An EfficientUNet Approach for Forest and Non-Forest Segmentation Using Sentinel-2 Satellite Imagery

Abstract

EfficientUNet is evaluated for binary forest/non-forest semantic segmentation of Sentinel-2 imagery over Meghalaya, Northeast India, and benchmarked against U-Net, ViT-Seg, and DeepLabV3+. As no manually annotated dataset exists for this region, Pseudo Labels were generated via an adaptive spectral voting algorithm on a 40-channel feature stack. EfficientUNet leads all metrics with 92.85% accuracy, an IoU of 0.8836, and an F1 score of 0.9097, using only 7.9M parameters, roughly four times fewer than competing architectures. The transformer baseline underperforms convolutional models by up to 4.80 percentage points, attributed to limited inductive bias on moderately sized training sets. EfficientUNet also recovers fine-grained Non-Forest detail absent from the Pseudo Labels, indicating meaningful generalization beyond its supervisory signal. Reported metrics reflect agreement with algorithmically generated pseudo-labels rather than manually verified ground truth [11].