Deep Learning Segmentation Models Evaluation for Deforestation Monitoring Embedded Systems

This work evaluates deep learning segmentation models to propose a deforestation monitoring embedded system. The approach stands for environmental monitoring using remote sensing imagery, edge computing, and a deep learning segmentation model. Thus, the performance of you only look once architecture version 8 (YOLOv8) and Mask Region-based convolutional neural networks (Mask R-CNN) embedded in Raspberry Pi Model 4 regarding Intersection over Union (IoU), mean Average Precision (mAP), and time per image processing metrics is compared. The models are combined with a pixel-based algorithm that analyzes the temporal segmented images to define their forest area percentage for deforestation monitoring and detection. The results demonstrate YOLOv8x model achieved an IoU of 0.762, with a time per image of 0.4777 seconds, while Mask R-CNN R101 FPN 3x obtained an IoU of 0.763, with a time per image of 0.2669 seconds. The average times for YOLOv8 ranged from 0.0434 to 0.4777 seconds, and for Mask R-CNN from 0.1969 to 0.2669 seconds. Finally, this work proposes evaluating the model's performance when working with generative AI models Dall-e, Craiyon, and Tess-AI to create a synthetic dataset to augment the initial one with synthetic samples and improve the model's training with a large dataset. The Dall-e has been shown to outperform the others regarding the IoU metric, which was suggested to augment datasets with synthetic samples.