Integrating Deep Learning and the Google Earth Engine Could Computing Platforms for Fast and Accurate Cloud Masking

In dealing with optical satellite images, accurate and efficient cloud positioning and masking are often the prerequisites for the subsequent tasks. However, the current cloud masking algorithms have difficulty in achieving these two goals at the same time. On the one hand, though many physically based models such as the Fmask algorithms have been proposed and widely applied, the performance of these methods still has some limitations including manually adjusted parameters in selecting the proper index and poor performance in thin cloud detection. This situation is much worse when applying the built-in Fmask in cloud-computing platforms such as the Google Earth Engine (GEE). On the other hand, an increasing number of sophisticated algorithms based on Deep Learning such as Convolutional Neural Networks (CNN) and Transformers have been proposed, they are, in most cases, deployed in local environments and require huge amounts of computational capacity to accomplish the tasks, which is inefficient and cannot be quickly utilized in large-scale and long-term studies, especially shows limitations with the trial of transferring the model into the GEE platforms. To solve the aforementioned dilemmas, the present study proposes a novel method for cloud masking by integrating deep learning and cloud-computing GEE. First, we construct a cloud dataset that is composed of globally selected cloud-contaminated pixels with the Fmask algorithm. Then, we use this cloud dataset to train a lightweight and flexible deep learning model based on LeNet. Last to screen out cloud pixels. Last, the developed model is transferred into the cloud-computing GEE platform and used to conduct cloud masking for each optical satellite image. The results show that in comparison to the conventional Fmask algorithm, the performance of the proposed exists superior in both detecting thick and thin clouds. More importantly, cloud masking can be achieved without bureaucratic procedures such as first downloading images and uploading the cloud masking, which is often required by locally developed deep learning models. By utilizing the proposed method, accurate and fast cloud detection can be achieved on the GEE platform that can be used in subsequent tasks like image compositing. For example, the generated monthly mean composting images show much better performance in visualizing ground objects when compared to those images based on the Fmask method, as the remaining cloud pixels that cannot be detected by the built-in Fmask algorithms can be more accurately examined. Through the usage of the proposed cloud mask methods, the merits of the powerful strength in data fitting and model optimization of deep learning algorithms and high efficiency in dealing with big data of GEE are naturally integrated, which is expected to shed light on cloud masking and other remote sensing modeling tasks.