RainForest: A random forest algorithm for quantitative precipitation estimation over Swizerland

Wolfensberger, Daniel; Gabella, Marco; Boscacci, Marco; Germann, Urs; Berne, Alexis

doi:10.5194/amt-2020-284

Cited by 3 publications

(3 citation statements)

References 37 publications

(45 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It seems likely that producing biased extreme values is an inherited behavior of the RF model. This is because the final prediction value is obtained by the averaging of the prediction of each tree, which could reduce the value range due to the smoothing effect of averaging [87,88]. Therefore, post processing, which entails the rescaling of the downscaled SM to in situ SM observations using approaches such as linear regression or cumulative distribution function (CDF) matching, could correct the observed bias of extreme SM values if a in situ measurements network was available.…”

Section: Advantages and Limitations Of This Studymentioning

confidence: 99%

Enhancing Spatial Resolution of SMAP Soil Moisture Products through Spatial Downscaling over a Large Watershed: A Case Study for the Susquehanna River Basin in the Northeastern United States

Wakigari

Leconte

2022

Remote Sensing

View full text Add to dashboard Cite

Soil moisture (SM) with a high spatial resolution plays a paramount role in many local and regional hydrological and agricultural applications. The advent of L-band passive microwave satellites allowed for it to be possible to measure near-surface SM at a global scale compared to in situ measurements. However, their use is often limited because of their coarse spatial resolution. Aiming to address this limitation, random forest (RF) models are adopted to downscale the SMAP level-3 (L3SMP, 36 km) and SMAP enhanced (L3SMP_E, 9 km) SM to 1 km. A suite of predictors derived from the Sentinel-1 C-band SAR and MODIS is used in the downscaling process. The RF models are separately trained and verified at both spatial scales (i.e., 36 and 9 km) considering two experiments: 1) using predictors derived from the MODIS and Sentinel-1 along with other predictors such as elevation and brightness temperature and 2) using all predictors of the first experiment except for the Sentinel-1 predictors. Only dates when the Sentinel-1 images were available are considered for the comparison of the two experiments. The comparison of the results of the two experiments indicates that the removal of Sentinel-1 predictors from the second experiment only reduces the R value from 0.84 to 0.83 and from 0.91 to 0.86 for 36 and 9 km spatial scales, respectively. Among the predictors used in the downscaling, the brightness temperature in VV polarization is identified as the most important predictor, followed by NDVI, surface albedo and API. On the contrary, the Sentinel-1 predictors play a less important role with no marked contribution in enhancing the predictive accuracy of RF models. In general, the two experiments have limitation, such as a small sample size for the training of the RF model because of the scarcity of Sentinel-1 images (i.e., revisit time of 12 days). Therefore, based on this limitation, a third experiment is proposed, in which the Sentinel-1 predictors are not considered at all in the training of the RF models. The results of the third experiment show a good agreement between the downscaled L3SMP and L3SMP_E SM, and in situ SM measurements at both spatial scales. In addition, the temporal availability of the downscaled SM increased. Moreover, the downscaled SM from both SMAP products presented greater spatial detail while preserving the spatial patterns found in their original products. The use of the two SMAP SM products as background fields for the downscaling process does not show marked differences. Overall, this study demonstrates encouraging results in the downscaling of SMAP SM products over humid climate with warm summers dominated by vegetation.

show abstract

Section: Advantages and Limitations Of This Studymentioning

confidence: 99%

Enhancing Spatial Resolution of SMAP Soil Moisture Products through Spatial Downscaling over a Large Watershed: A Case Study for the Susquehanna River Basin in the Northeastern United States

Wakigari

Leconte

2022

Remote Sensing

View full text Add to dashboard Cite

show abstract

“…As shown in Figure 5, RF has underestimated large values and overestimated small values. This is due to its own principles, RF tends to intermediate predicted values, the extreme observations are estimated using averages of response values that are closer to those observations (Zhang and Lu, 2012;Wolfensberger et al, 2021). The result of LR is similar to the GAM, which is significantly lower than that of the other three methods (Sun et al, 2015;Masinda et al, 2021;Yu et al, 2021).…”

