1. Introduction
Agriculture represents one of the major fields in remote sensing applications. Precise crop mapping is needed at various scales for users from individual agricultural companies to national and international entities. Crop mapping has therefore become one of the key remote sensing applications in agriculture. Due to the high diversity and complexity of agricultural landscapes, accurate identification of crop types is challenging. Moreover, agricultural fields are often divided into multiple cropping patches with different varieties or even different crops. Therefore, mixed targets are commonly seen in an image pixel when the spatial resolution does not allow for separating the patches. This is especially true in China because farmers have parcels of 0.2 ha on average. The number of mixed pixels decreases as the spatial resolution of satellite data improves. However, a high spatial resolution also implies increased costs and a smaller area being covered by one image. In addition, it is crucial to acquire frequent images to accurately differentiate the spectral profiles of different crops across the season [
1]. All in all, developing precise crop mapping methods is still a challenging task in remote sensing for agricultural monitoring.
Crop type classification with satellite image time series is an important tool for precise and timely crop mapping. A number of open-source, cloud-based tools have been developed for crop type classification in recent years. The Sen2Agri system is one such open-source tool available for free download, allowing users to generate near-real-time products tailored to their needs at their own premises or using cloud computing infrastructure [
1]. A variety of crop classification methods at different scales and with various levels of accuracy can be found in the literature [
2,
3,
4]. Meng [
5] collected all the available cloud-free Sentinel-2 multispectral images for winter wheat and rapeseed growth periods in the study area in southern China and used the random forest (RF) method as the classifier to identify the optimal temporal window. Song [
6] presented an approach integrating object-based image analysis with RF for mapping in-season crop types based on multitemporal GF-1 satellite data at a spatial resolution of 16 m. Song [
7] proposed a method to examine a total of 13 spectral variables from GF-1 images and three classifiers, the Bayesian discriminant (BD), Mahalanobis distance (MD), and RF, for classification of six LULC types at East Dongting Lake, Hunan, China. Li [
8] utilized deep learning-based frameworks, including DenseNet, ResNet, VGG, SegNet, and DeepLab v3+, for cotton crop field identification with GF-1 high-resolution (16 m) images. Fan [
4] reported on crop type classification with high-resolution satellite data with a wide field of viewer (WFV) onboard GF-1, the Multispectral Instrument (MSI) onboard Sentinel 2 (S2), and the Operational Land Imager (OLI) onboard Landsat 8 (L8), and found that these satellite data may be used for crop type classification within the growing season with very good accuracy if the training datasets were well-tuned. A combination of S2 and L8 for classification has been confirmed to be suitable for crop type classification [
9,
10,
11,
12].
Low-resolution satellite data have been applied to agricultural monitoring in the past three decades. It has been commonly understood that the spatial resolution of satellite data is one of the key factors determining the quality and accuracy of their applications for crop monitoring. To improve the accuracy of crop monitoring, the spatial resolution of several types of EO data has been steadily evolving. European satellite, the Project for On-Board Autonomy–Vegetation (PROBA-V) [
13], brought the world a new data source for agricultural monitoring. PROBA-V data has a spatial resolution of 100 m and the ability to cover the entire surface of the globe. Moreover, the data may be downloaded free of charge. In addition, PROBA-V provides satellite data with 300 m and 1 km pixel scales under nadir. The Chinese meteorological satellite provides 250-m-resolution data globally and free of charge since late 2008 [
14]. FY-3 MERSI [
15] is a Medium-Resolution Spectral Imager (MERSI) carried aboard the third FY (“Wind and Cloud”) series of meteorological satellites developed in China. FY-3A/B/C MERSI has 20 spectral bands, five bands in 250 m, and 15 bands in 1000 m, while FY-3D MERSI has 25 spectral bands, six in 250 m, and 19 bands in 1000 m. The 250 m bands of FY3 MERSI are similar to PROBA-V in the visible and near-infrared.
