Rice recognition ability basing on GF-1 multi-temporal phases combined with near infrared data

Abstract: Near-infrared wave bands have double sensitivities on water bodies and crop growth. Using near-infrared wave band on rice identification has notable advantage, and it is a key wave band in crop identification. The study chose 5 counties of Yinchuan City, Ningxia Hui Autonomous Region as the study area and took the near-infrared (770-890 nm) data of GF-1/WFV (wide field view) images on May 18, June 16, July 30, and September 13, 2016 as its data sources. By using decision-tree classification method, the study achieved the rice identification of 7 temporal combinations, including 4 single temporal data in May, June, July, and September, as well as 3 multi-temporal data of May/June, May/July, and May/June/July/September, and made a comparison with supervision classification results of full wave bands (0.45-0.52, 0.52-0.59, 0.63-0.69, 0.77-0.89 μm) of corresponding GF-1/WFV data. The rice identification accuracies of near-infrared wave bands of single temporal data in May, June, July, and September were 83.63%, 57.40%, 75.82% and 62.61% respectively. Except that the accuracy of May data was 5.75% higher than the full wave bands, the identification accuracies of other temporal phases were lower than that of full wave bands. The highest accuracy deviation was in June, 30.23%, and the lowest accuracy deviation was in July, 1.58%. The deviations in May and September were 5.75% and 25.47% respectively. The rice identification accuracies under 3 multi-temporal near-infrared combinations of May/June, May/July, May/June/July/September were 83.76%, 93.93%, and 94.03% respectively. The accuracies of near-infrared combinations of May/July, and May/June/July/September were 8.58% and 0.73% higher than that of full wave band data results respectively, but the accuracy of May/June was 5.47% lower. Regardless of near-infrared data or full wave band data, the minimum value, average value and maximum value of single temporal data identification accuracies were 57.40%, 76.31% and 88.10% respectively, with the Kappa coefficients of 0.22, 0.44 and 0.64 respectively; the minimum value, average value and maximum value of multi-temporal data identification accuracies were 83.76%, 89.98% and 94.03% respectively, with the Kappa coefficients of 0.52, 0.68 and 0.77 respectively; all identification accuracies of single temporal data were lower than the identification accuracies of the multi-temporal data. If using single temporal data as the data source of rice remote sensing identification, the identification accuracy can reach 75.82% based on the near-infrared WFV data of rice in middle growth period of July, which is consistent with the full wave band data result of this time phase; to reach the accuracy above 88.10%, it is necessary to use full wave band WFV data in September. If taking multi-temporal data of WFV data of temporal combinations of 2 rice growth periods i.e. early period of May and middle period of July as the data source of rice remote sensing identification, the identification accuracy can reach 93.93%, which is close to the highest identification accuracy of multi-temporal data. The study result shows that, for the near-infrared wave band, and in the applications with requirement on the rice identification accuracy of about 75%-85%, the 2 single temporal infrared WFV data in early growth period of May and middle growth period of July can be taken as the data sources of remote sensing identification. To reach the identification accuracy above 86%, it is necessary to choose the combination of near-infrared WFV data in early growth period of May and middle growth period of July as the remote sensing data sources, so as to achieve rice area spatial distribution results with relatively high identification accuracy. After further correction with visual observation, it can be taken as the crop area remote sensing supervision results.