As the second largest rice producer, India contributes about 20% of the world’s rice production. Timely, accurate, and reliable rice yield prediction in India is crucial for global food security and health issues. Deep learning models have achieved excellent performances in predicting crop yield. However, the interpretation of deep learning models is still rare. In this study, we proposed a transformer-based model, Informer, to predict rice yield across the Indian Indo-Gangetic Plains by integrating time-series satellite data, environmental variables, and rice yield records from 2001 to 2016. The results showed that Informer had better performance (R² = 0.81, RMSE = 0.41 t/ha) than four other machine learning and deep learning models for end-of-season prediction. For within-season prediction, the Informer model could achieve stable performances (R² ≈ 0.78) after late September, which indicated that the optimal prediction could be achieved 2 months before rice maturity. In addition, we interpreted the prediction models by evaluating the input feature importance and analyzing hidden features. The evaluation of feature importance indicated that NIR_V was the most critical factor, while intervals 6 (mid-August) and 12 (mid-November) were the key periods for rice yield prediction. The hidden feature analysis demonstrated that the attention-based long short-term memory (AtLSTM) model accumulated the information of each growth period, while the Informer model focused on the information around intervals 5 to 6 (August) and 11 to 12 (November). Our findings provided a reliable and simple framework for crop yield prediction and a new perspective for explaining the internal mechanism of deep learning models.

Light use efficiency (LUE) is a key indicator of vegetation photosynthesis, which provides important insights into how vegetation productivity responds to environmental conditions. The photochemical reflectance index (PRI) is based on reflectance at 531 and 570 nm, which reflects the xanthophyll cycle process of plants under different radiation conditions, and makes LUE related to plant optical characteristics. In this study, tower-based PRI and eddy covariance (EC) based LUEs were used to explore the ability of PRI to track LUE variations in a subtropical, evergreen mixed forest in South China. The results indicate that there is a stronger relationship between PRI and LUE, corrected by the bidirectional reflectance distribution function (BRDF), where R² = 0.46 before correction and R² = 0.60 after correction. Generally, PRI is able to capture diurnal and seasonal changes in LUE. Simultaneously, this study highlights a significant correlation between LUE and PRI, but there is also a large seasonal difference in its correlation. The correlation in winter was significantly stronger than summer. The strongest correlation is found in November (R² = 0.91) and the weakest is in July (R² = 0.34). Photosynthetically active radiation (PAR) had a strong influence on the LUE-PRI relationship, while vapor pressure deficit (VPD) and air temperature (Ta) had negative influences on the relationship between LUE and PRI. Terrestrial laser scanning is used to retrieve the vertical structure of forest crown. Our results show that the vegetation canopy structure (i.e., effective leaf area index, LAI_e), extracted from terrestrial laser scanning (TLS) point data in subtropical mixed forests, had a weak influence on LUE. Our research suggests that environmental factors and vegetation canopy structures should be considered when using PRI to accurately estimate LUE.

Fixed-wing unmanned aerial vehicles (UAVs) and multi-rotor UAVs are widely utilized in large-area (>1 km²) environmental monitoring and small-area (<1 km²) fine vegetation surveys, respectively, having different characteristics in terms of flight cost, operational efficiency, and landing and take-off methods. However, large-area fine mapping in complex forest environments is still a challenge in UAV remote sensing. Here, we developed a method that combines a multi-rotor UAV and a fixed-wing UAV to solve this challenge at a low cost. Firstly, we acquired small-scale, multi-season ultra-high-resolution red-green-blue (RGB) images and large-area RGB images by a multi-rotor UAV and a fixed-wing UAV, respectively. Secondly, we combined the reference data of visual interpretation with the multi-rotor UAV images to construct a semantic segmentation model and used the model to expand the reference data. Finally, we classified fixed-wing UAV images using the large-area reference data combined with the semantic segmentation model and discuss the effects of different sizes. Our results show that combining multi-rotor and fixed-wing UAV imagery provides an accurate prediction of tree species. The model for fixed-wing images had an average F1 of 92.93%, with 92.00% for Quercus wutaishanica and 93.86% for Juglans mandshurica. The accuracy of the semantic segmentation model that uses a larger size shows a slight improvement, and the model has a greater impact on the accuracy of Quercus liaotungensis. The new method exploits the complementary characteristics of multi-rotor and fixed-wing UAVs to achieve fine mapping of large areas in complex environments. These results also highlight the potential of exploiting this synergy between multi-rotor UAVs and fixed-wing UAVs.