Engenharia e extração de recursos em séries temporais de índice de vegetação para mapeamento de área com culturas anuais
| Ano de defesa: | 2023 |
|---|---|
| Autor(a) principal: |
Almeida, Luiz
 |
| Orientador(a): | |
| Banca de defesa: | , , |
| Tipo de documento: | Dissertação |
| Tipo de acesso: | Acesso aberto |
| Idioma: | por |
| Instituição de defesa: |
Universidade Estadual do Oeste do Paraná
Cascavel |
| Programa de Pós-Graduação: |
Programa de Pós-Graduação em Engenharia Agrícola
|
| Departamento: |
Centro de Ciências Exatas e Tecnológicas
|
| País: |
Brasil
|
| Palavras-chave em Português: | |
| Palavras-chave em Inglês: | |
| Área do conhecimento CNPq: | |
| Link de acesso: | https://tede.unioeste.br/handle/tede/6656 |
Resumo: | In Brazil, financial credit operations, costing and agricultural insurance are the main policies to encourage the development of agriculture. According to the Manual de Crédito Rural [Rural Credit Manual] (MCR), the granting financial institution is responsible for monitoring and inspecting rural credit operations, authorizing the use of geoprocessing and remote sensing techniques for this purpose. The measurement and validation of the farmable area of the property is one of the information necessary for compliance with the MCR. This demand can be met with the use of data mining techniques in time series vegetation index (STIV) for the classification of annual crops. With this panorama in mind, the objective of this research was to map and estimate areas with annual agricultural crops using data mining and engineering techniques and resource extraction in time series vegetation index. The study area comprises the western region of the state of Bahia, Northeastern Brazil, due to the availability of the LEM+ Dataset with land use and land cover classes for the region. The general methodology is based on the process known as Knowledge Discovery in Databases (KDD). The KDD process contains 5 steps, as follows: 1st stage - data selection with 48 vegetation index images from the MODIS sensor of the Terra and Aqua satellites, between the period of 09/22/2019 and 09/21/2020; 2nd stage - pre-processing: division of STIV values by 10,000 and stacking in STIV temporal cube; 3rd stage - transformation: generation of 2 groups of images, one with STIV smoothing which utilized the Savitzky-Golay filter (SG), and the other with STIV simplification by extracting its trend component (TD) with the Seasonal-Trend Decomposition Procedure Based on Loess algorithm (STL), applying feature engineering and feature extraction techniques in order to build 25 images of derived attributes (AD) in each treatment; 4th stage - data mining: 73 images (STIV + AD) of each SG and TD treatment were employed, and 10 combinations of attributes were created containing STIV, basic ADs and polar ADs. The Boruta algorithm was used to select attributes with greater importance for the annual crops classification task and the Random Forest (RF) classifier was optimized with Grid Search Cross Validation, with the purpose of finding the best classification model; 5 th stage - evaluation and interpretation, statistics were extracted from the 10 combinations of attributes to measure Accuracy, Kappa and Mapping Precision. The most promising results were identified in combination 8, which had 73 attributes, namely the STIV generated with the trend component, 15 AD extracted with basic metrics and 10 AD generated with polar metrics, in which, with the Boruta algorithm, only 60 attributes were selected. The mask built with RF mapped 2.53 million hectares with annual crops in the study region. With an overestimation of 12.5% in relation to official Brazilian Institute of Geography and Statistics (IBGE) data, the mask obtained an Accuracy of 92%, Kappa of 86% and Precision of approximately 92.2%. From the Cadastro Ambiental Rural [Rural Environmental Registry] (CAR), a rural property was selected, and a simulation of an agricultural credit inspection process was carried out. The area of consolidated use of the property, which had 852.82 hectares, was compared with cases of mapping the area with annual crops. One of the cases obtained a mapped area of 848.145 hectares with an Accuracy of 99.22%, Kappa of 98.32% and an approximate Precision of 100%. This mapping case was performed by applying the RF model trained on STIV and AD images cropped for said rural property. In conclusion, the techniques applied in mapping areas with annual crops using data mining, feature engineering and feature extraction in STIV allow agility in territorial management processes and for the validation of information on rural properties in agricultural credit carried out by financial institutions. |