A forest type-specific threshold method for improving forest disturbance and agent attribution mapping

Abstract

ABSTRACT Detecting forest disturbances and attributing their contributing agents with Landsat time series (LTS) images has advanced substantially in recent years; however, whether different forest types require individual disturbance indices or specific disturbance thresholds to accurately map forest disturbances and their causes over a vast region remains limited known. This study investigated the effectiveness of six spectral indices (SIs) and two threshold methods (a forest-specific threshold and a common threshold) for detecting forest disturbances among four forest types, namely, evergreen broad-leaved forests (EBFs), cold-temperate evergreen needle-leaved forests (CENFs), and subtropical evergreen needle-leaved forests mainly dominated by Pinus yunnanensis (SENF1) or Pinus kesiya (SENF2), across Yunnan Province with the Landsat-based Detection of Trends in Disturbance and Recovery (LandTrendr) algorithm and yearly Landsat time series (LTS) recorded from 1990 to 2020. A random forest (RF) model was applied to classify the forest disturbance agents from aggregated patches of disturbed forest pixels. The results indicated that the normalized burn ratio (NBR) outperformed the five other SIs and achieved consistently high overall accuracies (OAs; 93.04%±0.17% to 96.09%±0.28%) when mapping forest disturbances across all four forest types. The forest-specific NBR disturbance thresholds led to considerable increases in overall (0.14–3.92%), producer (0.25–13.47%) and user accuracies (0.88–3.01%). The total mapped area of disturbed forest was 9831.48 km2 (5.31%), of which approximately 79% occurred in the EBF and SENF1 distribution areas. Forest disturbances were predominantly caused by wildfires in CENF and SENF1 and by commodity-driven plantations in EBF and SENF2; these two agents together contributed approximately 93.15% of the forest disturbances in Yunnan province. These findings highlight that the optimal selection of SIs and forest-specific disturbance thresholds can significantly improve forest disturbance detection performance.