Inventory of a production forest using MSI/Sentinel-2 imagery and digital aerial photogrammetry
Name: RACHEL CLEMENTE CARVALHO
Publication date: 30/04/2021
Advisor:
Name |
Role |
|---|---|
| ADRIANO RIBEIRO DE MENDONÇA | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| ADRIANO RIBEIRO DE MENDONÇA | Advisor * |
| ANDRÉ QUINTÃO DA ALMEIDA | Co advisor * |
| GILSON FERNANDES DA SILVA | Internal Examiner * |
Summary: In the forestry sector, knowledge of forest productivity is obtained through forest inventories. However, the sampling techniques traditionally applied to forest inventories have a high demand for time and high cost of execution. Therefore, it is necessary to evaluate the use of alternative techniques to obtain this data, such as remote sensing applications. In this context, the use of remote sensing allows the acquisition of data in large areas quickly and at a reduced cost. This work had as main objective to estimate attributes of forest interest of a commercial planting of eucalyptus by orbital image (IO) and digital aerial photogrammetry (DAP) and to compare with the
results obtained by the traditional forest inventory. As a secondary objective, an evaluation of DAP products was carried out based on planting attributes collected in the field. For the inventory based on the IO, spectral bands of an image from the MSI/Sentinel-2 sensor were selected and various vegetation indices were calculated. The individual bands and the vegetation indexes were used as predictive variables for the modeling. To obtain the DAP data, a flight was performed with an unmanned aerial vehicle (UAV) for the generation of a three-dimensional point cloud by the SfM algorithm and also a digital terrain model (DTM) for its normalization. The quality of FAD's DTM was evaluated by comparing the values of dominant height of each plot
with the metrics representative of the maximum height of the normalized point cloud. Traditional height-based metrics were extracted for each plot, which were used as predictor variables. Multiple linear regression (MLR) and artificial neural networks (ANN) performed the basal area (G) and volume (V) estimation process. For the modeling, three data sources were considered, IO, DAP and the combination of IO and DAP. At the plot level, to estimate the G from IO and DAP data, the lowest values of RMSE in the validation occurred in the ANN modeling, being 13.22% and 13.36%,
respectively. For the combination of IO and DAP, the MLR presented a lower RMSE in the validation (RMSE = 12.46%). The same happened for V, with the lowest values of RMSE in the validation with data from IO (15.05%) and DAP (16.58%), obtained inANN modeling, for the combination of IO and DAP, the lower RMSE was obtained by MLR (14.14%). When performing the modeling for the entire area, it was possible to observe that the ANN presented greater capacity for generalization, with results closer to those obtained in the traditional forest inventory for all data sources. All the averages of G and V were close to the values obtained in the inventory, with a maximum of 3.2% difference. As in the plot level results, the combination of IO and DAP generated more accurate results for the whole area, with a difference of 0.3% for G and 0.4% for V, in relation to the inventory. The results obtained in this study indicate that IO and DAP data can be used for the inventory of G and V in eucalyptus plantations, with results compatible with those obtained in the traditional forest inventory.
Keywords: Forest inventory; artificial neural networks; 3D remote sensing; Structure from Motion; Forest 4.0.
