H the nearest pixel center of a satellite data grid, as shown byvariable arrowsmodel to handle the impact on the aircraft measurements as one more input the red in our in Figure 4. of vertical distribution along the column. For HAPs ground Combretastatin A-1 web in-situ data, we assigned 0 as the Compound 48/80 Purity Height. Figure 4 illustrates how the in-situ data were matched up using the satellite information spatially. The circle represents the center of every pixel of satellite information, and the brown lines Remote Sens. 2021, 13, x FOR PEER Critique 6 of 23 indicate the vertical projection of in-situ information. The in-situ data is matched with the nearest pixel center of a satellite data grid, as shown by the red arrows in Figure 4.Figure 4. Spatial matchups of in-situ information with satellite information. Figure 4. Spatial matchups of in-situ data with satellite data.two.1.three. Worldwide DEM Data Given that descriptive statistics showed a negative relationship among surface altitude and in-situ concentration, using a Pearson’s correlation of r = -0.3907 in our in-situ dataset,Remote Sens. 2021, 13,six of2.1.3. Global DEM Data Since descriptive statistics showed a negative connection involving surface altitude and in-situ concentration, with a Pearson’s correlation of r = -0.3907 in our in-situ dataset, we employed worldwide Digital Elevation Model (DEM) information as one of the input variables, “Altitude”, in order to estimate the ground-level concentration. The relationship involving the variables “Height” and “Altitude” is shown in Figure 3b. In our study, we employed the Shuttle Radar Topography Mission (SRTM) DEM solution and resampled it to a resolution of 0.05 . This dataset had an initial resolution of 90 m in the equator and was supplied in WGS84 projection with a resolution of 1 arc [48]. two.2. Data Processing Just after collecting and organizing data into formattable structure, we visualized and preprocessed these data. Then, two neural networks were implemented for point and interval estimations by using PyTorch, a well-known deep-learning framework. Our code is out there on line (https://github.com/dingyizhe2000/Interval-HCHO-ConcentrationEstimation accessed on 21 June 2021). The preprocessed data with all the ground truth from in-situ HCHO concentration have been then divided randomly into two groups; 90 in the dataset was applied to train our models and 10 was applied for validation. After that, international VCD data were fed in to the model to be able to derive global surface level HCHO concentration. two.2.1. Preprocessing In theory, a neural network is in a position to manage input information using a varied distribution; nevertheless, a significant defect was noticed within the instruction course of action with out preprocessing, owing for the extremely imbalanced and skewed distribution with the HCHO concentration (both column and in-situ). For that reason, we initially applied log-transformation to the raw information. As shown in Figure S1, the logarithm in the HCHO concentration data shows a bell-shaped distribution, and increments in estimation accuracy have also confirmed the effectiveness of log-transformation. 2.2.2. Neural Network Architecture As a universal function approximator, the neural network played a vital role in assisting us derive the point and interval estimations from the HCHO concentration. Having said that, rather of coaching a single network to obtain these estimations jointly, two separate neural networks were constructed for point and interval estimation, respectively, mainly because a number of experiments which we carried out indicated that a joint model normally has to compromise among point estimation and in.