Comparing Object-Based and Pixel-Based Machine Learning Models for Tree-Cutting Detection with PlanetScope Satellite Images: Exploring Model Generalization

Abstract

Despite utilizing various remote sensing datasets, precise tree-cutting detection remains challenging due to spatial and spectral resolution constraints in satellite imagery, complex landscapes, data integration issues, and the need for accurate multi-temporal reference datasets. This study investigates the utilization of PlanetScope (PS) satellite images, along with pixel-based (PBIA) and object-based (OBIA) image analysis, for accurate mapping of forest cover and detection of tree cuttings. Detailed multi-temporal reference datasets were collected based on airborne laser scanning (ALS)-derived canopy height models (CHM) and very high-resolution (VHR) aerial orthomosaics. Reference datasets were used to train three machine learning (ML) models: random forest (RF), support vector machine (SVM), and feed-forward neural network (Nnet) in two forest districts located in Western and Northern Poland. The study also assessed the generalization capabilities of the best model in both local and temporal contexts. Regarding forest cover mapping, the OBIA RF classifier outperformed all other models with an overall accuracy (OA) of 99.27 % and Kappa of 98.18 %, while the PBIA SVM model showed the lowest (OA = 97.18 %, Kappa = 94.35 %). The testing of the model's generalization confirmed the performance of the OBIA RF model, with the Dice Coefficient ranging from 95.86 % to 96.74 %. The methodology's effectiveness in tree-cutting detection was demonstrated, with the detection rate ranging from 96.20 % to 99.39 % for the total number of cuttings, and from 99.45 % to 99.86 % for the total volume. In conclusion, the integration of PS satellite images, spectral-textural features, and generalized ML models proves to be effective for tree-cutting detection.