Abstract:  Banana is an important tropical fruit in Hainan. Limited by meteorological disasters and the level of production technology, banana production has weak stability and large inter-annual fluctuations in Hainan. Remote sensing yield estimation is currently one of the most widely used crop yield estimation methods, especially suitable for large-area, uniformly distributed planting types. Therefore, carrying out research on banana yield estimation by remote sensing and accurately grasping the change trend of yield is of great significance to the planning and stable development of the banana planting industry. Based on Landsat-8 and MODIS data, the object-oriented method was first used to extract the spatial distribution of banana growing areas in Hainan Island, and then the Mahalanobis distance method based on the time series vegetation index was used for the second extraction, and finally the results of the two extractions were spatially overlaid. The method of field verification at random selected points was used to evaluate the classification accuracy. Aiming at the problem of the small number of regional yield estimation samples, the MODIS data from 2014 to 2015 and the banana regional yield data in 2015 were collected. The growth of banana throughout the growth period was used as an input variable to establish an information diffusion model for regional yield estimation. NDVI data was calculated by daily MODIS images, and synthesized to monthly data. The monthly composite values of NDVI was summed to obtain the cumulative value of NDVI throughout the growth period. Using the obtained vector files of banana planting areas, the cumulative NDVI values of bananas were extracted in 18 counties during the whole growth period. As the input and output variables of the information diffusion model, the cumulative value of NDVI and yield data were logarithmically normalized. The normal diffusion function was used to diffuse the sample information into the whole field, and an information matrix composed of information increments was established. Then the fuzzy relationship between the cumulative value of NDVI and the yield was obtained from the information matrix, that was the fuzzy relationship matrix. Finally, through the fuzzy approximate reasoning method, the simulated yield was obtained from the input samples. The cross-validation method was used to evaluate the accuracy of production estimation, and at the same time, the accuracy of the production estimation model for the simulation of production change trends was evaluated. Three kinds of information diffusion estimation schemes were constructed by combining multiple growth stages: scheme I was the NDVI cumulative value modeling scheme for the whole growth period; scheme Ⅱ was the scheme of joint input of vegetative growth stage and reproductive growth stage; scheme Ⅲ was the scheme of joint input during the budding stage and fruit development stage. The estimation accuracy of each plan was compared at last. The results showed that: (1) the comprehensive classification method using object-oriented and Mahalanobis distance had a higher accuracy. The total classification accuracy and Kappa coefficient were 82.5% and 0.7338 respectively, and the result of the consistency test was better. (2) The information diffusion model based on banana growth during the whole growth period had high yield estimation accuracy, with an average relative error of 26.0%, a coefficient of determination of 0.9216, and good explanatory power and stability; the accuracy of the estimation of inter-annual yield change trends reached 83.3%. (3) The univariate information diffusion estimation scheme based on the entire growth period had the highest accuracy, and the average relative error was reduced by 2.9 and 10.4 percentage point respectively compared with the other two multivariate modeling schemes. Based on the above results, it could be found that a fuzzy relationship was established with the information diffusion method between banana growth and yield by constructing a fuzzy set. The model was estimated with high accuracy, good explanatory ability and stability, and the accuracy rate for predicting the inter-annual yield change trend was also high, which could meet actual business needs. Through the establishment of a phase-by-growth yield estimation program to compare the impact of each developmental stage on the yield estimation accuracy, it was found that the effect of information diffusion model with whole growth period as the input variable was the best, which could take into account the impact of meteorological disasters and sensitive growth periods on yield. The accuracy of the yield estimation scheme II, which inputted both the vegetative growth period and the reproductive growth period, was higher than that of the program III, which only considered the reproductive growth period. The performance of information diffusion method with advantages in the processing of small sample data and ability to simulate nonlinear relationships was better. Applicability of yield estimation model based on full growth period in banana planting areas was judged satisfactory. The yield estimation model could be applied in the late stage of banana fruit expansion, and could also be used to carry out regular rolling forecasts after budding, to improve timeliness and provided scientific basis for agricultural departments and farmers to rationally arrange production and sales.

Key words:  Information diffusion, Estimated model, Remote sensing, Cash crops, Banana