No abstract

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 spectral and spatial resolutions of modern optical Earth observation data are continuously increasing. To fully utilize the data, integrate them with other information sources and create applications relevant to real-world problems, extensive training data are required. We present TAIGA, an open dataset including continuous and categorical forestry data, accompanied by airborne hyperspectral imagery with a pixel size of 0.7 m. The dataset contains over 70 million labeled pixels belonging to more than 600 forest stands. To establish a baseline on TAIGA dataset for multitask learning, we train and validate a convolutional neural network to simultaneously retrieve 13 forest variables. Due to the size of the imagery, the training and testing sets were independent, with strictly no overlap for patches up to 45×45 pixels. Our retrieval results show that including both spectral and textural information improves the accuracy of mapping key boreal forest structural characteristics, compared with an earlier study including only spectral information from the same image. TAIGA responds to the increased availability of hyperspectral and very high resolution imagery, and includes the forestry variables relevant for forestry and environmental applications. We propose the dataset as a new benchmark for spatial-spectral methods that overcomes limitations of widely used small-scale hyperspectral datasets.

Europe has acknowledged the need to develop a very high precision digital model of the Earth, a Digital Twin Earth, running on cloud infrastructure to bring data and end-users closer together. We present results of an investigation of a proposed submodel of the digital twin, simulating the worlds' forests. We focus on the architecture of the system and the key user needs on data content and access. The results are based on a user survey showing that the forest-related communities in Europe require information on contrasting forest variables and processes, with common interest in the status and forecast of forest carbon stock. We discuss the required spatial resolution, accuracies, and modelling tools required to match the needs of the different communities in data availability and simulation of the forest ecosystem. This, together with the knowledge on existing and projected future capabilities, allows us to specify a data architecture to implement the proposed system regionally, with the outlook to expand to continental and global scales. Ultimately, a system simulating the behaviour of forests, a digital twin, would connect the bottom-up and top-down approaches of computing the forest carbon balance: from tree-based accounting of forest growth to atmospheric measurements, respectively.

&lt;p&gt;Three-quarters of Finland&amp;#8217;s land surface area (22.8 million hectares) is filled with forests. The role of remote sensing in large area inventories is crucial. The forests of Finland serve as an important resource for the nation&amp;#8217;s nature conservation as well as for the forestry industry. Furthermore, forests are significant carbon sinks and play a great role in climate change mitigation. Research on vegetation parameter retrieval is of special relevance in order to extend our knowledge about the vegetation dynamics and terrestrial carbon stocks at regional and global scales.&lt;/p&gt;&lt;p&gt;In future in addition to multispectral satellites, hyperspectral satellite missions will start to provide remote sensing data to support the needs of forestry and other natural resource management practices. We investigated the influence of spectral and spatial resolution of remote sensing data on retrieval of biomass and other forest properties. The study contributed to better information productivity on forest variables in boreal forest ecosystem.&lt;/p&gt;&lt;p&gt;We used the remote sensing data by Sentinel-2 (10 bands, resolution 10 m) and hyperspectral AISA imager (128 bands, 400&amp;#8211;1000 nm, resolution 0.7 m). As reference data, we used new forest resource dataset provided by the Finnish Forest Centre and additional independent in situ measurements. We applied kernel-based regression methods to relate the forest variables of interest with the remotely sensed data. Based on recent studies, we selected Gaussian process regression (GPR) and support vector regression (SVR), which have proven to work well with hyperspectral and multispectral remote sensing data. Regression estimations were performed for stem biomass, basal area, mean height, leaf area index (LAI) and main tree species. The estimation accuracies were examined with absolute and relative root-mean-square errors.&lt;/p&gt;&lt;p&gt;Successful forest variable estimations showed that kernel-based regression algorithms are suitable tools for quantification of forest structure and assessment of its change. The estimation accuracies between the two algorithms were similar. However, the faster SVR algorithm was found to be more practical, especially considering large scale mapping and future near real-time applications. Based on the study results, the additional value of hyperspectral remote sensing data in forest variable estimation in Finnish boreal forest is mainly related to variables with species-specific information, such as main tree species and LAI. The more interesting variables for forestry industry, such as basal area or stem biomass, can also be estimated accurately with more traditional multispectral remote sensing data.&lt;/p&gt;

Very high resolution remote sensing data of forests, where individual tree crowns are separable, contains structural information on tree size and density. Such information is complementary to the spectral signatures currently used in forestry applications. Advanced machine learning methods, e.g. convolutional neural networks (CNNs), offer an automated and standardized way of retrieving both spectral and structural information from imagery. A key characteristic in CNNs is patch size, which should be large enough to include dominant structural scale, yet as small as possible to avoid unnecessary averaging. Our results show that the patch should be larger than one tree, but increasing it excessively reduces retrieval accuracy. Furthermore, large patch sizes can cause loss of independence between training and validation data, leading to overestimating model performance.

No abstract

Physically-based reflectance models offer a robust and transferable method to assess biophysical characteristics of vegetation in remote sensing. Forests exhibit explicit structure at many scales, from shoots and branches to landscape patches, and hence present a specific challenge to vegetation reflectance modellers. To relate forest reflectance with its structure, the complexity must be parametrised leading to an increase in the number of reflectance model inputs. The parametrisations link reflectance simulations to measurable forest variables, but at the same time rely on abstractions (e.g. a geometric surface forming a tree crown) and physically-based simplifications that are difficult to quantify robustly. As high-quality data on basic forest structure (e.g. tree height and stand density) and optical properties (e.g. leaf and forest floor reflectance) are becoming increasingly available, we used the well-validated forest reflectance and transmittance model FRT to investigate the effect of the values of the “uncertain” input parameters on the accuracy of modelled forest reflectance. With the state-of-the-art structural and spectral forest information, and Sentinel-2 Multispectral Instrument imagery, we identified that the input parameters influencing the most the modelled reflectance, given that the basic forestry variables are set to their true values and leaf mass is determined from reliable allometric models, are the regularity of the tree distribution and the amount of woody elements. When these parameters were set to their new adjusted values, the model performance improved considerably, reaching in the near infrared spectral region (740–950 nm) nearly zero bias, a relative RMSE of 13% and a correlation coefficient of 0.81. In the visible part of the spectrum, the model performance was not as consistent indicating room for improvement.