The assessment of public welfare forests quality and the formulation of differentiated ecological compensation policies have long relied on ground-based sampling surveys. However, such surveys are unable to objectively and timely capture the dynamic changes in forest quality over large scales, leading to a risk of mismatch between compensation types delineated from static data and actual ecological conditions. To address this issue, a methodology was developed using multi-source remote sensing data acquired before the implementation of a differentiated ecological compensation policy in Guangdong Province, China (2024), which classifies public welfare forests into four types: basic and incentive compensations under general zones and special zones, respectively. A Quality Rating Index (QRI) for public welfare forests was constructed to assess the ecological baseline of the year 2023. With the aid of prior knowledge (i.e., public welfare forests boundaries and compensation labels), the continuous QRI values were converted into discrete compensation types, and a spatial matching analysis was conducted between the QRI-based types and the policy‑defined types to identify mismatched areas. Furthermore, a Fully Connected Neural Network (FCNN) model was designed to explore an intelligent identification approach that can replicate policy classification rules based solely on continuous remote sensing features, without relying on prior policy boundaries. The results show that: ① the QRI-based evaluation reveals that the quality of public welfare forests in Guangdong was predominantly moderate (accounting for 73% of the area), with significant spatial heterogeneity. ② the overall spatial matching rate between QRI results and policy-defined compensation types was 82%. Within the “basic compensation in special zones” category, about 20% of the area was mapped as “incentive compensation in general zones”, indicating a partial disconnect between policy zoning and actual forest quality. ③ driven by eight remote sensing indicators (stand age, canopy density, stock volume, biomass, carbon sequestration, tree height, slope, and aspect), the FCNN model achieved a reproduction accuracy exceeding 90% for the general‑zone compensation types (both basic and incentive), demonstrating that ecologically dominated classification rules can be automatically learned. ④ the model’s reproduction accuracy for the policy‑driven “incentive compensation in special zones” was relatively low (60%~78%), revealing the limitation of purely eco‑feature‑driven models. The proposed framework provides a methodological reference for large-scale dynamic assessment of public welfare forests quality, rationality verification of differentiated compensation policies, and intelligent recognition of compensation types.