Estimation of forest structural variables is essential to provide relevant insights for public and private stakeholders in forestry and environmental sectors. Airborne light detection and ranging (LiDAR) enables accurate forest inventory, but it is expensive for large area analyses. Continuously increasing volume of open Earth Observation (EO) imagery from high-resolution (<30 m) satellites together with modern machine learning algorithms provide new prospects for spaceborne large area forest inventory. In this study, we investigated the capability of Sentinel-2 (S2) image and metadata, topography data, and canopy height model (CHM), as well as their combinations, to predict growing stock volume with deep neural networks (DNN) in four forestry districts in Central Finland. We focused on investigating the relevance of different input features, the effect of DNN depth, the amount of training data, and the size of image data sampling window to model prediction performance. We also studied model transfer between different silvicultural districts in Finland, with the objective to minimize the amount of new field data needed. We used forest inventory data provided by the Finnish Forest Centre for model training and performance evaluation. Leaving out CHM features, the model using RGB and NIR bands, the imaging and sun angles, and topography features as additional predictive variables obtained the best plot level accuracy (RMSE% = 42.6%, |BIAS%| = 0.8%). We found 3×3 pixels to be the optimal size for the sampling window, and two to three hidden layer DNNs to produce the best results with relatively small improvement to single hidden layer networks. Including CHM features with S2 data and additional features led to reduced relative RMSE (RMSE% = 28.6–30.7%) but increased the absolute value of relative bias (|BIAS%| = 0.9–4.0%). Transfer learning was found to be beneficial mainly with training data sets containing less than 250 field plots. The performance differences of DNN and random forest models were marginal. Our results contribute to improved structural variable estimation performance in boreal forests with the proposed image sampling and input feature concept.

The demonstration conducted in this study lays foundation for further development of an operational large area forest monitoring system that allows annual reporting of forest biomass and carbon balance from forest stand level to regional analyses. The system is seamlessly aligned with process based ecosystem modelling, enabling forecasting and future scenario simulation.

Abstract:In land monitoring applications, clouds and shadows are considered noise that should be removed as automatically and quickly as possible, before further analysis. This paper presents a method to detect clouds and shadows in Suomi NPP satellite's VIIRS (Visible Infrared Imaging Radiometer Suite) satellite images. The proposed cloud and shadow detection method has two distinct features when compared to many other methods. First, the method does not use the thermal bands and can thus be applied to other sensors which do not contain thermal channels, such as Sentinel-2 data. Secondly, the method uses the ratio between blue and green reflectance to detect shadows. Seven hundred and forty-seven VIIRS images over Fennoscandia from August 2014 to April 2016 were processed to train and develop the method. Twenty four points from every tenth of the images were used in accuracy assessment. These 1752 points were interpreted visually to cloud, cloud shadow and clear classes, then compared to the output of the cloud and shadow detection. The comparison on VIIRS images showed 94.2% correct detection rates and 11.1% false alarms for clouds, and respectively 36.1% and 82.7% for shadows. The results on cloud detection were similar to state-of-the-art methods. Shadows showed correctly on the northern edge of the clouds, but many shadows were wrongly assigned to other classes in some cases (e.g., to water class on lake and forest boundary, or with shadows over cloud). This may be due to the low spatial resolution of VIIRS images, where shadows are only a few pixels wide and contain lots of mixed pixels.

A method to detect abrupt land cover changes using hierarchical clustering of multi-temporal satellite imagery was developed. The Autochange method outputs the pre-change land cover class, the change magnitude, and the change type. Pre-change land cover information is transferred to post-change imagery based on classes derived by unsupervised clustering, enabling using data from different instruments for pre- and post-change. The change magnitude and change types are computed by unsupervised clustering of the post-change image within each cluster, and by comparing the mean intensity values of the lower level clusters with their parent cluster means. A computational approach to determine the change magnitude threshold for the abrupt change was developed. The method was demonstrated with three summer image pairs Sentinel-2/Sentinel-2, Landsat 8/Sentinel-2, and Sentinel-2/ALOS 2 PALSAR in a study area of 12,372 km2 in southern Finland for the detection of forest clear cuts and tested with independent data. The Sentinel-2 classification produced an omission error of 5.6% for the cut class and 0.4% for the uncut class. Commission errors were 4.9% for the cut class and 0.4% for the uncut class. For the Landsat 8/Sentinel-2 classifications the equivalent figures were 20.8%, 0.2%, 3.4%, and 1.6% and for the Sentinel-2/ALOS PALSAR classification 16.7%, 1.4%, 17.8%, and 1.3%, respectively. The Autochange algorithm and its software implementation was considered applicable for the mapping of abrupt land cover changes using multi-temporal satellite data. It allowed mixing of images even from the optical and synthetic aperture radar (SAR) sensors in the same change analysis.

No abstract