Lightweight Multimodal Fusion for Urban Tree Health and Ecosystem Services

Rapid urban expansion has heightened the demand for accurate, scalable, and real-time methods to assess tree health and the provision of ecosystem services. Urban trees are the major contributors to air-quality improvement and climate change mitigation; however, their monitoring is mostly constrained to inherently subjective and inefficient manual inspections. In order to break this barrier, we put forward a lightweight multimodal deep-learning framework that fuses RGB imagery with environmental and biometric sensor data for a combined evaluation of tree-health condition as well as the estimation of the daily oxygen production and CO₂ absorption. The proposed architecture features an EfficientNet-B0 vision encoder upgraded with Mobile Inverted Bottleneck Convolutions (MBConv) and a squeeze-and-excitation attention mechanism, along with a small multilayer perceptron for sensor processing. A common multimodal representation facilitates a three-task learning set-up, thus allowing simultaneous classification and regression within a single model. Our experiments with a carefully curated dataset of segmented tree images accompanied by synchronized sensor measurements show that our method attains a health-classification accuracy of 92.03% while also lowering the regression error for O₂ (MAE = 1.28) and CO₂ (MAE = 1.70) in comparison with unimodal and multimodal baselines. The proposed architecture, with its 5.4 million parameters and an inference latency of 38 ms, can be readily deployed on edge devices and real-time monitoring platforms.