Solar-induced chlorophyll fluorescence (SIF), as a direct probe of vegetation photosynthesis, has recently been an effective indicator for crop yield estimation in late-season. Spatio-temporal prediction of SIF (STP-SIF) from mid- to late-season could be a promising solution to forecast crop yields in mid-season. However, STP-SIF has not been well explored in assessing its applicability in crop yield forecasting. In the study, we first explored the potential driving mechanism and designed time-series data-driven deep learning approaches for the STP-SIF issue. We compared the feasibility of four styles of explanatory variables, six network structures, and the corresponding eight approaches in SIF prediction in the main cotton-planting area in Northern Xinjiang, taking the two tasks of one and two months before harvest as examples. The STP-SIF products of the five approaches, namely Spe_CNN-LSTM (a hybrid network combining convolutional neural network and long short-term memory considering spectra), Spe_CNN3D (a three-dimensional CNN considering spectra), SpaSpeSIF_CNN3D (a three-dimensional CNN considering spectra and SIF), SpaSpeSIF_HCNN (a hybrid network combining CNN2D and CNN3D considering spectra and SIF), and SIF_CNN-LSTM (a CNN-LSTM network considering SIF), showed a similar spatial pattern to the referenced SIF, indicating that these proposed approaches could be better for STP-SIF. Discussion of the spatial scale effect of the STP-SIF issue preliminary showed that our proposed methods had little spatial scale dependence and hence could be suitable for STP-SIF in different spatial extents at diverse resolutions. Further, we performed the experiments for the regional-scale cotton yield forecast one to two months before harvest based on the STP-SIF products of August and September. The best cotton yield prediction accuracies, with R2 of 0.70 and 0.66 for one and two months before harvest, were obtained respectively by the combination of the known SIF and NDWI (Normalized Difference Water Index) and the predicted SIF by SIF_CNN-LSTM. This study offers a baseline for the STP-SIF issue and reveals the feasibility of the STP-SIF products for accurate crop yield forecast.