Automated Monitoring Platform for Forest Restoration with Sensors and AI

This project develops a web-based system to monitor forest restoration using drones and AI, evaluating the effectiveness of restoration efforts at the UHE São Simão reservoir and providing more accurate results. Two field campaigns were conducted across 20 areas, totaling 208 hectares mapped using remote sensing techniques. The monitoring employed a DJI Matrice 210 V2 drone, equipped with a Livox Avia LiDAR sensor and a Sentera 6X multispectral sensor. The multispectral sensor captured images in six spectral bands (R, G, B, red edge, NIR, and RGB) with a spatial resolution of 4.8 cm per pixel, enabling the extraction of indices such as NDVI, SAVI, EVI, and AVI. NDVI proved effective in identifying areas with higher vegetative vigor, while AVI excelled at detecting invasive species. The images were processed and converted into orthomosaics, which were then cropped into smaller sections to facilitate algorithm execution. In total, 8,263 orthophotos, each 550x550 pixels, were generated, with 6,611 used for training and 1,652 for validating the AI models. Semantic segmentation was performed using the Awesome algorithm, and instance segmentation was conducted using Yolact, enabling detailed classification of vegetation, soil, tree species, and wildlife presence. The model achieved 82% accuracy for certain classes and 90% pixel-level accuracy in the final validation. The LiDAR data, with an average density of 200 pts/m², were processed using the R lidR package for manipulating LiDAR data. This enabled tree counting, the generation of DTMs (Digital Terrain Models) and CHMs (Canopy Height Models), and the extraction of metrics such as wood volume and biomass. Validation of the results was carried out using conventional inventories in 100 m² plots, serving as reference points for remote sensing estimates. The data are integrated into an interactive web platform featuring dashboards and interactive maps, which facilitate access to indicators and support real-time decision-making.