Early detection of drought-stressed stands in Mediterranean forests using remote sensing and machine learning classification models in a rainfall exclusion experiment

Abstract

Climate change-driven droughts increasingly threaten Mediterranean forests. Early detection is crucial for mitigating long-term impacts; yet, conventional methods are limited in spatial and temporal coverage. Remote sensing offers a large-scale solution, but its application at the individual-tree level remains limited, particularly in mixed-species forests.

We combined a controlled rainfall exclusion experiment with drone-based hyperspectral imaging and machine learning to classify drought stress at the individual-tree level in a semi-arid Mediterranean forest (Yishi Forest, Israel). Six 0.05-ha plots with five co-occurring tree species were monitored over two hydrological years. Hyperspectral data (274 bands, 400–1000 nm) were used as is and after synthetically simulating Planet, VENµS, and Sentinel-2 bands in three machine learning classification models.

Results show that rainfall was reduced by nearly half in treated plots. Standard physiological metrics—leaf water potential, carbon assimilation, and transpiration—showed limited treatment sensitivity across most species and seasons, whereas hyperspectral-driven machine learning classification models accurately distinguished between drought-treated and control stands. Logistic Regression (LR) outperformed Support Vector Machines (SVM) and Random Forest (RF), reaching an accuracy of 0.85, a recall of 0.94, and an F1 score of 0.83 in classifying treated stands on a held-out test set. High performance persisted after reducing input to 21 bands. Simulated satellite spectral data showed that SVM performed best using VENµS bands (accuracy = 0.74, F1 = 0.73). When applied to real VENµS imagery from three independent forest sites, the model identified areas of high drought risk one to two years before visible canopy decline.

The presented approach offers a scalable and transferable tool for real-time forest drought monitoring, supporting early warning systems amid growing climate pressures.