Pengyu Hao

Feature Selection of Time Series MODIS Data for Early Crop Classification Using Random Forest: A Case Study in Kansas, USA

Currently, accurate information on crop area coverage is vital for food security and industry, and there is strong demand for timely crop mapping. In this study, we used MODIS time series data to investigate the effect of the time series length on crop mapping. Eight time series with different lengths (ranging from one month to eight months) were tested. For each time series, we first used the Random Forest (RF) algorithm to calculate the importance score for all features (including multi-spectral data, Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and phenological metrics). Subsequently, an extension of the Jeffries–Matusita (JM) distance was used to measure class separability for each time series. Finally, the RF algorithm was used to classify crop types, and the classification accuracy and certainty were used to analyze the influence of the time series length and the number of features on classification performance; the features were added one by one based on their importance scores. Results indicated that when the time series was longer than five months, the top ten features remained stable. These features were mainly in July and August. In addition, the NDVI features contributed the majority of the most significant features for crop mapping. The NDWI and data from multi-spectral bands also contributed to improving crop mapping. On the other hand, separability, classification accuracy, and certainty increased with the number of features used and the time series length, although these values quickly reached saturation. Five months was the optimal time series length, as longer time series provided no further improvement in the classification performance. This result shows that relatively short time series have the potential to identify crops accurately, which allows for early crop mapping over large areas.

An improved high spatial and temporal data fusion approach for combining Landsat and MODIS data to generate daily synthetic Landsat imagery

Two key weaknesses of STDFA including sensor difference and spatial variability were adjusted.Three wildly used spatial and temporal fusion methods were compared.The correlation coefficient r had a negative exponential relationship with ratio of land cover change pixels.The accuracy of ISTDFA method had a logarithmic relationship with the size of applied area. Because of low temporal resolution and cloud influence, many remote-sensing applications lack high spatial resolution remote-sensing data. To address this problem, this study introduced an improved spatial and temporal data fusion approach (ISTDFA) to generate daily synthetic Landsat imagery. This algorithm was designed to avoid the weaknesses of the spatial and temporal data fusion approach (STDFA) method, including the sensor difference and spatial variability. A weighted linear mixed model was used to adjust the spatial variability of surface reflectance. A linear-regression method was used to remove the influence of differences in sensor systems. This method was tested and validated in three study areas located in Xinjiang and Anhui province, China. The other two methods, the STDFA and the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), were also applied and compared in those three study areas. The results showed that the ISTDFA algorithm can generate daily synthetic Landsat imagery accurately, with correlation coefficient r equal to 0.9857 and root mean square error (RMSE) equal to 0.0195, which is superior to the STDFA method. The ISTDFA method had higher accuracy than ESTARFM in areas greater than 200?× 200 MODIS pixels while the ESTARFM method had higher accuracy than the ISTDFA method in small areas. The correlation coefficient r had a negative power relation with ratio of land-cover change pixels. A land-cover change of 20.25% pixels can lead to a reduced correlation coefficient r of 0.295 in the blue band. The accuracy of the ISTDFA method indicated a logarithmic relationship with the size of the applied area, so it is recommended for use in large-scale areas.

The Potential of Time Series Merged from Landsat-5 TM and HJ-1 CCD for Crop Classification: A Case Study for Bole and Manas Counties in Xinjiang, China

Time series data capture crop growth dynamics and are some of the most effective data sources for crop mapping. However, a drawback of precise crop classification at medium resolution (30 m) using multi-temporal data is that some images at crucial time periods are absent from a single sensor. In this research, a medium-resolution, 15-day time series was obtained by merging Landsat-5 TM and HJ-1 CCD data (with similar radiometric performances in multi-spectral bands). Subsequently, optimal temporal windows for accurate crop mapping were evaluated using an extension of the Jeffries–Matusita (JM) distance from the merged time series. A support vector machine (SVM) was then used to compare the classification accuracy of the optimal temporal windows and the entire time series. In addition, different training sample sizes (10% to 90% of the entire training sample in 10% increments; five repetitions for each sample size) were used to investigate the stability of optimal temporal windows. The results showed that time series in optimal temporal windows can achieve high classification accuracies. The optimal temporal windows were robust when the training sample size was sufficiently large. However, they were not stable when the sample size was too small (i.e., less than 300) and may shift in different agro-ecosystems, because of different classes. In addition, merged time series had higher temporal resolution and were more likely to comprise the optimal temporal periods than time series from single-sensor data. Therefore, the use of merged time series increased the possibility of precise crop classification.

