4.1.2. Classification Results
(1) Classification Accuracy and Its Variation
Figure 5 and
Figure 6 illustrate the classification accuracies generated from spectral features, textural features and the combined spectral and textural features across varying spatial resolution imagery (a part of spatial resolutions) in Jizhou and Guyuan. It can be found that differences in accuracies (including the OA, PA and UA) were observed in different feature sets and study areas.
In Jizhou, for the spectral based classification, the OA reached the highest level of 91.16% at a spatial resolution of 10 m, and the PA and UA achieved the highest value of 88.30% and 88.59% at a spatial resolution of 8 m respectively. For the textural based classification, the OA reached the highest level of 95.95% at a spatial resolution of 6 m, and the PA and UA achieved the highest value of 90.85% and 93.58% at a spatial resolution of 6 m and 8 m respectively. For the combined spectral and textural feature based classification, the OA, PA and UA reached the highest level of 96.05%, 90.99% and 94.22% all at a spatial resolution of 6 m.
In Guyuan, for the spectral based classification, the OA, PA and UA achieved the highest value of 82.79%, 87.10% and 82.86% all at a spatial resolution of 6 m. For the textural based classification, the OA reached the highest level of 89.01% at a spatial resolution of 12 m, and the PA and UA achieved the highest value of 88.58% and 90.38% at the spatial resolution of 12 m and 8 m, respectively. For the combined spectral and textural based classification, the OA, PA and UA reached the highest level of 89.511%, 88.34% and 90.92% all at a spatial resolution of 12 m.
For the variation in OA, PA and UA for mapping PMF over varying spatial resolution imagery, we can see that when the spatial resolution changed from 2 m to 250 m, decreases were observed in OA, PA and UA as a general tendency. However, the varying trend in accuracies depends on the different feature sets and different study areas.
For Jizhou, as the spatial resolution coarsened from 2 m to 32 m, the OA, PA and UA generated from spectral features alone decreased by 1.59%, 2.95% and 1.46% respectively. The OA, PA and UA generated from textural features alone decreased by 5.27%, 3.36% and 9.01% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 4.75%, 3.25% and 9.46% respectively. As the spatial resolution coarsened from 32 m to 100 m, the OA, PA and UA generated from spectral features alone decreased by 17.94%, 12.21% and 5.70% respectively. The OA, PA and UA generated from textural features alone decreased by 22.07%, 8.10% and 11.81% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 23.02%, 8.10% and 12.36% respectively. As the spatial resolution coarsened from 100 m to 250 m, the OA, PA and UA generated from spectral features alone decreased by 40.57%, 52.86% and 54.59%. The OA, PA and UA generated from textural features alone decreased by 38.28%, 61.90% and 51.74%. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 38.28%, 61.90% and 51.74% respectively. Furthermore, as the spatial resolution coarsened from 2 m to 250 m, the OA, PA and UA generated from spectral features alone decreased by 60.10%, 68.02% and 61.75% respectively. The OA, PA and UA generated from the textural features alone decreased by 65.61%, 73.36% and 72.56% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 66.04%, 74.05% and 73.56% respectively.
For Guyuan, as the spatial resolution coarsened from 2 m to 32 m, the OA, PA and UA generated from the spectral features alone decreased by 2.34%, 1.53% and 2.58% respectively. The OA, PA and UA generated from the textural features alone decreased by 3.39%, 4.00% and 6.58% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 2.90%, 2.09% and 4.22% respectively. As the spatial resolution coarsened from 32 m to 100 m, the OA, PA and UA generated from the spectral features alone decreased by 13.25%, 23.59% and 13.39% respectively. The OA, PA and UA generated from the textural features alone decreased by 15.94%, 20.92% and 20.78% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 15.11%, 19.49% and 15.90% respectively. As the spatial resolution coarsened from 100 m to 250 m, the OA, PA and UA generated from the spectral features alone decreased by 12.88%, 54.07% and 41.67%. The OA, PA and UA generated from the textural features alone decreased by 10.22%, 46.38% and 26.02%. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 11.77%, 43.10% and 39.53% respectively. As the spatial resolution coarsened from 2 m to 250 m, the OA, PA and UA generated from the spectral features alone decreased by 28.46%, 79.19% and 57.64% respectively. The OA, PA and UA generated from the textural features alone decreased by 29.55%, 71.30% and 53.38% respectively. In addition, the OA, PA and UA generated from the combined spectral and textural features decreased by 29.78%, 64.68% and 59.65% respectively.
