Aboveground wheat biomass estimation from a low-altitude UAV platform based on multimodal remote sensing data fusion with the introduction of terrain factors

Aboveground biomass is an important indicator used to characterize the growth status of crops, as well as an important physical and chemical parameter in agroecosystems. Aboveground biomass is an important basis for formulating management measures such as fertilization and irrigation. We selected four irrigated wheat fields in a region near Kaifeng, Henan Province, for this study. The terrain in that region was undulating and had spatial differences. We used a low-altitude unmanned aerial vehicle (UAV) remote sensing platform equipped with a multispectral camera, thermal infrared camera, and RGB camera to simultaneously obtain different remote sensing parameters during the key growth stages of wheat. Based on the extracted spectral reflectivity, thermal infrared temperature, and digital elevation information, we calculated the spatial variability of remote sensing parameters and growth indices under different terrain characteristics. We also analyzed the correlations between vegetation indices, temperature parameters, structural topographic parameters and aboveground biomass. Three machine learning methods were used, including the multiple linear regression method (MLR), partial least squares regression method (PLSR) and random forest regression method (RFR). We compared the aboveground biomass (AGB) estimation capability of single-modal data versus multimodal data fusion frameworks. The results showed that slope was an important factor affecting crop growth and aboveground biomass. We therefore analyzed several remote sensing parameters for three different slope scales. We found significant differences among them for soil water content, water content of plants, and aboveground biomass at four growth stages. Based on the strength of their correlation with aboveground biomass, seven vegetation indices (NDVI, GNDVI, NDRE, MSR, OSAVI, SAVI, and MCARI), four canopy structure parameters (CH, VF, CVM, SLOPE) and two temperature parameters (NRCT, CTD) were selected as the final input variables for the model. There was some variability in the accuracy of the models at different growth stages. The average accuracy of the models was anthesis stage > booting stage > filling stage > jointing stage. For the single-modal data framework, the model constructed with the vegetation indices was better than the aboveground biomass model constructed using the temperature or structure parameters, and the highest accuracy was obtained with an RFR model based on vegetation indices at the anthesis stage (R2 = 0.713). For the double modal data fusion approach, the highest accuracy resulted at the anthesis stage, using the structural parameters combined with the vegetation indices of the RFR model (R-2 = 0.842). Even higher accuracies were obtained using the multimodal data fusion approach with an RFR model based on vegetation indices, temperature parameters and structure parameters at the anthesis stage (R-2 = 0.897). By introducing terrain factors and combining them with the RFR algorithm to effectively integrate multimodal data, the complementary and synergistic effects between different remote sensing information sources could be fully exerted. The accuracy and stability of the aboveground biomass estimation models were effectively improved, and a high-throughput phenotype acquisition method was explored, which provides a reference and basis for real-time monitoring of crop growth and decoding the correlation between genotype and phenotype.

Automated summary

This study focuses on the estimation of aboveground biomass in irrigated wheat fields using remote sensing. The research used a low-altitude unmanned aerial vehicle equipped with various sensors to collect remote sensing parameters at different growth stages. Machine learning methods were employed to construct biomass estimation models. The results showed that slope was an important factor affecting crop growth and aboveground biomass, and the fusion of multimodal data improved the accuracy and stability of the models.