Machine learning and Global Vegetation: Random Forests for Downscaling and Gapfilling

Abstract

Abstract. Drought is a devastating natural disaster, where water shortage often manifests itself in the health of vegetation. Unfortunately, it is difficult to obtain high-resolution vegetation drought impact, which is spatially and temporally consistent. While remotely sensed products can provide part of this information, they often suffer from data gaps and limitations in spatial or temporal resolutions. A persistent feature among remote sensing products is tradeoffs between spatial resolution and revisiting times, where high temporal resolution is met by coarse spatial resolution and vice verse. Machine learning methods have been successfully applied in a wide range of remote sensing and hydrological studies. However, global applications to resolve drought impacts on vegetation dynamics still need to be made available, while there is significant potential for such a product to aid improved drought impact monitoring. To this end, this study predicted global vegetation dynamics based on the Enhanced Vegetation Index (evi) and the popular Random Forest algorithm (RF) at 0.1°. We assessed the applicability of RF as a gap filling and downscaling tool to generate spatial and temporal consistent global evi estimates. To do this, we trained an RF regressor with 0.1° evi data using a host of features indicative of water and energy balances experienced by vegetation and we evaluated the performance of this new product. Next, to test whether the RF is robust in terms of spatial resolution, we downscale global evi, the model trained on 0.1° data is used to predict evi at 0.01° resolution. The results show that the RF can capture global evi dynamics at both the 0.1° (RMSE: 0.02–0.4) and at the finer 0.01° (RMSE: 0.04–0.6) resolution. Overall errors were higher in the down-scaled 0.01° compared to the 0.1° product. Yet, relative increases remained small, thus demonstrating that RF can be used to create downscaled and temporally consistent evi products. Additional error analysis reveals that errors vary spatiotemporally, with underrepresented landcover types and periods of extreme vegetation conditions having the highest errors. Finally, this model is used to produce global spatially continuous evi products at both the 0.1° and 0.01° spatial resolution for 2003–2013 at an 8-day frequency.

Affine combination‐based over‐sampling for imbalanced regression

“…In the real world, there are also many prediction problems involving imbalanced continuous target values, such as extreme weather, 9 electricity forecasting, 10 water resources, 11 and financial markets. 12 Additionally to these common fields, there are new areas of research such as biology (e.g., estimation of foliage transpiration 13 ), chemistry (e.g., prediction of aquatic toxicity 14 ), geography (e.g., prediction of vegetation conditions 15 ), and metabolomics. 16 The objective of addressing imbalanced regression problems is to obtain a model from a set of characteristics and continuous response variables, ensuring that the model can accurately predict rare values.…”

ASER: Adapted squared error relevance for rare cases prediction in imbalanced regression

“…Handling imbalanced data is a challenging technical issue that is important in various application scenarios, such as medical area (e.g., medical diagnosis 4 ), informatics (e.g., text categorization 5 ), and financial markets (e.g., fraudulent credit card transactions 6 ). In addition to these more traditional fields, there are also some new ones, such as meteorology (e.g., wind speed forecasts 7 ), biology (e.g., estimation of plant transpiration 8 ), geography (e.g., prediction of vegetation conditions 9 ), and metabolomics. 10 The problem of data imbalance has always been a challenge that many scientists strive to solve, but only a few aspects related to imbalanced learning have been addressed with some solutions.…”