1. Introduction
Studies on the extent of rice areas provide useful information for food security, water resources management and environmental sustainability. Rice is the major food for nearly half of the world’s seven billion people, mostly in developing countries in Asia, Africa and Latin America [
1]. Rice agriculture is mostly irrigated and consumes 24%–30% of world’s developed fresh water resources [
2]. Therefore, determining the total rice area is an important input for the effective management of global water resources. On the other hand, the world’s rice production is not increasing significantly, and present annual rice demand exceeds annual production; thus, food security remains uncertain [
3]. Rapid urbanization, industrialization, changing patterns of precipitation and rising global temperature affect the land and water resources for rice production [
4]. Hence, it is urgent to monitor the rice area to meet the growing food demand, efficient water resource management and environmental sustainability.
Remote sensing (RS) has demonstrated its potential for rice area mapping employing either optical or synthetic aperture radar (SAR) images [
5,
6,
7,
8,
9,
10]. Time series multi-spectral and multi-temporal data from MODIS, NOAA-AVHRR, LANDSAT, IRS, SPOT and Chinese HJ-1A/B were used by many researchers at the national and sub-national scale for rice mapping [
11,
12,
13,
14,
15]. Though optical data cover a wide range of spatial and temporal resolutions, it is weather dependent, and cloud cover hampers its operation. To overcome this drawback, weather-independent SAR has been used [
9,
10,
16]. Unfortunately, the low temporal resolution and high price of SAR data limit its use at a large scale. A number of techniques have been applied to discriminate rice area, mainly the thresholding method [
17,
18,
19], the supervised classification method [
20,
21] and phenology-based mapping [
11,
22]. However, these methods have limitations. Thresholding requires appropriate thresholding values for accurate classification. Supervised classification requires training data for each year, which is a challenge for large area mapping. The phenology-based mapping requires a long and continuous time series. Moreover, all of these methods underperform in the areas where the crop fields are small.
Compared to traditional pixel-based classification, object-based image analysis (OBIA) considers a group of pixels instead of a single pixel for classification [
23]. The OBIA produces more meaningful and reliable classification by contributing additional information, including spectral, textural and geometric features [
24]. The OBIA includes two steps: image segmentation and classification. In segmentation, the entire image is partitioned into regions or objects that are more internally uniform and homogeneous than neighboring objects [
25]. The segmented image generates extra spectral, textural and geometric information of objects. In the classification process, each object is assigned to a specific class according to its spectral, textural, geometric and customized properties. Previous studies have demonstrated the successful use of OBIA in many applications [
24,
26,
27,
28]. Small fragmented rice fields are common in India [
29]; therefore, the OBIA approach is particularly useful by focusing on objects or parcels instead of concentrating only on the properties of single pixels.
Rice mapping for a large area is generally performed at low spatial resolution due to the insufficient availability of high resolution temporal data [
17,
19]. To obtain the temporal data at high spatial resolution, image fusion or blending algorithms are commonly applied combing high and low spatial resolution images. The fusion algorithm can predict accurate surface reflectance [
30] and has shown the potential of dense time series generation for phenology studies [
31].
Phenology refers to the timing of growing events of plants, such as bud-burst, leafing, peak growth stage, flowering and abscission [
32]. In recent studies, phenology has been used for crop classification [
11,
22,
33,
34]. However, the capability of phenology has not been investigated yet for the OBIA-based rice classification. This study focuses on the use of phenology for rice classification under the OBIA framework. Its main objective is to investigate the applicability of phenology for OBIA-based rice mapping.
2. Study Area
The study area includes five districts (Bongaigaon, Barpeta, Goalpara, Nalbari and Kamrup) of Assam in northeast India, with an area about 14,000 square km. The center of the area is located at 26°23′N and 91°09′E (
Figure 1). Assam is mainly the flood plain of two rivers: Brahmaputra and Barak. The region experiences heavy rainfall with an average of 3000 mm annually. Rice, the principal crop of the region, is cultivated during very wet summers, as well as in very dry winters. The area under rice cultivation is 25 million hectares, which is 71% of the total cultivated area [
35].