Section: Model Evaluation and Comparisonmentioning

confidence: 96%

A comparison of five models in predicting surface dead fine fuel moisture content of typical forests in Northeast China

Fan

Ren

et al. 2023

Front. For. Glob. Change

View full text Add to dashboard Cite

IntroductionThe spread and development of wildfires are deeply affected by the fine fuel moisture content (FFMC), which is a key factor in fire risk assessment. At present, there are many new prediction methods based on machine learning, but few people pay attention to their comparison with traditional models, which leads to some limitations in the application of machine learning in predicting FFMC.MethodsTherefore, we made long-term field observations of surface dead FFMC by half-hour time steps of four typical forests in Northeast China, analyzed the dynamic change in FFMC and its driving factors. Five different prediction models were built, and their performances were compared.ResultsBy and large, our results showed that the semi-physical models (Nelson method, MAE from 0.566 to 1.332; Simard method, MAE from 0.457 to 1.250) perform best, the machine learning models (Random Forest model, MAE from 1.666 to 1.933; generalized additive model, MAE from 2.534 to 4.485) perform slightly worse, and the Linear regression model (MAE from 2.798 to 5.048) performs worst.DiscussionThe Simard method, Nelson method and Random Forest model showed great performance, their MAE and RMSE are almost all less than 2%. In addition, it also suggested that machine learning models can also accurately predict FFMC, and they have great potential because it can introduce new variables and data in future to continuously develop. This study provides a basis for the selection and development of FFMC prediction in the future.

show abstract

“…First, due to the features of the algorithm itself, RF averages the results for all regression trees. The underestimation of extreme values and overestimation of small values appears to be a common problem for RF regression models (C ̌eh et al, 2018;Zimmerman et al, 2018;Wolfensberger et al, 2021). Second, the training data sets contained very few extreme pCO 2 values and they were underrepresented in the RF model, thereby leading to a more mean-biased output from the RF model.…”

Section: Advantages and Limitations Of Rf Modelmentioning

confidence: 99%

Remote sensing and machine learning method to support sea surface pCO2 estimation in the Yellow Sea

Liu

Zhai

et al. 2023

Front. Mar. Sci.

View full text Add to dashboard Cite

With global climate changing, the carbon dioxide (CO2) absorption rates increased in marginal seas. Due to the limited availability of in-situ spatial and temporal distribution data, the current status of the sea surface carbon dioxide partial pressure (pCO2) in the Yellow Sea is unclear. Therefore, a pCO2 model based on a random forest algorithm has been developed, which was trained and tested using 14 cruise data sets from 2011 to 2019, and remote sensing satellite sea surface temperature, chlorophyll concentration, diffuse attenuation of downwelling irradiance, and in-situ salinity were used as the input variables. The seasonal and interannual variations of modeled pCO2 were discussed from January 2003 and December 2021 in the Yellow Sea. The results showed that the model developed for this study performed well, with a root mean square difference (RMSD) of 43 μatm and a coefficient of determination (R2) of 0.67. Moreover, modeled pCO2 increased at a rate of 0.36 μatm year-1 (R2 = 0.27, p < 0.05) in the YS, which is much slower than the rate of atmospheric pCO2 (pCO2air) rise. The reason behind it needs further investigation. Compared with pCO2 from other datasets, the pCO2 derived from the RF model exhibited greater consistency with the in-situ pCO2 (RMSD = 55 μatm). In general, the RF model has significant improvement over the previous models and the global data sets.

show abstract

RainForest: A random forest algorithm for quantitative precipitation estimation over Swizerland

Cited by 3 publications

References 37 publications

Enhancing Spatial Resolution of SMAP Soil Moisture Products through Spatial Downscaling over a Large Watershed: A Case Study for the Susquehanna River Basin in the Northeastern United States

Enhancing Spatial Resolution of SMAP Soil Moisture Products through Spatial Downscaling over a Large Watershed: A Case Study for the Susquehanna River Basin in the Northeastern United States

A comparison of five models in predicting surface dead fine fuel moisture content of typical forests in Northeast China

Remote sensing and machine learning method to support sea surface pCO2 estimation in the Yellow Sea

Contact Info

Product

Resources

About