. However, given its large extension (> 1.1 MI Km2) and scattered pattern of native forest fragments, in-situ checking of forest successional stages by law agents is impracticable. Our goal is thus to understand the spatial distribution of successional stages, using an extensive database of reference forest plots to be classified according to the law, and use this data to provide an upscaled high-resolution view of the spatial configuration of forest successional stages for the entire extent of São Paulo Atlantic Forest. We will start with the classification (according to law parameters) of a large dataset of 350 forest plots from 30 landscapes in São Paulo, collected within the NewFor Project (FAPESP #2018/18416-2). Then, in the next two steps, we will test, train, and validate deep learning (Convolutional Neural Networks and Deep Morphological Networks) to spatially upscale the classification of the successional stages. First, using data from a drone-borne lidar sensor acquired for 14 of these landscapes, we will extrapolate plot-level stages classification to the landscape level; with this, we can extend sampling areas. Then a final upscale to São Paulo Atlantic Forest level spatialization, we will be using plots (for landscapes without lidar data) and landscapes to classify the whole area using radar satellite images (Sentinel 1, ALOS 2 PALSAR mosaics) and optical images (Landsat and Sentinel 2, Planet NICFI). As the main result, we will determine the spatial configuration of the secondary forest types and successional stages. With our results, we intend to support monitoring secondary vegetation successional stages and decision-making by the environmental agencies. (AU)