Monitoring Forest Interventions for Supporting Regulatory Compliance Using PlanetScope Time Series and High-Resolution Satellite Imagery: A Case Study of Coppice Forests in Southern Italy

In the context of sustainable forest management, there is a growing need for scalable monitoring systems capable of detecting forest interventions, verifying post-harvest compliance, and supporting traceability throughout forest-wood supply chains. To address these needs, this study proposes Satellite-based Traceability and Environmental Monitoring (STEM), a remote sensing framework that integrates multi-temporal PlanetScope imagery, change detection, functional time-series representation learning, machine learning classification, and high-resolution imagery for compliance assessment within a unified workflow. The STEM framework follows a sequential workflow encompassing intervention detection, classification, and post-harvest verification. First, forest interventions are detected from vegetation-index time series by learning normal vegetation dynamics (i.e., NDVI/GNDVI) in reference areas using the deep learning-based forecasting model NeuralProphet. Persistent deviations between predicted and observed trajectories are then identified as indicators of forest disturbance. To classify the detected interventions under coppice forest conditions, STEM evaluates multiple temporal representations, including local descriptors, seasonal summaries, Functional Principal Component Analysis (FPCA), and Multivariate Functional Principal Component Analysis (MFPCA). Finally, the detected disturbed areas are further analyzed using the Retained-Tree Detection (RTD) module, which automatically identifies and characterizes retained seed trees from high-resolution satellite imagery and terrain data, supporting post-harvest compliance verification. Experimental evaluation across five coppice forest sites in Basilicata, Southern Italy, shows that interventions involving extensive biomass removal generate clear and distinguishable temporal signatures, whereas weaker and spatially heterogeneous disturbances remain more challenging to detect. The RTD module achieved an overall F1-score of 95.32%, with a precision of 92.34% and a recall of 98.49% within the sampled validation tiles, demonstrating accurate detection of retained seed trees within harvested stands. By integrating large-scale disturbance detection with site-level compliance verification, STEM provides a scalable and operational framework for supporting sustainable forest management, forest traceability, and regulatory compliance.