基于Transformer光谱匹配的星机协同滨海盐渍化农田土壤含盐量监测方法

Satellite-UAV Collaborative Monitoring Method for Soil Salinity in Coastal Saline Farmland Based on Transformer Spectral Matching

摘要:

星机协同遥感为滨海盐渍化农田土壤含盐量（Soil salinity content，SSC）空间监测提供了重要途径，但该类区域含盐量空间异质性强、地表覆盖复杂，Sentinel-2多光谱成像仪（Multispectral instrument，MSI）影像受波段设置、光谱分辨率和混合像元效应制约，含盐量敏感信息表达有限。为缓解跨传感器光谱差异对Sentinel-2 MSI影像SSC估算的影响，本文提出一种基于Transformer光谱匹配的星机协同土壤含盐量监测方法。以河北省黄骅市滨海盐渍化农田为研究区，同期获取实测SSC、无人机高光谱影像和Sentinel-2 MSI影像。将无人机高光谱影像重采样至10 m，并与Sentinel-2 MSI影像空间对齐;以无人机高光谱匹配波段反射率为参考，采用多层感知机（Multilayer perceptron，MLP）和Transformer构建Sentinel-2 MSI至无人机高光谱参考空间的非线性光谱匹配模型;最后构建原始波段和盐分光谱指数，筛选含盐量敏感特征，采用随机森林（Random forest，RF）模型估算SSC，并基于留一交叉验证评价模型精度。结果表明，Sentinel-2-RF模型决定系数（R2）为0.37，均方根误差（Root mean square error，RMSE）和平均绝对误差（Mean absolute error，MAE）分别为0.82 g/kg和0.62 g/kg;经MLP光谱匹配后，R2提高至0.45，RMSE和MAE分别为0.77 g/kg和0.61 g/kg;经Transformer光谱匹配后，R2进一步提高至0.49，RMSE和MAE分别为0.74 g/kg和0.58 g/kg。与Sentinel-2-RF模型相比，Sentinel-2_Transformer-RF模型R2提高0.12，RMSE降低0.08 g/kg。研究结果表明，Transformer光谱匹配能增强Sentinel-2 MSI影像的含盐量敏感信息表达，可为滨海盐渍化农田SSC空间分布推算提供方法层面参考。

Abstract:

Satellite-UAV collaborative remote sensing provides an important approach for spatial monitoring of soil salinity content (SSC) in coastal saline farmland. However, in such regions, soil salinity shows strong spatial heterogeneity and surface cover is complex. Sentinel-2 multispectral instrument (MSI) imagery was limited by band configuration, spectral resolution and mixed-pixel effects, resulting in limited expression of salt-sensitive information. To alleviate the influence of cross-sensor spectral differences on SSC estimation from Sentinel-2 MSI imagery, a satellite-UAV collaborative soil salinity monitoring method was proposed based on Transformer spectral matching. Coastal saline farmland in Huanghua City, Hebei Province, China, was selected as the study area, and measured SSC, UAV hyperspectral imagery and Sentinel-2 MSI imagery were acquired during the same period. Firstly, the UAV hyperspectral imagery was resampled to 10 m and spatially aligned with Sentinel-2 MSI imagery. Then, using UAV hyperspectral matched-band reflectance as a reference, multilayer perceptron (MLP) and Transformer models were used to construct nonlinear spectral matching models from Sentinel-2 MSI to the UAV hyperspectral reference space. Finally, original bands and salinity spectral indices were constructed, salt-sensitive features were selected, SSC was estimated by using a random forest (RF) model, and model accuracy was evaluated by leave-one-out cross-validation. Results showed that the coefficient of determination (R2) of the original Sentinel-2-RF model was 0.37, and the root mean square error (RMSE) and mean absolute error (MAE) were 0.82 g/kg and 0.62 g/kg, respectively. After MLP spectral matching, R2 was increased to 0.45, and RMSE and MAE were 0.77 g/kg and 0.61 g/kg, respectively. After Transformer spectral matching, R2 was further increased to 0.49, and RMSE and MAE were 0.74 g/kg and 0.58 g/kg, respectively. Compared with the original Sentinel-2-RF model, the Sentinel-2_Transformer-RF model increased R2 by 0.12 and reduced RMSE by 0.08 g/kg. The results indicated that Transformer spectral matching can enhance the expression of salt-sensitive information in Sentinel-2 MSI imagery and provide a methodological reference for SSC spatial distribution estimation in coastal saline farmland.