.png){kind=small}
Effective Downstream Regression Model Construction and its Tranferability Exploration: Take the LAI Estimation as a Case
Ms GU Xiangfeng
(Supervisor: Prof Shunlin Liang)
Abstract:
The foundation model, trained on large-scale data and containing millions of parameters, can be adapted and fine-tuned to various downstream tasks. The Foundation Model enables an unprecedented level of understanding and reasoning and achieves remarkable transfer performance in finishing downstream tasks. Remote Sensing Foundation models (RSFM) are derived from computer visual Foundation models. They are data-oriented, unsupervised, and pre-trained on tons of remote sensing image datasets, and ultimately integrate and store the learned features for downstream tasks application. The downstream tasks construction focuses on semantic segmentation, change monitoring, and object detection tasks related to image classification. While Foundation models theoretically support the construction of downstream regression tasks, little research has made it a practice and the effective practice of regression tasks remains questionable. This study aims to construct an effective regression downstream model based on the RSFM, Prithvi, to estimate 30m resolution LAI and explore the transferability of the regression model. The regression model utilizes a multi-temporal visual transformer architecture and takes the convolutional neck and regressive decoder for construction. To improve the robustness of the model, the research will modify aspects like feature extractors, auxiliary data pipelines, and selective parameters fine-tuning for better results. The data input to the downstream model includes HLS multi-spectral images as well as 30m Hi-GLASS LAI products as labels. The in-situ validation dataset is used to verify the estimation effect. The performance of the downstream regression task shows that it can complete the reproduction of the basic contour of LAI label images, but many differences exist in the details and numerical prediction. The preliminary result proves that RSFM can support the construction and use of regression tasks, but still lots of questions to improve for reliable index estimation. This research expects to generate a regression downstream model for various remote sensing indexes estimation. This study can provide some practical support for the overall construction and application of regression tasks of Foundation models, and provide other potential directions for high-resolution remote sensing index estimation methods.