Comparison of Hybrid Classifiers for Crop Classification Using Normalized Difference Vegetation Index Time Series: A Case Study for Major Crops in North Xinjiang, China

A range of single classifiers have been proposed to classify crop types using time series vegetation indices, and hybrid classifiers are used to improve discriminatory power. Traditional fusion rules use the product of multi-single classifiers, but that strategy cannot integrate the classification output of machine learning classifiers. In this research, the performance of two hybrid strategies, multiple voting (M-voting) and probabilistic fusion (P-fusion), for crop classification using NDVI time series were tested with different training sample sizes at both pixel and object levels, and two representative counties in north Xinjiang were selected as study area. The single classifiers employed in this research included Random Forest (RF), Support Vector Machine (SVM), and See 5 (C 5.0). The results indicated that classification performance improved (increased the mean overall accuracy by 5%~10%, and reduced standard deviation of overall accuracy by around 1%) substantially with the training sample number, and when the training sample size was small (50 or 100 training samples), hybrid classifiers substantially outperformed single classifiers with higher mean overall accuracy (1%~2%). However, when abundant training samples (4,000) were employed, single classifiers could achieve good classification accuracy, and all classifiers obtained similar performances. Additionally, although object-based classification did not improve accuracy, it resulted in greater visual appeal, especially in study areas with a heterogeneous cropping pattern.

Reconstruction of Daily 30 m Data from HJ CCD, GF-1 WFV, Landsat, and MODIS Data for Crop Monitoring

With the recent launch of new satellites and the developments of spatiotemporal data fusion methods, we are entering an era of high spatiotemporal resolution remote-sensing analysis. This study proposed a method to reconstruct daily 30 m remote-sensing data for monitoring crop types and phenology in two study areas located in Xinjiang Province, China. First, the Spatial and Temporal Data Fusion Approach (STDFA) was used to reconstruct the time series high spatiotemporal resolution data from the Huanjing satellite charge coupled device (HJ CCD), Gaofen satellite no. 1 wide field-of-view camera (GF-1 WFV), Landsat, and Moderate Resolution Imaging Spectroradiometer (MODIS) data. Then, the reconstructed time series were applied to extract crop phenology using a Hybrid Piecewise Logistic Model (HPLM). In addition, the onset date of greenness increase (OGI) and greenness decrease (OGD) were also calculated using the simulated phenology. Finally, crop types were mapped using the phenology information. The results show that the reconstructed high spatiotemporal data had a high quality with a proportion of good observations (PGQ) higher than 0.95 and the HPLM approach can simulate time series Normalized Different Vegetation Index (NDVI) very well with R2 ranging from 0.635 to 0.952 in Luntai and 0.719 to 0.991 in Bole, respectively. The reconstructed high spatiotemporal data were able to extract crop phenology in single crop fields, which provided a very detailed pattern relative to that from time series MODIS data. Moreover, the crop types can be classified using the reconstructed time series high spatiotemporal data with overall accuracy equal to 0.91 in Luntai and 0.95 in Bole, which is 0.028 and 0.046 higher than those obtained by using multi-temporal Landsat NDVI data.

Potential of multitemporal Gaofen-1 panchromatic/multispectral images for crop classification: case study in Xinjiang Uygur Autonomous Region, China

Abstract. Gaofen-1 panchromatic/multispectral (GF-1 PMS) data have both high spatial and temporal resolutions, and this research aims at evaluating the potential of GF-1 PMS data for crop classification. Three PMS images (at days 110, 192, and 274) were acquired in Manas County of Xinjiang. The images were first segmented and all objects were then visually interpreted based on ground reference data. Some indices and textual features were then extracted at the object level. Subsequently, the Jeffries–Matusita (JM) distance was employed to estimate the class separability among all pair-wise comparisons of each time period. Afterward, a random forest algorithm was used to calculate importance scores of all features and classify crop types for every possible image combination. Additionally, to evaluate the influence of feature number on classification accuracy, features were added one by one based on the importance of scores. The result showed that GF-1 PMS images with high-spatial resolution had the potential to identify the boundary of the crop fields. Relatively high JM distance (above 1.5) and classification accuracy (above 90%) indicated that day 192 image contributed the most to the crop identification in the study area. For multi-image combinations, days 110 to 192 combination can achieve high overall accuracy (around 93%) and more images cannot substantially improve the classification performance. As for features, normalized difference vegetation index and near infrared (NIR) band had the highest importance scores and textual features contributed to distinguishing tree from crop land. Finally, classification accuracy increased together with the augmentation of feature number when only a few features were used. After accuracies reached saturation points, however, more features only slightly improved the classification performance.