PROBA-V supports applications such as land use, worldwide vegetation classification, crop monitoring, famine prediction, food security, disaster monitoring, and biosphere studies. The 100-m PROBA-V data has been widely used for landscape classification. Eberenz [
16] demonstrated land cover mapping over West Africa with PROBA-V 100 m time series data for the 2014–2015 season, using temporal metrics and cloud filtering in combination with in situ training data and machine learning, implemented on the ESA Cloud Toolbox infrastructure. Lambert [
17] proposed a fully automated classification method to deliver the first cropland map at a 100 m scale for the Sahel and Sudan region using PROBA-V images with an overall accuracy of 84% and an F-score of 74%. Durgun [
18] developed a method inspired by spectral matching techniques (SMTs) and based on phenological characteristics of different crop types using 100-m PROBA-V normalized difference vegetation index (NDVI) data from the 2014–2015 season for crop area mapping. Zhang [
19] studied crop classifications based on Iterative Self-Organizing Data Analysis Technique (ISODATA) clustering, the maximum likelihood method (MLC), and a similarity analysis. Roumenina [
20] assessed the crop mapping performance provided by the PROBA-V 100 m data in comparison with coarser-resolution data (e.g., PROBA-V 300 m data) in Bulgaria with the maximum likelihood classification (MLC) and Iterative Self-Organizing Data Analysis Technique (ISODATA). Dimitrov [
21] presented the results of a subpixel classification of crop types in Bulgaria from PROBA-V 100 m NDVI time series using two subpixel classification methods, an artificial neural network (ANN), and support vector regression (SVR). Moreover, VITO [
22] developed an operational Mission Exploitation Platform (MEP) to drastically improve the exploitation of the PROBA-V Earth Observation (EO) data archive. FY-3 MERSI data were also made available through the research cooperation [
14] and used to make the crop type map in the North China Plain with the decision tree algorithm [
23] as well as to monitor the crop growth [
24,
25,
26,
27,
28,
29,
30,
31,
32,
33,
34,
35].
The recent development of high- and coarse-resolution satellite data in Europe and China provides an opportunity to investigate the potential of using both types of data for precise crop mapping at a regional scale. In this paper, we present the results of the Dragon 4 project 32194. The aim of the project is to evaluate the potential of crop mapping with the combined use of European and Chinese satellite data. It was divided into two sub-projects, namely, crop mapping with (i) high-resolution Sentinel-2 and Gaofen-1 data and (ii) medium-resolution PROBA-V and FY-3 MERSI data. The main objectives of the first sub-project were to (i) compare S2 and GF-1 time series data in a perspective of crop type mapping, (ii) develop a crop type mapping method based on both high-resolution satellites time series, (iii) validate and extend the crop mapping approach of ESA Sen2Agri in an irrigated area in northwest China, and (iv) improve GF-1 satellite data processing. The aims of the second sub-project were to (i) develop a method for crop mapping with both PROBA-V and FY-3 MERSI data and (ii) improve FY-3 MERSI data processing and product generation.
4. Discussion
In the big data era, various kinds of satellite data are increasingly made easily and/or freely available. These satellite images have become rich data sources for crop type mapping with machine learning algorithms. As an open sources tool, the Sen2Agri system has the user community benefiting from the automatic downloading, processing, and applying of sentinel satellite data for the crop type classification, but the Chinese GF satellite data is not yet ready to be explored in this system. After the implementation of the Sen2Agri system in this study area, we observed that training datasets, input features, and classifier algorithms are three key factors that determine the quality and accuracy of the classified results. Thus, we made improvements for our approach of crop type classification with GF-1 and S2 in this study to quickly achieve peak accuracy.
We observed that some studies paid more attention to the evaluation and comparison of various classifier algorithms [
48,
49,
50], while others paid more attention to the selection of [
5,
46,
51], or adding more [
52,
53], input features. In general, the RF did perform better than other conventional classifiers, like Maximum Likelihood, Neural Network, Support Vector Machine, etc. [
49]. The algorithms of various deep learning methods also perform well, but the computation time for tuning and implementing is often unexpectedly long for the crop type classification of large areas. We followed the Sen2Agri system approach in our study and used the RF [
54] as the classifier in this project. We did not intensively test other conventional classifiers and deep learning algorithms since the RF worked well in our study area. Although, there are many options to create and increase the features from the input satellite images, we found in our studies that the classification of all NDVI-like indices combined with the input bands gained higher accuracy than those that used the same spectral bands and only a few selected indices, like NDVI, NDWI, and brightness, that the Sen2Agri system used [
1]. The studies [
52,
53] and the review on crop classification [
55] also concluded that the use of vegetation indices improves classification performance, but there is no conclusion which indices have positive contributions on the classification. Therefore, we recommend using all proposed indices if the computation is allowed since RF is able to deal with high dimension data arrays and neglect the duplications of input data by the random selection of input features.
