Accurate spatial mapping of bamboo land cover is essential for resource management, ecological monitoring, and livelihood planning in Northeast India. However, the detection of fragmented bamboo-dominated patches in complex mountainous terrain remains challenging. This study develops and evaluates a deep learning-based semantic segmentation framework for mapping bamboo land cover in Mizoram using a nine-channel multi-sensor composite. The composite integrates high-resolution LISS-IV (5.8 m) surface reflectance data with Sentinel-2 Level-2A red-edge and narrow near-infrared bands (B5, B6, B7, B8A), as well as derived vegetation indices (NDVI, EVI2). The integration of high spatial resolution and red-edge spectral information was designed to enhance discrimination of bamboo-dominated vegetation within a multi-source mapping framework. A U-Net architecture with a ResNet-34 encoder was trained on 813 labelled patches acquired across multiple states of Northeast India. The model achieved a test Intersection over Union (IoU) of 0.8456, an F1-score of 0.9163, a precision of 0.9200, and a recall of 0.9127, evaluated at the patch level on a held-out test subset. Independent-site assessment across Mizoram, Manipur, and Assam indicates consistent model performance across diverse landscape conditions. Deployment across nine LISS-IV tiles covering Mizoram produced a bamboo land cover map at 5.8 m spatial resolution, with an estimated extent of approximately 6262 km 2 (29.7% of the state’s geographical area). The results demonstrate the potential of integrating high-resolution imagery, red-edge spectral information, and deep learning for bamboo land cover mapping in complex mountainous environments and provide a potentially transferable framework for regional-scale bamboo resource assessment.
Islam et al. (Sun,) studied this question.