Downscaling solar-induced chlorophyll fluorescence for field-scale cotton yield estimation by a two-step convolutional neural network

As the largest cotton-growing region in China, Xinjiang has contributed more than 80% of the total national cotton production in recent years. Timely and accurate estimation of cotton yield in Xinjiang is important for sustainable agricultural development and food security. However, most current studies have been devoted to the linkage of crop yield with remotely sensed reflectance and climate parameters. This has caused numerous uncertainties due to that these explanatory variables are unable to quickly reflect the actual photosynthetic dynamics of crops. Solar-induced chlorophyll fluorescence (SIF), as a direct proxy of plant photosynthesis (gross primary productivity, GPP), has recently been suggested to be a promising method for crop yield estimation, but the spatial resolution of current SIF products derived from satellites is usually very low (such as Global OCO-2 SIF (GOSIF): 0.05°). This greatly limited the ability of SIF to accurately estimate field-scale cotton yield in Xinjiang. Here, we first proposed a two-step convolutional neural network (CNN) strategy to downscale the monthly GOSIF products sequentially from 0.05°, 0.005° to 0.0005° to match the size of cotton field parcels, and then linear regression and random forest (RF) regression were respectively conducted using the monthly downscaled SIF product (CNN-SIF) to assess its feasibility to estimate field-scale yield. Results showed that the proposed stepwise approach for downscaling GOSIF worked well, indicating a high goodness of fit (R2 > 0.85) with the referenced SIF as well as strong correlations to both GPP products and fraction of photosynthetically active radiation (FPAR) (the median r > 0.90). On this basis, preferable accuracies (the optimal R2 = 0.62 and the ratio of prediction to deviation = 1.64) were also achieved for our proposed cotton yield estimation models in the Mosuowan region, Xinjiang only by the 0.0005° SIF products. With the assistance of NDVI (normalized difference vegetation index), the higher performance was given (R2 = 0.67 and RPD = 1.72). This study reveals the importance of finer-resolution SIF products for accurate crop yield estimation and offers a promising and practical approach for estimating agricultural yield, especially for fragmented farmlands.