Aplicação de inteligência artificial na identificação automática de mudas de restauração florestal em imagens RGB e multiespectrais provenientes de aeronaves remotamente pilotadas

Abstract

Forest restoration corresponds to several actions and techniques aimed at repairing and maintaining forest ecosystems. In Brazil, forest restoration plays an important role in the restitution of the Atlantic Forest biome and only gained ground in the 21st century with the introduction of laws that demand compensation and repair of environmental damage. In this sense, forest restoration can not be applied aiming only at the techniques that best serve the study area, but also it is essential to consider its long-term monitoring to verify the progression of the restoration process. Nowadays, restoration monitoring uses equipment such as the Remotely Piloted Aircraft System (RPAS) that through loco activities and image processing in the laboratory aids the periodic inspection of the restored area. It enables information such as the number of resilient individuals, species classification, and vegetation height of the given area. Such monitoring practices require a large investment of time and human resources. Those practices may contribute to greater cost-effectiveness of the restoration process if they are optimized. In front of the current impediments in optimizing forest restoration monitoring, this project aims to apply the artificial intelligent algorithm Mask R-CNN to automate the counting and delineation of restoration seedlings in images from RPAS, comparing the performance of this algorithm in RGB and multispectral (MSP) images. The results obtained by the study showed that the algorithm was more efficient when applied to RGB images, identifying 717 seedlings out of a total of 1069. On the other side, the MSP images obtained a lower performance, identifying 61 seedlings out of a total of 1011. Another outcome of the study was the Jaccard index, which refers to the percentage of hits in the delineation (overlap/intersection analysis) of identified objects (seedlings), in this case, such as to the images RGB as to the MSP, the algorithm presented an index of intersection ranging from 50 to 95%. The contribution of this study to future academic research in monitoring and mapping the vegetation shows that it is possible to automate the counting of seedlings delineation in RPAS derived from images. Moreover, it saves resources and time, acknowledging that the use of artificial intelligence algorithms represents a promising method for accomplishing this function. It is also concluded that the greater efficiency of the algorithm in RGB images rules out the need to use a multispectral sensor for this purpose, lowering the cost of equipment and bringing alternatives to make monitoring forest restoration more viable.