Biomass mapping for wildlife management using UAV-satellite integration and deep learning in Kui Buri National Park, Thailand

JEDSADA NOOWONG (Special Research Unit for Wildlife Genomics (SRUWG), Department of Forest Biology, Faculty of Forestry, Kasetsart University. 50 Phahonyothin Rd., Chatuchak District, Bangkok 10900, Thailand)
WARONG SUKSAVATE (Special Research Unit for Wildlife Genomics (SRUWG), Department of Forest Biology, Faculty of Forestry, Kasetsart University. 50 Phahonyothin Rd., Chatuchak District, Bangkok 10900, Thailand)
VISANUWIT THONG-ON (World Wide Fund for Nature International-Thailand (WWF-Thailand). Pradiphat Rd., Phayathai District, Bangkok 10400, Thailand)
NOPPORN AKKAMANEE (Department of National Parks, Wildlife and Plants Conservation. 61 Phaholyothin Rd., Chatuchak District, Bangkok 10900, Thailand)
ATTAPONG PAO-ON (Department of National Parks, Wildlife and Plants Conservation. 61 Phaholyothin Rd., Chatuchak District, Bangkok 10900, Thailand)
KORNSORN SRIKULNATH (Special Research Unit for Wildlife Genomics (SRUWG), Department of Forest Biology, Faculty of Forestry, Kasetsart University. 50 Phahonyothin Rd., Chatuchak District, Bangkok 10900, Thailand)
PRATEEP DUENGKAE (Animal Genomics and Bioresource Research Unit (AGB Research Unit), Faculty of Science, Kasetsart University. 50 Phahonyothin Rd., Chatuchak District, Bangkok 10900, Thailand)

Biodiversitas Journal of Biological Diversity
Vol. 26 No. 9 (2025)

Abstract. Noowong J, Suksavate W, Thong-On V, Akkamanee N, Pao-On A, Srikulnath K, Duengkae P. 2025. Biomass mapping for wildlife management using UAV-satellite integration and deep learning in Kui Buri National Park, Thailand. Biodiversitas 26: 4577-4597. In recent years, remote sensing has become a widely adopted tool for grassland management, offering advantages in assessing spatiotemporal dynamics. However, sample collection within satellite grids presents limitations because ground samples may not fully represent the corresponding satellite pixels. To address this challenge, the present study aimed to reduce sampling error by bridging ground-scale observations with satellite-scale data using digital images, Unmanned Aerial Vehicle (UAV) imagery, and Sentinel-2 for grass biomass estimation. A Convolutional Neural Network (CNN) was used to classify biomass from digital images and UAV imagery, whereas a random forest algorithm was utilized to link these classifications to freely available Sentinel-2 imagery. The study was conducted in cultivated grasslands managed for mitigating human-wildlife conflict in Kui Buri National Park, Thailand. The results showed that a pre-trained CNN model based on digital images (MAE±0.351 classes) successfully transferred to UAV imagery using fine-tuning. When scaled to the satellite level, the model explained 94% of the variance (R2), with RMSE = 8.56 g/m2. The grassland yield was lowest during the dry season, with a minimum value in March of 34.82±0.09 g/m² (67.12±0.33 tons/month), while it reached a peak during the wet season in November at 110.03±0.32 g/m² (212.06±1.19 tons/month). These finding demonstrate the ecology of ruzi grass under natural conditions with free-ranging wildlife grazing. Overall, the study highlights a viable strategy for bridging ground and satellite scales to reduce sampling error and proposes a novel approach to monitoring grassland yield at very high spatial resolution and high precision, while providing evidence of overgrazing and gaur overpopulation in the grasslands. Grazing rotation management was suggested to restore degraded grassland, and enhance the potential yields of the grassland. By integrating ecological insights with practical management recommendation, this study contributes to sustainable grassland restoration and wildlife conservation strategies.