The authors develop a deep learning model to generate the new version of global 250 m spatio-temporal continuous GLASS FAPAR product. This product has great application potential as it is the first 250-m product and has the highest validation accuracy among existing FAPAR products according the evaluation results.

The paper gives details of the model, data, and product validation. Although the authors have done a good job, some modifications are needed to further their manuscript. Please see below for my specific comments:

1. Reference for Ameriflux should be given.
2. L222: In 2000, MODIS reflectance has no data before DOY49, did it affect the estimates in 2000? If added other 500-m bands, how the model accuracy changes?
3. In Fig. 4, the meaning of P values should be explained.
4. In Fig.6, “Direct validation of (the a) MODIS” should be “Direct validation of the a) MODIS”.
5. How does the author aggregate the 250-m FAPAR to 500-m and 3-km? If using all MODIS surface reflectance bands to estimate 500-m FAPAR, can the accuracy be iimproved?
6. From Fig. 9, the GLASS V6 FAPAR was smoothed, did the author reprocess the estimated FAPAR curve, or the training samples are inherently smooth? How to reprocess them?
7. In Fig. 7, please check the number of the sub figures. The range of legend should be 0-1.
8. In Fig. 9, note the legend of field FAPAR (two triangles)
9. 9: Do these four FAPAR products have the same definitions?
10. L210: What is the resolution of input GLASS V6 LAI and output FAPAR? The input MODIS surface reflectance is 250 m, the output FAPAR should be also 250 m. How to get the 250-m FAPAR samples?
11. The GLASS V6 LAI product has uncertainty, the error will be brought into the estimation model. If only MODIS reflectance is used for training data, this error can be ignored, the accuracy of fAPAR maybe improved.

The authors presented a method to generate a global FAPAR product at 250 m by exploiting MODIS surface reflectance data and deep learning model. The generated product is extremely interesting and useful. The validation analysis demonstrates the quality of the obtained product.

However, I have some comments to make to the authors:

1. The reasoning for choosing the bi-LSTM method is not clear. Currently, deep learning models with transformer have outperformed LSTMs when working with both short and long time series of data.

2. It would be interesting to have more images like Figure 8 to really see (also qualitatively) the product obtained compared to the existing one.

3. The authors stated that the obtained product has better spatiotemporal consistency. It would be nice to have figures showing how the FAPAR changes spatially for sequential instances of the time series (e.g., to see the FAPAR generated at time t1, t2, t3 and so on).

4. What is the computational burden required to generate a new FAPAR product on a global scale once new MODIS surface reflectance data are available?