Optimizing forest defoliation detection using remote sensing data: a multi–resolution approach using machine learning algorithms

Abstract

Remote sensing technologies, particularly satellite-based imagery, offer an effective means to monitor large-scale forest disturbances including pest-induced defoliation. This study investigates the use of multi-source remote sensing data, specifically PlanetScope (3 m spatial resolution) and Sentinel-2 (10 m and 20 m spatial resolution), combined with derived spectral vegetation indices (SVIs), to detect disturbances like spruce budworm (SBW)-induced defoliation in northeastern USA. We comprehensively evaluated the performance of various model building scenarios incorporating sensor types as well as spatial and spectral resolutions for SBW defoliation detection. Three machine learning algorithms—random forest (RF), support vector machine (SVM), and multi-layer perceptron (MLP)—were applied to model defoliation at a landscape scale. The RF algorithm outperformed the others in defoliation detection. Five models using various Sentinel-2 and PlanetScope band combinations were evaluated, and all produced low error rates. The model based on Sentinel-2 (20 m resolution), using all bands and SVIs (Model V), provided the best performance with a modeling error of 6.1 % followed by Sentinel-2 variables (20 m resolution) with bands and SVIs comparable to PlanetScope (Model III) with a modeling error of 7.6 %. The PlanetScope-based model (Model I) had a modeling error of 11.4 %. The Soil Adjusted Vegetation Index (SAVI) and Modified Simple Ratio (MSR) were the most effective indices for detecting defoliation. However, SVIs selection was affected by the timing of image acquisition. Models based on a combination of SVIs related to canopy structure, stress, and biochemistry resulted in similar accuracy levels (error range: 6.1 %-9.1 %). The models suggested in this study can be used for timely SBW defoliation at landscape scale.