Abstract:

Organic matter is an important indicator of soil nutrient status in rubber plantations. The establishment of a fast and highly precise estimation model for organic matter content can better guide the fine production management of rubber plantations. In this study, 225 rubber plantation soil samples were collected from Dongfeng Farm in Jinghong City, and the spectral reflectance and organic matter content data of the soil samples were obtained. After noise band removal and resampling of the spectral reflectance, three methods [log(1/R), MSC and SNV] were applied to perform spectral transformation processing on resampled spectral reflectance R, and then SG smoothing or derivative transformation modes were used to optimize the spectral reflectance R and the spectral data of three transform forms. The best spectral transformation mode was obtained as log(1/R) combined with SG smoothing transformation. The SG smoothing transformation mode had the derivative order, SG filtering window, and polynomial degree of 0, 5, and 2 or 3, respectively. Based on the best spectral transformation data and soil organic matter content data, three methods including CARS, SPA and CARS-SPA were selected to extract characteristic wavelength, and the hyper-spectral estimation model for soil organic matter was constructed by MLR, PLSR and SVR. The results showed that the estimation accuracy of CARS-SVR model was the highest. The R², RMSE and RPD were respectively 0.897, 3.990 g/kg and 2.947. Therefore, the established hyper-spectral optimal estimation model for soil organic matter content in Yunnan mountainous rubber plantations, with RPD between 2.5 and 3.0, has good estimation ability.

Key words: rubber plantation, soil organic matter, hyper-spectral, competitive adaptive reweighted sampling, support vector regression