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 $3\times 3$ 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.

Many cloud and shadow detection methods have been proposed already, but improvements can be made on accuracy or automation. In this study, we propose a Fully Convolutional Network model for the detection of clouds and shadows in optical satellite images. The proposed model was trained on 165 Landsat images in Finland, and tested on an independent set of images. The cloud and shadow detection accuracy reached 95%, outperforming both quantitatively and qualitatively a selection of other deep learning architectures.

We studied the possibility of replacing a complex forest growth and productivity model with a deep learning model with sufficient accuracy. We used three different neural network architectures for emulating the prediction task of the PREBASSO (Mäkelä 1997; Minunno et al. 2016) forest growth model: 1) Recurrent Neural Network (RNN) Encoder-decoder network, 2) RNN encoder network, and 3) Transformer encoder network. The PREBASSO forest growth model was used to produce 25-year predictions for forest variables: tree height, stem diameter, basal area, and the carbon balance variables: net primary production (NPP), gross primary production per tree layer (GPP), net ecosystem exchange (NEE) and gross growth (GGR) to train the machine learning models. The Finnish Forest Centre provided the data for 29 619 field inventory plots in continental Finland that were used as the initial state of the forest sites to be simulated. Climate data downloaded from Copernicus Climate Data Store were used to provide realistic climate scenarios. We emphasized the importance of low bias in long term predictions and set the goal for the emulator prediction relative bias to be within ±2%. The RNN encoder model produced the best results with the mean of the yearly bias values within the specified ±2% limit over the 25-year prediction period. The study shows that emulating the operation of analytical forest growth models is feasible using state-of-the-art machine learning methods and indicates the potential of using such emulators for producing long time span simulations for e.g. digital twins.

Our objective was to develop a method for the assessment of forest area and structural variables for cases in which the availability of representative ground reference data is poor and these data are not collected from the whole area of interest. We implemented two independent approaches to the estimation of the forest variables of a European boreal forest: (i) the computation of wall-to-wall estimates using moderate- to low-resolution VIIRS imagery from the Suomi NPP mission; and (ii) the visual interpretation of plots of samples from very high resolution (VHR) satellite data obtained via a two-stage design. Our focus was on the statistical comparison of forest resources at a country or larger level. The study area was boreal forest ranging from Norway to the Ural Mountains in Russia. We computed a seamless mosaic from 111 VIIRS images. From the mosaic, we computed predictions for the forest area, growing stock volume, height of the dominating tree layer, proportion of conifers and broadleaved trees, site fertility class, and leaf area index. The reference data for the VIIRS imagery were national forest inventory (NFI)-based raster maps from Finland. The first stage sample of VHR data included 42 images; of these, a second stage sample of 2690 plots was visually interpreted for the same variables. The forest area prediction from VIIRS for the whole study area was 1.2% higher than the VHR-based result. All other structural variable predictions using VIIRS fitted within the 95% confidence intervals computed from the VHR sample except for estimates of the main tree species groups, which were outside the limits. A comparison of VIIRS-based forest area estimates using Finnish and Swedish NFI data indicated overestimations of 10.0% points and 4.6% points, whereas the total growing stock volumes were overestimated by 8% and underestimated by 3.4%, respectively. The correlation coefficients between the VIIRS and VHR image predictions at the 42 VHR image locations varied from 0.70 to 0.85. The VIIRS maps strongly averaged the local predictions due to their coarse spatial resolutions. Based on our findings, the approach using two independent estimations yielded similar figures for the central forest variables for the European boreal forest. A model computed using reference data from a small part of the area of interest can provide satisfactory predictions for a much larger area with a similar biome. Therefore, our concept is applicable to the estimation and overall mapping of a forest area and central structural variables at regional to national levels.

Retrieval of soil organic carbon (SOC) from uncrewed aerial vehicle (UAV) hyperspectral imagery depends on how laboratory measurements are linked to pixels at centimetre resolution. This study presents a quality-controlled workflow in which radiometric calibration, harmonisation against proximal spectroradiometer measurements, pixel-level screening, spectral support definition and model validation are treated as interdependent components of a single procedure. Imagery was acquired with a tunable Fabry–Pérot interferometer camera in six bands between 670 and 1100~nm over eight bare soil fields in Flanders, Belgium (85 topsoil samples) and five fields in Central Macedonia, Greece (42 samples). Single pixel spectra, neighbourhood aggregation and local pixel augmentation were compared at radii of 0.25, 0.50 and 1.00~m using Random Forest, XGBoost, a 1D convolutional neural network and TabPFN, with all pixels from a given sampling location assigned to the same cross-validation fold. In Flanders, 23 of 24 neighbourhood configurations reduced RMSE relative to single pixel extraction, the best reaching $R^2 = 0.784$ and RMSE =0.155% SOC. In Central Macedonia, adverse acquisition conditions and a narrow target range prevented a usable SOC model, although electrical conductivity, CaCO₃ and pH reached $R^2$ between 0.61 and 0.69. These results indicate that the spectral support should be defined to match the field sampling support, and that this correspondence belongs alongside radiometric calibration, pixel screening and validation in a reproducible workflow for field-scale soil mapping.