Spatiotemporal changes of urban impervious surface area and land surface temperature in Beijing from 1990 to 2014

This study examined changes in urban expansion and land surface temperature in Beijing between 1990 and 2014 using multitemporal TM, ETM+, and OLI images, and evaluated the relationship between percent impervious surface area (%ISA) and relative mean annual surface temperature (RMAST). From 1990 to 2001, both internal land transformation and outward expansion were observed. In the central urban area, the high-density urban areas decreased by almost 7 km2, while the moderate- and high-density urban land areas increased by 250 and 90 km2, respectively, outside of the third ring road. From 2001 to 2014, high-density urban areas between the fifth and sixth ring roads experienced the greatest increase by more than 210 km2, and RMAST generally increased with %ISA. During 1990–2001 and 2001–2014, RMAST increased by more than 1.5 K between the south third and fifth ring roads, and %ISA increased by more than 50% outside of the fifth ring road. These trends in urban expansion and RMAST over the last two decades in Beijing can provide useful information for urban planning decisions.

Modeling Spatial Patterns of Soil Respiration in Maize Fields from Vegetation and Soil Property Factors with the Use of Remote Sensing and Geographical Information System

To examine the method for estimating the spatial patterns of soil respiration (Rs) in agricultural ecosystems using remote sensing and geographical information system (GIS), Rs rates were measured at 53 sites during the peak growing season of maize in three counties in North China. Through Pearsons correlation analysis, leaf area index (LAI), canopy chlorophyll content, aboveground biomass, soil organic carbon (SOC) content, and soil total nitrogen content were selected as the factors that affected spatial variability in Rs during the peak growing season of maize. The use of a structural equation modeling approach revealed that only LAI and SOC content directly affected Rs. Meanwhile, other factors indirectly affected Rs through LAI and SOC content. When three greenness vegetation indices were extracted from an optical image of an environmental and disaster mitigation satellite in China, enhanced vegetation index (EVI) showed the best correlation with LAI and was thus used as a proxy for LAI to estimate Rs at the regional scale. The spatial distribution of SOC content was obtained by extrapolating the SOC content at the plot scale based on the kriging interpolation method in GIS. When data were pooled for 38 plots, a first-order exponential analysis indicated that approximately 73% of the spatial variability in Rs during the peak growing season of maize can be explained by EVI and SOC content. Further test analysis based on independent data from 15 plots showed that the simple exponential model had acceptable accuracy in estimating the spatial patterns of Rs in maize fields on the basis of remotely sensed EVI and GIS-interpolated SOC content, with R2 of 0.69 and root-mean-square error of 0.51 µmol CO2 m−2 s−1. The conclusions from this study provide valuable information for estimates of Rs during the peak growing season of maize in three counties in North China.

Analyzing the Sensitivity of Crops Classification Accuracy Based on MODIS EVI Time Series and History Ground Reference Data

Abstract An improved spectral profile–based classification method was developed to discriminate corn, alfalfa, and winter wheat in the U.S. state of Kansas. Unlike other classification procedures, this method uses historical field reference data as training samples. An artificial immune network (AIN) algorithm, namely the artificial antibody network (ABNet), was tested as a classifier, combining historical field reference data and moderate-resolution imaging spectroradiometer (MODIS)-enhanced vegetation index (EVI) images. Historical field reference data from the years 2009 to 2012 were used to classify the three crops for 2013 data. A new method was developed to select the purest pixels from cropland data layer (CDL). Historical reference data were used in two different methods to classify crops in 2013: (i) single-year historical data and (ii) multiyear data used in four different combinations. Using method (i), classification was most accurate when the most recent year of training data was utilized. The accuracy of method (ii) increased with the number of years of data used for training the classifier. Results ranged from 81% to 92% overall accuracies, with the exception of the year 2012, where a severe drought created anomalous spectral profiles for all crops in the study area.

Using historical NDVI time series to classify crops at 30m spatial resolution: A case in Southeast Kansas

Most crop classification work use the ground reference data to training the classifier; but sometimes, the ground reference data cannot be obtained. In this paper, we tried to use the NDVI time series obtained during 2006 and 2013 to classify crop types in 2014 at 30 m spatial resolution. The experiment was conducted in Southeast Kansas, USA. Firstly, we extracted the NDVI time series using ground reference data between 2006 and 2013 from MODIS NDVI time series. Then, the composed Landsat NDVI data were transformed to MODIS NDVI using the linear correlation between the two data sets. Next, Random Forest (RF) was employed to classify crop types at 30 m resolution. The result showed that this procedure could accurately identify the major crops in the study area as the overall accuracy was 92.22% and the Kappa coefficient was 0.8758. In addition, two subsets of the study area showed that the result obtained in this study was similar to that of Crop Data Layer (CDL) provided by National Agricultural Statistics Service (NASS). Thus, the method proposed in this study could be an alternative way for crop classification when ground reference data cannot be acquired.