2024

Gyawali, Arun; Adhikari, Hari; Aalto, Mika; Ranta, Tapio

Accurate canopy cover estimation is essential for mature and early-stage young forests, as it guides forest management and silvicultural activities necessary for their growth and regeneration. However, obtaining precise measurements of canopy cover in the field is time-consuming and challenging, especially at the regional and landscape levels. Remote sensing techniques offer a promising alternative to traditional field-based measurements for estimating forest canopy cover. In this study, our objective is to estimate forest canopy cover using vegetation indices derived from the multispectral bands of PlanetScope (Planet Lab, Inc., San Francisco, CA, USA). To the best of our knowledge, this is the first study to utilise PlanetScope imagery data for estimating canopy cover in young boreal forests. Based on the analysis of four bands (green, blue, red, and near-infrared) from PlanetScope imagery, 43 vegetation indices, including four spectral bands and 13 salinity indices, were computed to select predictors in canopy cover modelling. Six regression models were employed to model canopy cover: linear, elastic net, support vector machine, random forest, extreme gradient boosting, and light gradient boosting machine. All the models demonstrated good performance for both the training dataset (R2 = 0.58–0.69) and the testing dataset (R2 = 0.59–0.64, RMSE = 0.16–0.18, rRMSE = 22%–23%, and MAE = 0.12–0.14). Based on the fit statistics in the training and testing datasets and the paired t-test, our study identified the light gradient boosting machine as the most suitable model for predicting canopy cover in young boreal forests. For the light gradient boosting machine, the R2 value was 0.69 (training), and for testing data, the R2 = 0.64, RMSE = 0.16, rRMSE = 22%, and MAE = 0.12. Therefore, we recommend that future researchers utilise Planet multispectral data and the light gradient boosting machine regression to estimate forest canopy cover at a higher spatial resolution. However, exploring additional machine learning algorithms and explicitly boosting methods when computing forest canopy cover using satellite remote sensing is strongly advised.