chapter_4_conclusions.qmd

# Conclusions and Future Work

## Synthesis

This research evaluated the advantages and limitations associated with
estimating GPP using traditional VIs like NDVI, EVI, and recent ones like CCI
and NIRv. Specifically, the study focused on VIs derived from MODIS, chosen for
its extensive time records. To assess GPP estimates through the VIs, we employed
traditional methods such as linear regression and GAM, as well as ML techniques.

Traditional methods, such as LM and GAM, offer the advantage of being more
straightforward to interpret in relation to their predictions and associated
errors. In contrast, ML techniques like regression RF and autoML pursue
predictions through methods that are harder to elucidate, a phenomenon referred
to as "black model boxes." Despite their complexity, these ML methods have the
potential to achieve superior accuracy in their predictions.

To evaluate the capacity of VIs to estimate GPP,Productivity (GPP), our analysis
spanned three temporal aggregations: daily, weekly, and monthly. This assessment
considered the utilization of VIs individually or collectively as covariates in
the model. For ML methods, the same temporal aggregations were employed, with an
assessment involving all indices and the entire set of MODIS bands as
predictors.

The results indicate that, across all models, monthly temporal aggregation
consistently yields better outcomes in comparison to weekly and daily
aggregations. This superiority can be attributed to the reduced variation
present in monthly aggregated data, resulting in lower errors in the models.
According to LM and GAM,  incorporating all indices as covariates in the models
enhances predictive capability, suggesting that a single VI in isolation may
fall short of capturing the entire variation in the data.

Examining the outcomes of models utilizing individual VIs reveals minimal
differences. Nonetheless, VIs consistently displaying a slightly superior
performance in terms of explained variation and lower error are typically EVI
and CCI, whereas NDVI tends to exhibit lower performance. This trend is also
evident in ML models, with EVI and CCI frequently occupying top positions in
variable of importance rankings. In ML models, the inclusion of all bands as
predictors highlights that B02 (NIR) often ranks among the top variables in
importance over other bands.

In comparing the various models and their configurations (LM and GAM with VIs as
covariates; RF and autoML with VIs and complete bands as predictors), it becomes
evident that ML methods exhibit a better performance across different time
aggregations. For daily models, assessed by R² and RMSE metrics, RF surpassed
autoML, GAM, and LM, in that sequence. For the weekly models, autoML led,
followed by GAM, LM, and RF. Regarding the monthly models, RF had better
performance, followed by LM, autoML, and GAM.

## Limitations

The uncertainties in VIs, in this case, for GPP estimation, may be due to their
inability to capture variations in GPP or inherent limitations falling into two
categories:  artifacts or external factors. Artifacts originate from the sensors
used for deriving VIs, encompassing calibration, quality control, or sensor
degradation [@zeng_optical_2022]. In this study, we addressed these effects by
excluding pixels with suboptimal quality using data from quality assurance (QA)
and quality control (QC). These variables offer insights into the quality of
each pixel, aiding in the selection of pixels potentially harboring erroneous
values. However, the accuracy of these data may be compromised by the presence
of pixels with poor quality not identified as such by postprocessing algorithms,
or influenced by external factors such as clouds, cloud shadows, inaccuracies in
snow detection, atmospheric pollution, or technical issues in image acquisition
[@pesquer_spatial_2019].

Another potential source of uncertainty in GPP estimates is related to the
representation of the flux tower footprint within the pixel area. EC towers
exhibit a flux footprint that can vary over time. The footprint is defined as
the extent to which measurements taken at a specific time and location
accurately reflect the actual flows in time and specific area. The monthly
climatologies of the footprint can exhibit variations, usually falling within
the range of 100 m to 450 m. [@chu_representativeness_2021]. The heterogeneity of
the landscape beyond this footprint may affect flow conditions differently than
those measured within the footprint. In our methodology, we established a square
area of 3 km² around the EC tower to ensure that the flux footprint climatology
aligns with this region, considering the homogeneity in land cover type.
However, the resolution of MODIS might not be adequate for capturing nuanced
fine-scale spatial variations [@robinson_terrestrial_2018].

