It is important for aquaculture monitoring, scientific planning, and management to extract offshore aquaculture areas from medium-resolution remote sensing images. However, in medium-resolution images, the spectral characteristics of offshore aquaculture areas are complex, and the offshore land and seawater seriously interfere with the extraction of offshore aquaculture areas. On the other hand, in medium-resolution images, due to the relatively low image resolution, the boundaries between breeding areas are relatively fuzzy and are more likely to ‘adhere’ to each other. An improved U-Net model, including, in particular, an atrous spatial pyramid pooling (ASPP) structure and an up-sampling structure, is proposed for offshore aquaculture area extraction in this paper. The improved ASPP structure and up-sampling structure can better mine semantic information and location information, overcome the interference of other information in the image, and reduce ‘adhesion’. Based on the northeast coast of Fujian Province Sentinel-2 Multispectral Scan Imaging (MSI) image data, the offshore aquaculture area extraction was studied. Based on the improved U-Net model, the F1 score and Mean Intersection over Union (MIoU) of the classification results were 83.75% and 73.75%, respectively. The results show that, compared with several common classification methods, the improved U-Net model has a better performance. This also shows that the improved U-Net model can significantly overcome the interference of irrelevant information, identify aquaculture areas, and significantly reduce edge adhesion of aquaculture areas.

Grassland gross primary productivity (GPP) is an important part of global terrestrial carbon flux, and its accurate simulation and future prediction play an important role in understanding the ecosystem carbon cycle. Machine learning has potential in large-scale GPP prediction, but its application accuracy and impact factors still need further research. This paper takes the Mongolian Plateau as the research area. Six machine learning methods (multilayer perception, random forest, Adaboost, gradient boosting decision tree, XGBoost, LightGBM) were trained using remote sensing data (MODIS GPP) and 14 impact factor data and carried out the prediction of grassland GPP. Then, using flux observation data (positions of flux stations) and remote sensing data (positions of non-flux stations) as reference data, detailed accuracy evaluation and comprehensive trade-offs are carried out on the results, and key factors affecting prediction performance are further explored. The results show that: (1) The prediction results of the six methods are highly consistent with the change tendency of the reference data, demonstrating the applicability of machine learning in GPP prediction. (2) LightGBM has the best overall performance, with small absolute error (mean absolute error less than 1.3), low degree of deviation (root mean square error less than 3.2), strong model reliability (relative percentage difference more than 5.9), and a high degree of fit with reference data (regression determination coefficient more than 0.97), and the prediction results are closest to the reference data (mean bias is only −0.034). (3) Enhanced vegetation index, normalized difference vegetation index, precipitation, land use/land cover, maximum air temperature, potential evapotranspiration, and evapotranspiration are significantly higher than other factors as determining factors, and the total contribution ratio to the prediction accuracy exceeds 95%. They are the main factors influencing GPP prediction. This study can provide a reference for the application of machine learning in GPP prediction and also support the research of large-scale GPP prediction.

It is still difficult to obtain high-resolution and fast-updated NDVI data, and spatiotemporal fusion is an effective means to solve this problem. The purpose of this study is to carry out the comparative analysis and comprehensive trade-off of spatiotemporal fusion models for NDVI generation and to provide references for scholars in this field. In this study, four spatiotemporal fusion models (STARFM, ESTARFM, FSDAF, and GF-SG) were selected to carry out NDVI image fusion in grassland, forest, and farmland test areas, and three indicators of root mean square error (RMSE), average difference (AD), and edge feature richness difference (EFRD) were used. A detailed evaluation and analysis of the fusion results and comprehensive trade-off were carried out. The results show that: (1) all four models can predict fine-resolution NDVI images well, but the phenomenon of over-smoothing generally exists, which is more serious in high-heterogeneity areas; (2) GF-SG performed well in the evaluation of the three indicators, with the highest comprehensive trade-off score (CTS) of 0.9658. Followed by ESTARFM (0.9050), FSDAF (0.8901), and STARFM (0.8789); (3) considering the comparative analysis and comprehensive trade-off results of the three test areas and the three indicators, among the four models, GF-SG has the best accuracy in generating NDVI images. GF-SG is capable of constructing NDVI time series data with high spatial and temporal resolution.