Research On Remote Sensing Classification For Main Fruiters Species Based On Canopy Hyperspectral Data In Southern Xinjiang Basin

The characteristic forestry and fruit industry of Xinjiang developed rapidly in the past ten years, so far, its area has more than about 22,000,000 mu, and it has become the dominant industry in Xinjiang to solve the "agricuture issues". However, in the process of industrialization in fruit, its informatization construction has lagged behind. Extracted fruit information based on hyperspectral remote sensing has the characteristcs of celerity, accuracy and wide range of application so that has important practical significance in th e sustainable development of fruit industry. Therefore, in this study we use four kinds of main fruiters species( Apple, Pear, Walnut, Date) as our target fruit tree species. Based on the principle of hyperspectral remote sensing, by measuring the fruiters species canopy hyperspectral data in different periods, we analyzed the dynamic changes of fruiters species canopy spectral’s curve in different periods. By using different dimension reduction methods of hyperspectral data, we selected and analyzed identifiable effective band and band features. Based on the precision testing of different classification method, we confirmed the identifiable effective band and the best period. The main conclusions are as follows:(1) Spectral characteristics existed deviations which are different on band and at various phenological periods of fruit development: the four kinds of main fruiters species canopy’s spectrum exhibited “down-up” feature. The overall trend of four kinds of main fruiters species canopy’s spectrum that is measured in different period in South XinJiang basin is similar, while in 525~575 nm, 650~700 nm and 750~850 nm band range, the difference is significant.(2) Using hyperspectral data transfer approach can improve the identification accuracy of trees. When using the Optimum Index Factor method to filter the sensitive band, the ideal data transfer method is d(log(R)), and the average classification accuracy is 95.92%. While using Band Index meth od to filter the sensitive band, the ideal data transfer method is d(N(R)), and the average classification accuracy is 96.88%.(3) Using band selection approach can improve the identification accuracy of the four fruit tree species. Refering to the Optimum Index Factor method, the best band combination is 690~699 nm, 880~889 nm, 890~899 nm. Refering to the Band Index method, the best band combination is 530~539 nm, 720~729nm, 760~769nm. When using BP Artificial Neural Network to evaluate the best band combination of the two referring method above, the former classification accuracy is 97.99% while the later is 98.66%. The Band Index method is better than Optimum Index Factor methed.(4) By applying BP Artificial Neural Network to identify four kinds of main fruiters species growed up in South XinJiang basin produce the best, most consistently accurate results. The average classification accuracy is 98.66% that is better than the Suppor Vector Machine with the accuracy of 24.01% and the Stepwise Discriminant Analysis with the accuracy of 22.15%. Meanwhile, the best period to filter four kinds of main fruiters species in South XinJiang basin by using analysis and comparison method is September.