The decrease in accuracy was larger for the textural feature alone than the spectral feature alone, from 2 m to 32 m and from 32 m to 100 m; and was larger for the spectral feature alone than the textural feature alone, from 100 m to 250 m. From 2 m to 250 m, the larger decrease in accuracy was observed for the textural feature alone. The smallest decrease in accuracy was observed at the range from 2 m to 32 m, the larger decrease was observed from 32 m to 100 m and from 100 m to 250 m, regardless of the study areas and the feature sets.
The textural features from the fine resolution imagery were available for describing the PMF. Joining the textural features and spectral features with the finer spatial resolution imagery for mapping PMF yields better results compared to using spectral or textural features alone. In addition, at higher resolution imagery, the contribution of the textural feature was greater than the spectral feature. The plastic mulch is 1 m in width, and separated by 0.2–0.5 m bare soil, and the objective of our study is to map the PMF at a paddock scale. Therefore, at the higher spatial resolution, the imagery obtains more detailed information with respect to the spectral feature, and increases the within-class differences. The textural features imagery is obtained by the statistics in a fixed moving window size, thus textural features can control the salt and pepper effect and highlight the spatial characteristic, and increase the inter-class differences. With the coarsening spatial resolution, the textural feature of PMF is lost, because the fixed moving window size to compute the textural feature corresponds to a larger field size and the patch size of PMF is mostly small and fragmented. Therefore, the contribution of the textural feature depends largely on the spatial resolution of the imagery rather than the spectral feature. As the spatial resolution becomes coarser, the contribution of the textural feature was decreased significantly.
(2) Spatial distribution of PMF obtained from different spatial resolution imagery
The spatial distribution of PMF in our study areas (
Figure 7 and
Figure 8) was generated from a SVM with the aid of the developed three feature sets over the varying spatial resolution imagery.
The spatial distributions of PMF in Jizhou extracted by SVM using different feature sets are displayed in
Figure 7, taking PMF generated from different feature sets with spatial resolutions of 6 m, 64 m, 100 m and 250 m as examples. We can observe that most of the PMF in Jizhou was distributed in the central region and dispersed in the northern and southern regions. Marked differences were observed among the results generated from different feature sets across the varying spatial resolution. The distribution of PMF generated from the coarser spatial resolution imagery included more mismatching results than that from the finer resolution imagery. The spatial structure of PMF obtained from textural features from finer spatial resolution imagery was more obvious and reasonable than that from the spectral features. Over the varying spatial resolutions, the distribution of PMF from the coarser resolution imagery was extended further than that from the finer resolution imagery due to the greater commission error resulting from mixed pixels and blurred structural information. Such a phenomenon was similar across different feature sets but was more serious in the textural feature and less serious in the other two feature sets.
The spatial distributions of PMF in Guyuan extracted by SVM using different feature sets are displayed in
Figure 8, taking PMF generated from the different feature sets with spatial resolutions of 6 m, 64 m, 100 m and 250 m as examples. We can observe that most of the PMF in Guyuan, Ningxia, were dispersed across whole regions. There are significant differences between feature sets over the varying spatial resolution. At the spatial resolution of 64 m and 100 m, the distribution area of PMF was larger than that from the finer resolution imagery due to the greater commission error resulting from mixed pixels and blurred structural information. Furthermore, at the spatial resolution of 250 m, the PMF is almost undetectable, particularly when using the textural or combined spectral and textural features.