The limitations of the study include the unavailability of data from a sensor
like Sentinel-2, which offers better spatial resolution compared to the MODIS
product employed. This data could have allowed us to explore whether, despite
estimating in areas with homogeneous land cover, it is feasible to capture
nuanced fine-scale spatial variations. Such an exploration could potentially
lead to improved GPP estimates, thereby reducing uncertainty. The restriction in
using Sentinel-2 data stems from the fact that calibrated data is accessible
only from 2017 onward. However, the GPP data processed by ONEFlux for each site
commenced in 2015, concluding for Bartlett and Michigan in 2017 and 2018,
respectively, thus resulting in the exclusion of a significant portion of the
data.

Finally, it is important to acknowledge that the sites selection can constitute
a potential limitation in this study. The focus on deciduous broadleaf forest
sites in the northern hemisphere, encompassing only three locations, restricts
the generalization of the findings to this specific ecosystem type. Expanding
the scope to include more sites representing other various ecosystem types could
further contribute to understanding which indices have limitations or can
perform better in estimating GPP across a broader range of scenarios.

## Future work 

As uncertainties persist in GPP estimation models relying on VIs as predictors,
the potential for improvement lies in models that can consider critical stress
factors influencing photosynthesis [@rogers_response_2020; @ryu_what_2019;
@xiao_remote_2019]. While incorporating meteorological data can address this, it
poses challenges in various regions around the globe. In such circumstances, a
more advantageous approach involves the utilization and validation of VIs that
demonstrate sensitivity to stress factors affecting photosynthetic activity, not
just photosynthetic capacity. Remote sensing provides a means to achieve this by
integrating data products such as thermal information, which can indirectly
address water deficit in ecosystems, for example [@pabon-moreno_potential_2022].
Despite potential uncertainties associated with these products, there is a
prospect for enhanced estimations by incorporating considerations beyond
traditional reflectance values.

Addressing the inherent variability across ecosystems, it is essential to
recognize that no single VI will be the best, and their representation of
vegetation functioning has limitations [@zeng_optical_2022].  Nonetheless, the
continual introduction of new VIs derived from both traditional and advanced
sensors [@montero_standardized_2023] presents an opportunity to enhance the
estimation of GPP on a global level. A potentially effective strategy would
involve establishing a standardized guide of VIs or combinations tailored to
different ecosystem types, providing researchers with a reliable foundation for
accurate assessments. Furthermore, the field could benefit from adopting a
qualitative rating of model performance, such as the one used by for the FAO
model AquaCrop [@raes_aquacrop_fao]. This practice involves standardizing model
evaluations based on their performance within specific research domains. Such
standardized benchmarks can provide valuable guidance for researchers seeking
optimal model selection and interpretation within their respective contexts.

Given the growing reliance on ML methods to enhance GPP predictions in recent
years, a notable challenge arises from their tendency to obscure the
understanding of the underlying mechanisms explaining the relationship between
VIs and GPP [@molnar2020interpretable]. Hence, it becomes crucial for new
studies to articulate their research objectives with clarity and consistency.
Specifically, researchers should explicitly state whether their primary focus is
to explore mechanistic relationships between VIs or if their aim is to improve
GPP predictions. This distinction is vital, whether the goal is to predict GPP
in sites lacking validation data (flux towers) or to advance forecasting
capabilities [@meyer_importance_2019].

The implementation of explainable ML methods could enhance our understanding of
the variables influencing predictions [@molnar2020interpretable]. Additionally,
it is well-known that a prediction model's ability to generalize the rules it
has learned from the training data set to a new (unseen) data set is poor.
Diverse scenarios, such as droughts, intense rainfall, and vegetation mortality,
among others, can affect predictive capacity. Therefore, consistent model
predictive power assessment over time and retraining as necessary are crucial
[@kuhn2022tidy]. This underscores the significance of having flux towers and
extending their presence to underrepresented sites to acquire validation data.
These efforts contribute to refining future global GPP estimation models on a
broader scale [@meyer_importance_2019].

\newpage

## References

<!-- ::: {#refs} -->
<!-- ::: -->