Cloud Base Height Inversion Based on a Heterogeneous Stacked Integrated Network

To address the challenge of accurately inverting cloud base height (CBH) from passive remote sensing, this study proposes Hetero-CBHNet, a heterogeneous stacked integrated network framework for CBH inversion. The framework uses random forest (RF), convolutional neural network (CNN), and Transformer models as a pool of candidate base models and employs the RidgeCV algorithm to automatically learn the fusion weights. A total of 1,011,620 samples from FY-4A and MODIS, covering the period from May 2019 to May 2020, were selected (with CloudSat/CALIPSO 2B-CLDCLASS-LIDAR as the ground truth), and the dataset was split chronologically. Across the two data sources under the chronological splitting protocol, Hetero-CBHNet achieves RMSE values of 2966.12 m and 2797.64 m and R values of 0.630 and 0.659 on the test sets, outperforming the individual models. Feature analysis shows that incorporating operational passive cloud type products does not yield a stable gain, indicating that current passive cloud type information is not yet sufficient to replace active satellite labels. In addition, the RMSE values under the random splitting protocol are 1510.01 m and 1488.41 m, substantially better than those under chronological splitting, which confirms the strong dependence of the conclusions on the evaluation protocol.