The crop is cultivated in many environmental conditions in the region based on the hydrological characteristics. The most commonly-practiced cultivation types are: (1) rainfed lowland rice, where the flooding of the fields is non-continuous and water depth varies around 1 cm–50 cm; the water availability of the fields primarily depends on the rainfall; (3) irrigated rice, where the water depth of the fields varies between 5 cm and 10 cm; the fields are kept constantly flooded throughout the growing seasons; two crops are harvested per year from these fields, and often, the production is high; (3) flood-prone rice, which is cultivated in deep-water areas with a water depth of 0.5 m–3 m; this rice is generally grown near the rivers and submerged frequently; the production is relatively low due to the effect of floods; and (4) upland rice, which is mostly cultivated on the hill slopes of the region; it is practiced in the rainfed conditions without the flooding and accumulation of water in the fields; the water requirement is low for this type of rice.
The region has two distinct growing seasons of rice: rabi (February/March–June/July) and kharif (June/July–November/December) (
Figure 2). Most farmers prefer to plant kharif rice due to the abundant rain and favorable temperature. Rabi rice is mainly distributed near the river and the lakes to take advantage of the captured or retained water of the monsoon rain or the irrigation infrastructure. The distribution of double-cropped rice is scattered on the north bank of Brahmaputra River, where irrigation is available. The selected study area has all of the rice types, and all were considered for the mapping.
6. Discussion
The integration of crop phenology has advantages in the rice crop classification. The grassland and the other crops that are spectrally similar to rice often create confusion for rice discrimination. Moreover, among the rice fields, variability exists due to the differences in crop planting time, local weather conditions and other factors [
24]. In this study, classification was performed on spectral data with and without using the phenology. The main differences in the results of the two approaches were the misclassification of grasslands and crop types. Using only the spectral features, the grasslands and other crops were misclassified as rice. The reason for the misclassification was that both the grassland and other crops had spectral properties similar to those of rice. On the other hand, the use of phenology along with the spectral features improved the classification significantly. The main reason behind the improvement was that the different crop types had different seasonal behavior, which is captured through the phenological variables. Additionally, the five phenological variables (base value, largest value, seasonal amplitude, large and small integral) are actually the statistical values of NDVI, meaning that they are calculated either by integrating, averaging, differencing or taking the maximum value of annual NDVI. These values are not dependent on the crop planting and harvesting date, thereby improving the classification by better capturing the crop growth profiles.
Discriminating a specific crop from the various vegetation types using a single image is a challenge [
12,
37]. However, the images at the key crop growth stages effectively improve the classification accuracy. In the study area, the agricultural fields are generally small, fragmented and irregular. Therefore, a stepwise method has been established to identify the rice fields in a heterogeneous landscape using an object-based image analysis method.
This study demonstrates the applicability of multi-source remote sensing data (HJ-1A/B and MODIS) to accurate mapping of rice. The rice usually grows during the rainy season in a humid tropical climate where obtaining multi-temporal cloud-free optical images for the whole growing season is a constraint. This constraint becomes particularly critical when considering a large area. Therefore, optimal use of available cloud-free images is of great importance. This study demonstrates how the images from different sensors at peak crop growing stages help with achieving accurate rice classification. Additionally, the use of multi-source images increases the chance of getting more cloud-free images for the classification.
The top-down hierarchical classification approach yielded several advantages to discriminate rice fields. The general classes (e.g., water, forest, built-up, etc.) were classified by applying thresholds and then removed from further processing. The threshold condition for any class was adjusted to achieve higher classification accuracy without affecting the condition for the other classes. Along with the threshold condition, the nearest neighbor classification has also been applied for the initial classification of rice. The nearest neighbor classification considered the statistical distribution of the classes. It helped to identify the classes that could not be classified by the threshold condition. Moreover, the applied top-down hierarchical approach was logical to discriminate the rice class.
