基于STFCNN时空融合与VPM模型的农田总初级生产力反演研究

Inversion of Total Primary Productivity of Farmland Based on STFCNN Spatiotemporal Fusion and VPM Model

摘要:

针对MODIS GPP产品空间分辨率较低、难以满足农田尺度生态系统生产力精细化监测需求问题，以山西省寿阳县玉米种植区为研究对象，利用2014—2015年生长季MODIS和Landsat 8遥感数据，构建基于深度卷积神经网络的时空融合模型（Supervised-training fully-connected neural network，STFCNN），实现增强型植被指数（EVI）和地表水分指数（LSWI）的高时空分辨率重建，并以融合后植被指数驱动植被光合作用模型（Vegetation photosynthesis model，VPM）开展GPP反演。采用均方根误差（RMSE）、决定系数（R2）、结构相似性指数（SSIM）、通用图像质量指数（UIQI）、光谱角映射（SAM）和全局相对误差（ERGAS）等指标，比较了STFCNN与STARFM、ESTARFM、FSDAF和STNLFFM模型的融合性能，并利用涡度相关观测数据验证模型反演结果。结果表明：在EVI和LSWI时空融合中，STFCNN均表现出最佳的融合精度。EVI融合RMSE为0.02~0.09，R2为0.70~0.95，UIQI为0.82~0.98;LSWI融合RMSE为0.024~0.046，R2为0.54~0.88，整体优于其他模型。空间细节与散点分析表明，STFCNN能够有效保持农田边界和纹理结构信息，融合结果与真实Landsat影像具有较高一致性，像元级验证获得R2为0.985、RMSE为0.003，表现出最优的光谱重建能力和空间细节恢复能力。基于融合影像驱动VPM模型反演GPP，与实测GPP相比，反演结果R2由MODIS GPP产品的0.1682提高至0.3452，对植被生长旺盛期光合作用变化的响应能力明显增强，更准确地表征农田GPP时空变化。寿阳县2014—2015年GPP总体呈“西北部高、东南部低”的空间分布格局，平均GPP为585~827 gC/m2。研究结果表明，STFCNN能够充分融合MODIS数据的高时间分辨率和Landsat数据的高空间分辨率优势，可为区域农业生态系统生产力监测和碳循环研究提供高时空分辨率数据支撑。

Abstract:

Aiming to address the issue of low spatial resolution in MODIS GPP products, which makes it difficult to meet the needs of refined monitoring of ecosystem productivity at the farmland scale, focusing on a maize-growing area in Shouyang County, Shanxi Province, the MODIS and Landsat 8 remote sensing data from the 2014—2015 growing season was used, a supervised-training fully-connected neural network (STFCNN) was constructed to achieve high spatiotemporal resolution reconstruction of the enhanced vegetation index (EVI) and surface water index (LSWI). The fused vegetation index was then used to drive a vegetation photosynthesis model (VPM) for GPP inversion. The fusion performance of STFCNN with STARFM, ESTARFM, FSDAF, and STNLFFM models was compared using indicators such as root mean square error (RMSE), coefficient of determination (R2), structural similarity index (SSIM), universal image quality index (UIQI), spectral angle mapping (SAM), and global relative error (ERGAS). Eddy covariance observation data were used to verify the model inversion results. The results showed that STFCNN exhibited the best fusion accuracy in both EVI and LSWI spatiotemporal fusion. The RMSE of EVI fusion was 0.02~0.09, R2 was 0.70~0.95, and UIQI was 0.82~0.98;the RMSE of LSWI fusion was 0.024~0.046, and R2 was 0.54~0.88, which was better than that of other models overall. Spatial detail comparison and scatter analysis showed that STFCNN can effectively preserve farmland boundary and texture structure information. Its fusion results had high consistency with real Landsat images. Pixel-level verification yielded R2 of 0.985 and RMSE of 0.003, showing the best spectral reconstruction ability and spatial detail recovery ability. Based on the fusion image-driven VPM model, the correlation of the inverted GPP with the measured GPP was increased from 0.1682 in the MODIS GPP product to 0.3452, which significantly enhanced the response to changes in photosynthesis during the peak vegetation growth period and more accurately characterized the spatiotemporal changes of farmland GPP. The GPP of Shouyang County in 2014—2015 generally showed a spatial distribution pattern of “high in the northwest and low in the southeast”, with an average GPP range of 585~827 gC/m2. The result showed that STFCNN can fully integrate the high temporal resolution advantage of MODIS data and the high spatial resolution advantage of Landsat data, and can provide high spatiotemporal resolution data support for regional agricultural ecosystem productivity monitoring and carbon cycle research.