This study examined the effect of atmospheric, topographic, and Bidirectional Reflectance Distribution Function (BRDF) corrections of Sentinel-2 images implemented in Google Earth Engine (GEE) for use in land cover classification. The study was carried out in an area of complex orography in northern Spain and made use of the Spanish National Forest Inventory plots and other systematically located plots to cover non-forest classes. A total of 2991 photo-interpreted ground plots and 15 Sentinel-2 images, acquired in summer at a spatial resolution of 10–20 m per pixel, were used for this purpose. The overall goal was to determine the optimal level of image correction in GEE for subsequent use in time series analysis of images for accurate forest cover classification. Particular attention was given to the classification of cover by the major commercial forest species: Eucalyptus globulus, Eucalyptus nitens, Pinus pinaster, and Pinus radiata. The Second Simulation of the Satellite Signal in the Solar Spectrum (Py6S) algorithm, used for atmospheric correction, provided the best compromise between execution time and image size, in comparison with other algorithms such as Sentinel-2 Level 2A Processor (Sen2Cor) and Sensor Invariant Atmospheric Correction (SIAC). To correct the topographic effect, we tested the modified Sun-canopy-sensor topographic correction (SCS + C) algorithm with digital elevation models (DEMs) of three different spatial resolutions (90, 30, and 10 m per pixel). The combination of Py6S, the SCS + C algorithm and the high-spatial resolution DEM (10 m per pixel) yielded the greatest precision, which demonstrated the need to match the pixel size of the image and the spatial resolution of the DEM used for topographic correction. We used the Ross-Thick/Li-Sparse-Reciprocal BRDF to correct the variation in reflectivity captured by the sensor. The BRDF corrections did not significantly improve the accuracy of the land cover classification with the Sentinel-2 images acquired in summer; however, we retained this correction for subsequent time series analysis of the images, as we expected it to be of much greater importance in images with larger solar incidence angles. Our final proposed dataset, with image correction for atmospheric (Py6S), topographic (SCS + C), and BRDF (Ross-Thick/Li-Sparse-Reciprocal BRDF) effects and a DEM of spatial resolution 10 m per pixel, yielded better goodness-of-fit statistics than other datasets available in the GEE catalogue. The Sentinel-2 images currently available in GEE are therefore not the most accurate for constructing land cover classification maps in areas with complex orography, such as northern Spain.
Accurate forest monitoring is critical for achieving the objectives of the European Green Deal. While national forest inventories provide consistent information on the state of forests, their temporal frequency is inadequate for monitoring fast-growing species with 15-year rotations when inventories are conducted every 10 years. However, Earth observation (EO) satellite systems can be used to address this challenge. Remote sensing satellites enable the continuous acquisition of land cover data with high temporal frequency (annually or shorter), at a spatial resolution of 10-30 m per pixel. This study focused on northern Spain, a highly productive forest region. This study aimed to improve models for predicting forest variables in forest plantations in northern Spain by integrating optical (Sentinel-2) and imaging radar (Sentinel-1, ALOS-2 PALSAR-2 and TanDEM-X) datasets supported by climatic and terrain variables. Five popular machine learning algorithms were compared, namely kNN, LightGBM, Random Forest, MLR, and XGBoost. The study findings show an improvement in R2 from 0.24 when only Sentinel-2 data are used with MultiLinear Regression to 0.49 when XGboost is used with multi-source EO data. It can be concluded that the combination of multi-source datasets, regardless of the model used, significantly enhances model performance, with TanDEM-X data standing out for their remarkable ability to provide valuable radar information on forest height and volume, particularly in a complex terrain such as northern Spain.