Use of object-based image analysis showed an advantage in identifying fragmented rice fields. Fragmentation of cropland is common in India [
29]. During the period between 2001 and 2010, the operational field size has reduced from 1.33 Ha down to 1.15 Ha [
58]. Land is becoming fragmented due to the absence of land use planning, rapid population growth, economic development, urbanization and the limited availability of arable land [
13]. This study demonstrated that the fragmented rice fields were well identified by the use of high resolution (30 m) HJ-1A/B images under the object-based image analysis framework.
In the recent study of Tian
et al. [
59] and Jia
et al. [
60,
61], the number of base images used to produce the synthetic time series was much lower than the number used in this study, with six images in 12 years and one image in one year, respectively. In this study, three HJ CCD images of key rice growth stages were used as a base image for the ESTARFM fusion. When using different base images in the fusion process, the reported difference is minor between the synthetic and the actual images on the same date [
59,
62]. Therefore, in spite of the three base images in our study, the accuracy of the blended time series was not significantly decreased. The base images were acquired at the beginning, middle and the end of the growing seasons, which helped the ESTARFM to capture the reflectance changes caused by phenology, and the synthetic time series reflected the actual changes in the NDVI trend [
30]. Additionally, the index-then-blend approach was adopted, which produces more accurate synthetic images [
46]. The adopted fusion technique yielded accurate synthetic NDVI. However, it is advised to use dense time series base images in order to achieve high fidelity in the synthetic images generated by ESTARFM.
Earlier works on rice mapping methods using optical images can be categorized into three groups. The first group considers individual cloud-free images and uses image statistics approaches, e.g., the unsupervised classifiers, like the self-organizing data analysis technique (ISODATA) [
63,
64], and supervised classifiers, like maximum likelihood [
65] and support vector machine [
66].
The second group considers time series images and uses different algorithms, like threshold-based and spectral matching techniques (SMT) [
7,
67]. The third group is phenology- and pixel-based paddy rice mapping (PPPM) [
15,
17], where rice is identified for individual pixels based on the flooding signals of the rice transplanting phase by evaluating the differences between the Enhanced Vegetation Index (EVI)/NDVI and Land Surface Water Index (LSWI). The first two groups often generate maps that are difficult to compare for different regions, working groups and years, primarily due to the spectral heterogeneity, training sample selection, post-classification processing and the capabilities of the image interpreter [
68]. In comparison, the PPPM methods of the third category are less affected by these issues. The phenology-based methods can capture the growing stages of different crops and identify the unique signals according to crop calendar and management activities [
34]. In this study, rice growth stages were extracted by the analysis of 16-day MODIS NDVI and HJ CCD datasets, and they were used to improve the spectral-based mapping of rice. This study provides a rice mapping method combining phenology and multispectral satellite data (
Table 6).
To apply the proposed approach to other climatic regions and large regions, a few points need to be considered carefully, such as the appropriate selection of threshold conditions; parallel processing or cloud-computing is suggested for effective handling of large datasets [
71]; and if possible, more high resolution images should be used. Increasing availability of medium and high spatial resolution satellite sensors like SPOT 6, Landsat 8, Sentinel-2 and G-F1 is increasing the observation frequency. An increase in good quality observations would increase the efficiency of the methodology, especially in the tropical regions, where cloud cover is common [
72].
The identification of rice fields was affected by several potential factors. The problem of mixed pixels remains an issue. Some rice patches were too small to be identified with 30-m spatial resolution HJ-1A/B data. The widely-used, subtraction-based method could not be applied due to the unavailability of the SWIR (shortwave infrared) band in the HJ CCD sensor. The limited availability of cloud-free images for the whole growing season affected the rice identification. The vegetation dynamics could be better captured by using additional HJ-1A/B images during the sowing and harvesting period of rice. However, the use of phenology significantly reduces the necessity of the hyper-temporal image requirement for rice classification.