Improving DeepLabV3+ Using Normalized Satellite Indices in Land-Cover Segmentation

Abstract

The availability of land-cover segmentation and classification maps at multiple time frames is crucial for designing spatial and regional planning. At present, remote sensing and geo-graphic information system practitioners rely on object-based image analysis for land-cover segmentation/classification. Although deep learning methods are available, their use remains limited to satellite imagery datasets. DeepLabV3+ and U-Net are popular methods owing to their accuracy and speed. In this study, we propose a method for enhancing the accuracy of DeepLabV3+ to closely match ground truth datasets by integrating the normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI), and normalized dif-ference water index (NDWI) on the decoder side to correct land-cover segmentation. Testing of the proposed method in Karawang Regency, West Java, Indonesia demonstrated a 0.3% improvement in accuracy when the NDVI, NDBI, and NDWI were incorporated on the output side of DeepLabV3+.