Conditional Generative Adversarial Networks for Enhanced Aerial Forest Image Segmentation

<p dir="ltr">Forest ecosystems have long been one of the most important environments for our planet, providing key resources, promoting biodiversity, and fighting against global climate change. In order to facilitate effective forest monitoring and management, artificial intelligence can be used to address the need for data processing for real-time forest supervision systems. In this paper, conditional generative adversarial networks (CGANs) have been explored to synthesize accurate image segmentations of forest aerial images, mapping forested against non-forested areas. With 1000 training, 200 validation, and 100 test images subsetted from Kaggle’s Forest Aerial Images for Segmentation dataset, three CGANs of varying parameter number and upsampling and downsampling layers have been trained and evaluated. The results of training show that the smallest CGAN, with 37x less generator and 4x less discriminator parameters than the biggest CGAN, performed the best with an IoU of 0.701, Dice coefficient of 0.778, pixel accuracy of 0.781, recall of 0.919, and precision of 0.734 in the test set. Using a weighted scoring algorithm comparing inference time in addition to the five aforementioned metrics, the medium CGAN was determined to be the best, with a weighted score of 0.861 closely followed by the small CGAN’s 0.783 score for the dataset. These outputs signify the need for model complexity and dataset size compatibility, the importance quality labelled annotations for GAN conditioning, and most importantly, the potential of CGANs for accurate, automated, and effective segmentation of aerial forest images.</p>