Estudo de áreas úmidas de campos de murundus sob impacto de projeto de irrigação no noroeste goiano

Abstract

This research investigated Mound Field Wetlands next to an area where the Projeto de Irrigação Luiz Alves do Araguia (PILAA) irrigation project was installed using remote sensing and Random Forest algorithms. The region of interest of the study was the Parque Estadual do Araguaia and stages 2 and 3 of the PILAA, as representative areas of different degrees of anthropization. The first step on the research was making a bibliometric review about mound fields on Web of Science, Scopus and Biblioteca Digital Brasileira de Teses e Dissertações (BDTD) databases. This stage aimed to understand the current state of general research about these wetlands in Brazil, as well as identify gaps in knowledge to be addressed. During this stage 97 scientific papers and 21 post graduate research projects were analysed. The paper publications were scarce in the first decades, however there was a significant increase in publication numbers from 2010 onward. A prevalence of the Biological Sciences was identified on the examined papers, with a notable sub representative number of them being connected to the Earth Sciences. Most of the inspected works also had some relation with the Cerrado, since the biome’s name was used on many of the papers as keywords, also most of the postgraduate research was developed in institutions of the Cerrado. The second stage of this study investigated mound fields next to PILAA using Remote Sensing and Random Forest algorithms, aiming to identify the most relevant spectral indexes and classify the region of interest according to the anthropization level. 200 sample points were collected for the classification, divided in 4 classes: 1- preserved areas; 2- areas in regeneration; 3- degraded areas; 4- water. The variables for the classification were collected from Landsat 7 and 8, Sentinel 1 and 2 and ALOS, totalling 72 spectral and topographic variables , including spectral indexes generated from the passive sensors. The filtering of this data collection was done in Rstudio using Recursive Feature Elimination (RFE), reducing the dataset to the 5 most relevant variables. The validation for the model was done using confusion matrixes, global accuracies, Kappa index and F1-score. The RFE model gave the highest scores to the Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Pond Index (NDPI), Modified Normalized Difference Water Index (MNDWI) and Normalized Difference Turbidity Index (NDTI), emphasizing the importance of the vegetation’s and flooded areas’ spectral response on the study of these environments. The final classification model achieved a global accuracy of 84,06% (train) and 80,59% (test), with a Kappa of 0.786 and 0.739, respectively. The classification showed that most of the study area was covered by preserved areas (41,36%) and areas in regeneration (39,05%), focused in the PEA and stage 3 of PILAA. The anthropized areas (12,45%) were mostly present on the active parts of the irrigation complex and along the roads and artificial drainages. The biggest confusion made by the algorithm was between classes 1 and 2, which shows a spectral similarity among them. We concluded that Mound Field Wetlands show some resilience facing the advance of anthropic activities because of the spectral similarity found, which showed the possibility of regeneration after more moderate anthropic interventions. The good scores on the validation metrics show the possibility of generalization.