In this study, we used Spanish National Forest Inventory (SNFI) data, Sentinel-2 imagery and ancillary data to develop models that estimate forest variables for major commercial timber plantations in northern Spain. We carried out the analysis in two stages. In the first stage, we considered plots with and without sub-meter geolocation, three pre-processing levels for the Sentinel-2 images and two machine learning algorithms. In most cases, geometrically, radiometrically, atmospherically and topographically (L2A-ATC) corrected images and the random forest algorithm provided the best results, with topographic correction producing a greater gain in model accuracy as the average slope of the plots increased. Our results did not show any clear impact of the geolocation accuracy of SNFI plots on results, suggesting that the usual geolocation accuracy of SNFI plots is adequate for developing forest models with data obtained from passive sensors. In the second stage, we used all plots together with L2A-ATC-corrected images to select five different groups of predictor variables in a cumulative process to determine the influence of each group of variables in the final RF model predictions. Yield variables produced the best fits, with R2 ranging from 0.39 to 0.46 (RMSE% ranged from 44.6% to 61.9%). Although the Sentinel-2-based estimates obtained in this research are less precise than those previously obtained with Airborne Laser Scanning (ALS) data for the same species and region, they are unbiased (Bias% was always below 1%). Therefore, accurate estimates for one hectare are expected, as they are obtained by averaging the values of 100 pixels (model resolution of 10 m pixel−1) with an expected error compensation. Moreover, the use of these models will overcome the temporal resolution problem associated with the previous ALS-based models and will enable annual updates of forest timber resource estimates to be obtained.
28 September 2023 in Cuadernos de la Sociedad Española de Ciencias Forestales
El objetivo de este trabajo fue realizar una clasificación automática de coberturas forestales con una resolución espacial de 10 m/pixel en Asturias (Norte de España). Esta región tiene una vocación eminentemente forestal con una gran superficie de plantaciones forestales de Pinus pinaster, Pinus radiata y Eucalyptus globulus, cuyas cortas representan el 91% de la madera aprovechada anualmente en la región. La clasificación se basó en una serie de datos multitemporales del sensor Sentinel-2 y en datos auxiliares climáticos y topográficos. Como fuentes de datos se utilizaron: parcelas de campo (Inventario Forestal Nacional), clasificación de coberturas terrestres (Corine Land Cover), imágenes de satélite (constelación Sentinel-2) y varias fuentes de variables auxiliares. El código fue implementado en Google Earth Engine® (GEE) y para la clasificación se usó el algoritmo Random Forest. Se realizó una división aleatoria de la base de datos, destinando un 70% para el entrenamiento de los modelos y un 30% para su validación. Finalmente se clasificaron las coberturas forestales del territorio en 30 clases, obteniéndose un valor del estadístico Kappa de 0,5723. Al analizar la matriz de confusión, se observó que los fallos de clasificación eran principalmente consecuencia de la similitud de las firmas espectrales de las coberturas analizadas. En los próximos años, con la mayor diversidad y mejora de los datos capturados por sensores remotos, el desarrollo de plataformas como GEE, y el desarrollo de nuevos algoritmos, permitirán obtener clasificaciones de cobertura automáticas más precisas y probablemente se convertirán en herramientas indispensables.
Aim of study: In this study we developed machine learning models for estimating five forest stand variables, including total over bark volume, number of trees per hectare, dominant height, basal area and aboveground biomass, from remote sensing data and ancillary variables in forest plantations of Eucalyptus globulus Labill, Pinus pinaster Aiton and Pinus radiata D. Don. Area of study: The study was conducted in northern Spain (Autonomous Communities of Asturias, Cantabria, Basque Country and Galicia) where plantations of the species of interest are mainly concentrated due to the high productivity of forest stands in these areas. Material and methods: We used ground-truth data from Spanish National Forest Inventory plots, two sources of remote sensing data and images from the Sentinel-1 constellation (radar) and from the Sentinel-2 constellation (optical) combined with terrain and climate data. We used the Google Earth Engine platform using Random Forest algorithm with 10-fold cross-validation to obtain forest stand variables, including density, size and yield variables. Main results: Evaluation of the model accuracy and variable importance in estimating forest variables showed that the total volume models performed best, yielding R² values between 0.39 and 0.45 for the different tree species. Optical bands were very important in all cases, while radar bands were less important. Research highlights: The use of synthetic aperture radar, although promising, is limited in platforms like GEE, especially for complex terrains where shadowing and angles affect the image quality. The research contributions include the automation of preprocessing steps, and the findings highlight the need to develop more robust, adaptable models using AI and new radar sensors.