The preparation of the training dataset is the most time-consuming task in the process of classification. A field survey is often necessary for crop type classification, and the field data collection may help the experts gain the experience of visually interpreting the satellite images. The spatially even distribution and statistical balance of the training datasets for setting up the model should be seriously taken into account. It is difficult to achieve the spatially even distribution for the field samples due to access to fields and time consumption, but this can be improved at home in the process of visually interpreting training samples on the images with a grid frame. For the majority of crop types, training samples may be identified as numerously as possible. However, the minority crop types are difficult to identify, and these classes are poorly represented in the training dataset. This results in a statistically imbalanced dataset for the training sample. In some cases, we had better group some classes to increase the representativeness. This kind of problem can be found after we check the input images and the result from the first-round classification. The impact reduction with the imbalanced dataset for building the classification model is one of the main concerns in the field of machine learning. The effects of imbalanced training samples for the classification in this study were neglected and should be investigated in further studies. There is still another question without a clear answer, which is how many training samples are optimal for the classification. At this moment, what we can do is to collect training samples as numerously as possible.
In this study, the classification was separately conducted with each individual data source. The accuracy of the resulting classification reflected individual performances of targeted satellite data in the given growing season. For optical satellite images, the cloud is a major issue hampering the application. This study did not investigate the data fusion or simulation to make two similar satellite data sources compatible, and the time series can be increased to have more cloud-free images for better classification. In a further study, the joint application of similar satellite data should be enhanced.
5. Conclusions
Project 32194 under the Dragon 4 Program was successfully executed with a focus on crop type mapping, using high- and medium-resolution European and Chinese satellite data for YERID, Ningxia, northwest China. First, the Sen2Agri system was conducted in the study area so that the collection and processing of S2 satellite data were fully benefited from the system. The results demonstrated the good performance of the Sen2Agri system in the fully irrigated area of Ningxia. The medium-term report [
27] revealed that nine types of crops were classified, and the crop type map in 2017 was produced based on 35 S-2A/B images. The OA of the crop type map was high, up to 88%. Second, a further study was conducted with the increased input features of GF-1 WFV, S2, and other third-party data after the training dataset was well-tuned with expert knowledge and the ground truth samples. The results showed that crop type mapping with any of these satellite data types could achieve acceptable accuracy. The lowest OA in the tests was 94%, high enough to be acceptable. The relatively lower accuracy with the GF-1 WFV data was due to the limited spectral bands. Third, crop classification with medium-resolution satellite data, 100 m PROBA-V, and 250 m FY-MERSI data was implemented. The preliminary results demonstrated the promising crop assessment capability using 100 PROBA-V or 250 m FY-3B MERSI data as the medium-resolution satellite data produced crop type maps with reasonable accuracy at a regional scale. Classification with these data may produce crop type maps early in the season, as desired by many users. With a very high revisit rate (twice a day), the medium-resolution EO satellite can offer more valid optical satellite images for various applications, including agricultural monitoring.
In this study, Random Forest (RF) was used as the classifier, but the training datasets and input features were paid more attention and finally improved after we gained the experience from the Sen2Agri system. The results proved again that the accuracy of crop type mapping increases with the number of input features used for the classification. The acceptable and peak accuracy of the crop type map was achieved after all the special bands, and the potential associate indices (NDVI like) were jointly used as input features for the classification. The spatially even distribution and the statistical balance of the reference sample are of importance for building a classifier model. Classifications based on individual images result in varying accuracies due to the limited signatures for correctly identifying the crop types on each image. Satellite image time series, used together as the input, is the best option to produce a good classification with high accuracy because such use would ensure that all available information for the whole growing season is involved in the classification for crop type mapping.
Finally, crop type mapping is not only useful for agricultural production management. The classified images or crop type maps may also be very useful for purposes such as environmental studies and irrigation management. Results from coarse resolution crop type classification may, for example, be helpful for the quick running of a crop-specific yield model or pest diseases forecast. Irrigation management is of great importance in Ningxia. Information from the crop type mapping is required to accurately compute the water demand for irrigation in specific growing stages of crops at a regional scale. The methods developed in this study may contribute to making better water management decisions so that the water use efficiency in the area can be significantly improved. Moreover, precision agro-meteorological services also require crop type maps at a regional scale to create better agrometeorological forecasts for the various crop growing stages. In addition, crop type mapping can also provide useful information for early warning of potential agricultural or meteorological disasters occurring at a regional scale. The combination of high- and coarse-resolution satellite data for crop type classification would thereby be useful for these research and operation services. With this understanding and experience, we are able to apply the required crop type mapping practices in various ways. Therefore, our understanding and experience of crop type mapping with high- and medium-resolution satellite data have been widened and sharpened through the implementation of the Dragon 4 Program project.