NDVI-Augmented DeepLabV3+ with Dual-Level Attention for Amazon Deforestation Segmentation

Abstract

Accurate tropical deforestation mapping remains difficult because canopy gaps, degraded vegetation, cloud shadows, and exposed soil can blur class boundaries in visible-spectrum imagery. This study presents a multispectral semantic-segmentation framework that augments DeepLabV3+ with four coordinated components. First, per-band Z-score normalisation aligns Sentinel-2 channel statistics and stabilises optimisation. Second, the Normalised Difference Vegetation Index (NDVI), calculated from red and near-infrared reflectance, is injected as a fourth input channel to provide an explicit vegetation signal. Third, Squeeze-and-Excitation (SE) blocks recalibrate the four hierarchical outputs of a Mix Transformer B3 (MiT-B3) encoder and suppress spectrally irrelevant activations. Fourth, a Convolutional Block Attention Module (CBAM) applies channel and spatial attention after the decoder to refine features before pixel-level classification. Experiments used 619 Sentinel-2 image-mask pairs at 10 m spatial resolution and 512 × 512 pixels, divided into 459 training, 90 validation, and 70 test samples. The proposed framework achieved a test intersection-over-union (IoU) of 0.9459, Dice coefficient of 0.9721, and pixel accuracy of 0.9671. Relative to the DeepLabV3+ baseline and U-Net, the corresponding absolute IoU gains were 8.05 and 5.49 percentage points. Ablation results showed that NDVI injection restored 0.1850 IoU after normalisation, SE blocks produced the largest validation gain, and CBAM added 0.0905 test IoU. These results indicate that physics-informed spectral augmentation and dual-level attention jointly improve boundary-sensitive deforestation segmentation.