Multi-Modal Transformer Framework for Accurate Detection and Counting of Standing Dead Trees in Forest

Forest health monitoring is critical for ecological stability and carbon management. Dead trees serve as indicators of forest decline and increase the risk of wildfires. Traditional surveys are costly and limited in scale, making automated aerial detection vital. This paper presents a Multi-Modal Transformer Framework integrating RGB and Near Infrared data for accurate detection and counting of standing dead trees. The system employs three deep learning architectures U-Net, Feature Pyramid Network and U-Net++ with modality-specific encoders and transformer fusion layers to jointly optimize segmentation and counting tasks. Experimental evaluation using the Aerial Imagery for Standing Dead Tree Segmentation dataset demonstrates that the U-Net++ (ResNet-50) model achieved superior performance with an F1-score of 0.518, highlighting its capability for scalable and accurate forest health assessment.