In machine learning, one often assumes the data are independent when evaluating model performance. However, this rarely holds in practice. Geographic information datasets are an example where the data points have stronger dependencies among each other the closer they are geographically. This phenomenon known as spatial autocorrelation (SAC) causes the standard cross validation (CV) methods to produce optimistically biased prediction performance estimates for spatial models, which can result in increased costs and accidents in practical applications. To overcome this problem, we propose a modified version of the CV method called spatial k-fold cross validation (SKCV), which provides a useful estimate for model prediction performance without optimistic bias due to SAC. We test SKCV with three real-world cases involving open natural data showing that the estimates produced by the ordinary CV are up to 40% more optimistic than those of SKCV. Both regression and classification cases are considered in our experiments. In addition, we will show how the SKCV method can be applied as a criterion for selecting data sampling density for new research area.

Digital maps of forest resources are a crucial factor in successful forestry applications. Since manual measurement of this data on large areas is infeasible, maps must be constructed using a sample field data set and a prediction model constructed from remote sensing materials, of which airborne laser scanning (ALS) data and aerial images are currently widely used in management planning inventories. ALS data is suitable for the prediction of variables related to the size and volume of trees, whereas optical imagery helps in improving distinction between tree species. We studied the prediction of forest attributes using field data from National Forest Inventory complemented with ad hoc field plots in combination with ALS and aerial imagery data in Aland province, Finland. We applied feature selection with genetic algorithm and greedy forward selection and compared multiple linear and nonlinear estimators. Maximally around 40 features from a total of 154 were required to achieve the best prediction performances. Tree height was predicted with normalized root mean squared error value of 0.1 and tree volume with a value around 0.25. Predicting the volumes of spruce and broadleaved trees was the most challenging due to small proportions of these tree species in the study area.

In large-area forest inventories, a trade-off between the amount of data to be sampled and the corresponding collection costs is necessary. It is not always possible to have a very large data sample when dealing with sampling-based inventories. It is therefore important to optimize the sampling design with the limited resources. Whereas this sort of inventories are subject to these constraints, the availability of remote sensing (RS) data correlated with the forest inventory variables is usually much higher. For this reason, the RS and sampled field measurement data are often used in combination for improving the forest inventory estimation. In this study, we propose a model-based data sampling method founded on Bayesian optimization and machine learning algorithms which utilizes RS data to guide forest inventory sample selection. We evaluate our method in empirical experiments using real-world volume of growing stock data from the Aland region in Finland. The proposed method is compared against two baseline methods: simple random sampling and the local pivotal method. When a suitable model link is selected, the empirical experiments show on best case on average up to 22% and 79% improvement in population mean and variance estimation respectively over baselines. However, the results also illustrate the importance of model selection which has a clear effect on the results. The novelty of the study is in the application of Bayesian optimization in national forest inventory survey sampling.

The age of forest stand's growing stock is a vital variable for the planning of forest management and conservation. Age is rarely known in boreal forest without increment coring. However, age can be predicted by linking the age observations of the National Forest Inventory (NFI) plots to remote sensing data and other auxiliary data. We compare tabular foundation model (TabFM), convolutional neural network (U-Net), tree ensemble (GPBoost and RandomForest) and k-nearest neighbor models for the prediction of forest age at the national scale in Finland. Remote sensing data included photogrammetric point clouds derived from aerial images and Sentinel‐2 imagery, which were complemented with spatial and categorical variables describing, for example, temperature sum, plot coordinates and site type. When all predictor variable sets were included, root-mean-square errors (RMSEs) associated with forest age prediction ranged from 23 to 26 years depending on the modeling approach in separate test data (n = 5,571). TabFM achieved the smallest RMSEs, followed by GPBoost, random forest and convolutional neural network. Incorporating spatial and categorical predictor variables reduced RMSE by 13%, on average, compared with that of the models relying solely on remote sensing data. Spatial predictor variables were important for localizing models, which reduced systematic errors, i.e., mean differences (MDs). The convolutional neural network exhibited the largest MDs ranging from 2.3 to 6.9 years, whereas the other approaches produced MDs deviating from zero by less than one year. Although the inclusion of auxiliary variables positively affected the predictive performance of the forest age models, predicting forest age in forests older than 100 years remains challenging due to saturation of canopy height growth.