Harnessing Time-Series Satellite Data and Deep Learning to Monitor Historical Patterns of Deforestation in Eastern Himalayan Foothills of India

• Published in: Journal of the Indian Society of Remote Sensing Vol. 53; no. 4; pp. 993 - 1008
• Main Authors: Bhuyan, Jintu Moni, Nandy, Subrata, Padalia, Hitendra, Pebam, Rocky, Nath, Arun Jyoti, Kushwaha, S. P. S., Das, Ashesh Kumar
• Format: Journal Article
• Language: English
• Published: New Delhi Springer India 01.04.2025; Springer Nature B.V
• Subjects: Algorithms; Biodiversity; Biodiversity hot spots; Biodiversity loss; Carbon; Classification; Climate change; conservation areas; Data collection; Datasets; Deep learning; Deforestation; Earth and Environmental Science; Earth Sciences; Effectiveness; Emissions; energy; Environmental monitoring; Foothills; Forest management; forests; Himalayan region; Hydrologic cycle; image analysis; Image segmentation; India; Landsat satellites; Machine learning; Neural networks; Performance measurement; Python; Remote sensing; Remote Sensing/Photogrammetry; Research Article; Tensors; time series analysis; Training; Trends; Tropical forests; Vegetation; Wildlife; Wildlife conservation; Wildlife refuges; India; Himalayan region; Landsat time series; Northeast India; Tropical forests; Land use/land cover change; Google Earth Engine; U-Net
• ISSN: 0255-660X; 0974-3006
• DOI: 10.1007/s12524-025-02137-8

The Indian Eastern Himalaya, part of the Himalaya global biodiversity hotspot, is experiencing extensive deforestation which significantly contributes to biodiversity loss and climate change, highlighting the complex interconnections between tropical forests and the global carbon, energy, and water cycles. In this context, this pioneering study in the Eastern Himalayan region conducted the first comprehensive assessment of long-term, large-scale deforestation using the deep learning algorithm U-Net, employing a novel approach by modifying the model to accommodate multiple tensor inputs during the training, unlike the conventional single tensor input method. The U-Net model demonstrated a testing accuracy of 84% after 100 epochs, with no notable improvements observed even after extending training to 1000 epochs. Therefore, the model trained for 100 epochs was selected for further analysis, demonstrating stable performance without signs of overfitting. Key performance metrics indicate robust image segmentation capabilities, including a precision of 0.82, recall of 0.79, and intersection over union of 0.68. These results underscore the model’s effectiveness in capturing temporal patterns and the changes in forest cover. This study examined significant deforestation trends over the past 34 years (from 1990 to 2024), revealing a drastic reduction in forest cover by 581.92 km², from 1,091.82 km² in 1990 to 509.90 km² in 2024, with an annual deforestation rate of 2.24%. Charduar Reserved Forest experienced the highest deforestation, losing 126.11 km² with at an annual rate of 7.61%, while Sonai Rupai Wildlife Sanctuary lost 99.29 km² at an annual rate of 1.84%. This study highlights the effective use of advanced technology for monitoring deforestation. By leveraging free satellite data and deep learning models, it enables large-scale deforestation tracking, crucial for understanding forest carbon depletion and developing mitigation strategies. The insights gained can significantly advance environmental monitoring and conservation planning.