Extending Airborne Lidar-Derived Estimates of Forest Canopy Cover and Height Over Large Areas Using kNN With Landsat Time Series Data

Airborne light detection and ranging (lidar) data provide an accurate and consistent means to obtain reliable forest canopy cover (CC) and height measurements, which are important in determining forest stand structure, volume, and biomass. Extending CC and height measurements over larger areas by integration with satellite imagery increases the value of airborne lidar data. A typical approach has been to use multiple regression, machine-learning, or regression tree methods to determine relationships between the forest structure variables measured in the lidar data and available single-date or multitemporal Landsat sensor reflectance data, and if the relationships are strong enough, to extend those variables over much larger areas than is typically covered by lidar. Such methods can be difficult to apply because of the complexity of the extending models and algorithms, and the long processing times, which may become prohibitive. One machine-learning approach, which uses the k-nearest neighbor (kNN) algorithm, is presented here in a British Columbia, Canada forest environment. Our goal was to simplify the estimation of lidar-derived forest structure variables with available Landsat time series data and compare the results of the kNN model to the traditional multiple regression results, and to those obtained with more complex and computationally demanding random forest (RF) methods. We develop and test the kNN model with airborne lidar-derived estimates of forest CC and height in 1846 relatively young and mature forests. The best kNN model produced estimates of airborne lidar-derived CC in validation sample (n = 1132) of mature forest with three Landsat time series variables with an R² = 0.74 and RMSE of approximately 10%. This result is comparable to the results obtained in earlier work using more complex machine-learning approaches (in approximately 1/10 the time). Younger (i.e., recently disturbed) forest CC estimation was less successful in the kNN model because of the high degree of structural variability in these forests. Extending airborne lidar estimates of forest height to larger areas was also possible though, as expected, less successful than CC estimation using the kNN approach.