Mapping invasive Fallopia japonica by combined spectral, spatial, and temporal analysis of digital orthophotos

• Published in: International journal of applied earth observation and geoinformation Vol. 19; pp. 185 - 195
• Main Authors: Dorigo, Wouter, Lucieer, Arko, Podobnikar, Tomaž, Čarni, Andraž
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier B.V 01.10.2012; Elsevier
• Subjects: Alien species; Applied geophysics; Earth sciences; Earth, ocean, space; Exact sciences and technology; Internal geophysics; Multitemporal; NDVI; Random forest classifier; Texture; Southern Europe; Slovenia; Europe; Oceania; Alien species; Multitemporal; Random forest classifier; Texture; NDVI; cost; color; textures; reflectance; springs; probability; accuracy; remote sensing; vegetation; classification; forests; BASIC; agriculture; channels; America; ecosystems; hot spots; land cover
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2012.05.004

► We map the occurrence of invasive Fallopia japonica in Slovenia. ► We exploit spectral, textural, and temporal characteristics of digital orthophotos. ► We use a supervised random forest classifier. ► F. japonica can be well identified from low-cost RGB and false color infrared imagery. ► Results provide a valuable support tool for eradication measures. Japanese knotweed (Fallopia japonica) is listed among 100 of the World's worst invasive alien species and poses an increasing threat to ecosystems and agriculture in Northern America, Europe, and Oceania. This study proposes a remote sensing method to detect local occurrences of F. japonica from low-cost digital orthophotos taken in early spring and summer by concurrently exploring its temporal, spectral, and spatial characteristics. Temporal characteristics of the species are quantified by a band ratio calculated from the green and red spectral channels of both images. The normalized difference vegetation index was used to capture the high near-infrared (NIR) reflectance of F. japonica in summer while the characteristic texture of F. japonica is quantified by calculating gray level co-occurrence matrix (GLCM) measures. After establishing the optimum kernel size to quantify texture, the different input features (spectral, spatial, and texture) were stacked and used as input to the random forest (RF) classifier. The proposed method was tested for a built-up and semi-natural area in Slovenia. The spectral, spatial, and temporal provided an equally important contribution for differentiating F. japonica from other land cover types. The combination of all signatures resulted in a producer accuracy of 90.3% and a user accuracy of 98.1% for F. japonica when validation was based on independent regions of interest. A producer accuracy of 61.4% was obtained for F. japonica when comparing the classification result with all occurrences of F. japonica identified during a field validation campaign. This is an encouraging result given the very small patches in which the species usually occur and the high degree of intermingling with other plants. All hot spots were identified and even likely infestations of F. japonica that had remained undiscovered during the field campaign were detected. The probability images resulting from the RF classifier can be used to reduce the relatively large number of false alarms and may assist in targeted eradication measures. Classification skill only slightly reduced when NIR information was not considered, which is an important recognition with regard to transferability of the method to the most basic type of digital color orthophotos. The possibility to use orthophotos, which at most municipalities are commonly available and easily accessible, facilitates an immediate implementation of the approach in situations where intervention is urgent.