@article{fdi:010087663,
  title = {{M}achine-learning-based downscaling of hourly {ERA}5-land air temperature over mountainous regions},
  author = {{S}ebbar, {B}. {E}. and {K}habba, {S}. and {M}erlin, {O}. and {S}imonneaux, {V}incent and {E}l {H}achimi, {C}. and {K}harrou, {M}. {H}. and {C}hehbouni, {A}bdelghani},
  editor = {},
  language = {{ENG}},
  abstract = {{I}n mountainous regions, the scarcity of air temperature ({T}a) measurements is a major limitation for hydrological and crop monitoring. {A}n alternative to in situ measurements could be to downscale the reanalysis {T}a data provided at high-temporal resolution. {H}owever, the relatively coarse spatial resolution of these products (i.e., 9 km for {ERA}5-{L}and) is unlikely to be directly representative of actual local {T}a patterns. {T}o address this issue, this study presents a new spatial downscaling strategy of hourly {ERA}5-{L}and {T}a data with a three-step procedure. {F}irst, the 9 km resolution {ERA}5 {T}a is corrected at its original resolution by using a reference {T}a derived from the elevation of the 9 km resolution grid and an in situ estimate over the area of the hourly {E}nvironmental {L}apse {R}ate ({ELR}). {S}uch a correction of 9 km resolution {ERA}5 {T}a is trained using several machine learning techniques, including {M}ultiple {L}inear {R}egression ({MLR}), {S}upport {V}ector {R}egression ({SVR}), and {E}xtreme {G}radient {B}oosting ({X}gboost), as well as ancillary {ERA}5 data (daily mean, standard deviation, hourly {ELR}, and grid elevation). {N}ext, the trained correction algorithms are run to correct 9 km resolution {ERA}5 {T}a, and the corrected {ERA}5 {T}a data are used to derive an updated {ELR} over the area (without using in situ {T}a measurements). {T}hird, the updated hourly {ELR} is used to disaggregate 9 km resolution corrected {ERA}5 {T}a data at the 30-meter resolution of {SRTM}'s {D}igital {E}levation {M}odel ({DEM}). {T}he effectiveness of this method is assessed across the northern part of the {H}igh {A}tlas {M}ountains in central {M}orocco through (1) k-fold cross-validation against five years (2016 to 2020) of in situ hourly temperature readings and (2) comparison with classical downscaling methods based on a constant {ELR}. {O}ur results indicate a significant enhancement in the spatial distribution of hourly local {T}a. {B}y comparing our model, which included {X}gboost, {SVR}, and {MLR}, with the constant {ELR}-based downscaling approach, we were able to decrease the regional root mean square error from approximately 3 ? to 1.61 ?, 1.75 ?, and 1.8 ?, reduce the mean bias error from -0.5 ? to null, and increase the coefficient of determination from 0.88 to 0.97, 0.96, and 0.96 for {X}gboost, {SVR}, and {MLR}, respectively.},
  keywords = {reanalysis ; {ERA}5-{L}and ; air temperature ; downscaling ; complex terrain ; machine learning},
  booktitle = {},
  journal = {{A}tmosphere},
  volume = {14},
  numero = {4},
  pages = {610 [21 ]},
  year = {2023},
  DOI = {10.3390/atmos14040610},
  URL = {https://www.documentation.ird.fr/hor/fdi:010087663},
}