Extracting Corn Planting Area By Multi-Source Data With SVM Mixed-Field Decomposed Method

The corn planting area can reflect how the agricultural production make use of the agricultural resource in the space, by which we can adjust the agricultural structure, and can forecast the comprehensive productivity of agricultural resource and the carrying capacity of population. Accurate and timely updated-information for corn planting area is essential to yield estimation, agricultural management and food security. Comparing to the traditional ground investigation, using remote sensing to measure the planting area of corn can exclude the human disturbance greatly, and can save manpower, material resources, financial and time, which will have great economic and social benefits.Most current automatic classification techniques to obtain crop planting area from digital imagery operate on a per-pixel basis in isolation from other pertinent information. Therefore, per-pixel techniques often yield results with limited reliability on areas where parcel size is too small. The reliability of image classification can be improved by including a priori knowledge about the contextual relationships of the pixels in the classification process. For per-field classification, the geometry of the boundaries defines the spatial context between the pixels contained within, and enables those pixels to be processed in coherence. A final decision on the class assignment of pixels within each field is taken based on the coherent processing of these pixels. This is unlike per-pixel classification where the decision for each pixel is reached independently, and it can provide the best results.In order to obtain the planting area of corn, this study chose yuanyang county as experimental area and established field background database by high-resolution image. Then we pretreated the data and used the NDVI and reflectivity to carry through the classification according to the multi-temporal TM images. Then we integrated the classification results and vector field boundary, taken the area proportion of corn in the field as the hierarchical model to establish the hierarchical model, then we went out to investigate the real area proportion of corn in the selected field. In virtue of the vector field boundary, we extracted the eigenvector on TM images, then we integrated the spectral reflectivity, vegetation-index and texture to compose SVM mixed-field decomposed model. At last, we used the investigate results as standard to judge the per-field classification results, and compared the per-field classification results to the single-source data and per-pixel classification results. By analyzing the accuracy, this paper drew main conclusions as follows:(1) We established field background database on high-resolution image, which could insure the field boundary corresponding to the real field, and was easy to open surveying field-work. But for measuring corn area greatly, the work efficiency of this method was far worse than division-technique automatically, so how to improve the accuracy of the division-technique is the stress for our future work.(2) We got higher accuracy of corn planting area measuring when using multi-source data than single-source data, which showed that vegetation-index and texture could improve the accuracy of classification.(3) We got higher accuracy of corn planting area measuring when using SVM than the decision-tree classification, which showed that SVM had the ability to extend the high-dimension data, and fit classification with a few-swatch.(4) We selected yuanyang country as our experimental area and had a good accuracy ,but how to extend this method to the whole province is the stress for the future.