High-Resolution Seamless Mapping of the Leaf Area Index via Multisource Data and the Transformer Deep Learning Model

The leaf area index (LAI) is a critical parameter for monitoring vegetation health and studying climate change. The spatial resolutions of most LAI products range from 500–1,000 m. Only a few LAI products exhibit spatial resolutions ranging from 16–30 m, but notable missing data occur because of the revisit cycle of satellites and the effect of weather conditions. These methods cannot satisfy the requirements of LAI application communities. To address this issue, a workflow was proposed for high-resolution seamless mapping of the LAI on the basis of multisource data and the Transformer deep learning model. Jiangsu Province in China was chosen as the study area. In this area, numerous cloudy and rainy days occur annually. Harmonized Landsat and Sentinel-2 (HLS) and Moderate Resolution Imaging Spectroradiometer (MODIS) reflectance images were obtained. The MODIS and HLS images were first composited and spatially aligned. Then, on the basis of the MODIS images and the spatiotemporal Fusion Incorporating Spectral Autocorrection (FIRST) method, missing HLS image data were reconstructed, thus producing a reflectance product with a spatial resolution of 30 m and a temporal resolution of 12 days. On the basis of the reflectance product, a Transformer model was designed for LAI prediction and compared with models designed through backpropagation neural network (BPNN), convolutional neural network (CNN), long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) methods. These models were compared with and without transfer learning on the basis of an independent dataset. The best model was selected and employed to produce a LAI product for the study area, which was subsequently compared with an existing MODIS product from spatial and temporal perspectives. The results showed that the procedure for HLS reconstruction is effective, with errors varying between 4.00% and 15.93% for different bands. Among the LAI prediction models, the Transformer model consistently performed the best across all scenarios. Notably, the Transformer model trained via transfer learning yielded the best results, with a test <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.62, a root mean square error (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RMSE</i>) of 0.79 and a mean relative error (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MRE</i>) of 14.80%. The <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of the other models ranged from 0.31 to 0.59, the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RMSE</i> values ranged from 0.82 to 1.06, and the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MRE</i> values ranged from 15.41% to 22.28%. In addition, the HLS LAI established via the above best model provided greater spatiotemporal accuracy than did the MODIS LAI product. This study provides reference data for establishing seamless LAI products with high spatial and temporal resolutions, contributing to applications such as vegetation health monitoring and global change research.