Detection of thinnings and clearcuts in boreal forests using bi-temporal airborne laser scanning, aerial images and time series of Sentinel-2 images

Most forests in Finland are in commercial use, thus the amount of felled roundwood is several tens of millions cubic meter every year. Nevertheless, there are biodiversity hotspots protected by law and certification standards that should not be affected by cuttings. While clear cut areas are already well monitored by Finnish Forest Centre, the detection of thinnings has been more challenging when relying solely on spectral information from satellite imagery. Therefore, our study aimed to evaluate whether textural features could improve the detection of thinned stands. In this study, we used Haralick’s textural features, template matching as a line detection method, and spectral values. As data source, we used bi-temporal airborne laser scanning (ALS) data, aerial images and temporal series of Sentinel-2 images. Ordinal logistic regression with three classes (clear cut, thinning, and no change) was used in modelling. The models used in this study were all features together, Sentinel-2 and aerial images together, ALS, aerial images, Sentinel-2, and Sentinel-2 without SWIR and red edge bands. Two study areas were used to create models, and the third area was used as validation dataset. We had previous information about realized clearcuttings and thinnings for all study areas. The results showed that thinned stands were detected most accurately from ALS data (F₁ score 97.4%). Overall, ALS data yielded good results for all classes, whereas aerial images produced the poorest results. F₁ score for clear cuttings varied between 91.8% – 99.4%, for thinnings F₁ score varied from 35.7% – 97.4% and for unchanged values varied between 82.7% – 99.4%. Average F₁ score varied between 70.3% – 98.7% and weighted kappa varied from 0.79 to 0.99. Most misclassifications occurred between thinnings and unchanged stands, while clear cuttings were always predicted most accurately. Our results showed that ALS can produce highly accurate estimates of forest management activities, whereas aerial images were possibly more sensitive to shadows and thinning intensity.