Climate studies and effective environmental management require unbiased climate datasets. This study develops a new bias correction approach using a three-layer feedforward neural network to reduce the biases of climate variables (temperature and precipitation) over northern South America. Air and skin temperature, specific humidity, and net longwave and shortwave radiation are used as inputs to the network for bias correction of 6-hourly temperature. Inputs to the network for bias correction of monthly precipitation are precipitation at lag 0, 1, 2, and 3 months, and also the standard deviation of precipitation from 3 × 3 neighbors around the pixel of interest. The climate model data are provided by the Community Climate System Model, version 3 (CCSM3). Results show that the trained artificial neural network (ANN) can improve the estimation error and correlation of the variables for both calibration and validation periods even when there is a low temporal consistency between the time series of the model data and targets. The developed model is also able to modify the probabilistic structure of the variables although the quantile-based information is not directly considered in the network. The ANN model outperforms linear regression, which is used for comparison purposes. The new method can be used to produce bias-corrected climate variables that can be used as forcing to hydrological and ecological models.

Background: Climate change and human activities are two main forces that affect the intensity, duration, and frequency of wildfires, which can lead to risks and hazards to the ecosystems. This study uses machine learning (ML) as an effective tool for predicting wildfires using historical data and influential variables. The performance of two machine learning algorithms, including logistic regression (LR) and random forest (RF), to construct wildfire susceptibility maps is evaluated in regions with different physical features (Okanogan region in the US and Jamésie region in Canada). The models’ inputs are eleven physically related variables to output wildfire probabilities. Results: Results indicate that the most important variables in both areas are land cover, temperature, wind, elevation, precipitation, and normalized vegetation difference index. In addition, results reveal that both models have temporal and spatial generalization capability to predict annual wildfire probability at different times and locations. Generally, the RF outperforms the LR model in almost all cases. The outputs of the models provide wildfire susceptibility maps with different levels of severity (from very high to very low). Results highlight the areas that are more vulnerable to fire. The developed models and analysis are valuable for emergency planners and decision-makers in identifying critical regions and implementing preventive action for ecological conservation.