Manejo florestal orbital: uso de imagens de Satélite e machine learning para otimizar e Auxiliar a tomada de decisão

Abstract

Forest monitoring using images is a well-established practice in the forestry industry, utilizing both drones and remote sensing. However, there is a continuous need to enhance technologies and methodologies to achieve more accurate and effective results. Satellite images provide a rich source of information with spatial and temporal resolution, enabling constant monitoring of forests with reliable precision. This thesis aims to use remote sensing data economically and periodically, integrating this information with forest conditions. This includes the automated and periodic detection of forest anomalies, the identification of weed competition with young eucalyptus plantations, and predicting future growth and productivity potential. In the first article, the objective was to develop an autoregressive model using vegetation index information combined with cadastral variables such as climate zones and genetic material to detect anomalies in forest plantations. Machine learning (ML) models, including random forest (RF) and extreme gradient boosting (XGBoost), were employed to achieve this goal, with XGBoost delivering the best results. The proposed methodology proved to be highly effective and reliable for anomaly detection. The second article focused on detecting weed competition with young eucalyptus plantations using Sentinel-2 images. A methodology for variable selection using a genetic algorithm and classification of the presence or absence of competition with RF, known as GARF, was implemented. This method efficiently selected an optimal subset of variables, enabling the identification of weed competition using satellite images. The third article aimed to estimate the dominant height of eucalyptus forests at two years of age, utilizing satellite and climate information collected when the plantations were one year old. This early estimate of local quality facilitates the generation of a preliminary forest inventory. GARF was employed for variable selection and regression, incorporating climatic variables and biological influences into the model. This information will be integrated into a database to support decision-making, offering a solid basis for more efficient and sustainable forest management. The incorporation of computational intelligence, such as ML, further enhances the ability to constantly monitor forests and detect anomalies.