1. Introduction
Fishery in Bangladesh has been increasing rapidly in the last few decades as a major source of food and economic growth [
1]. According to the International Food Policy Research Institute (IFPRI), the fish farming market has grown 25 times in all aspects of the aquaculture industry in the last three decades. Shahin et al. [
2] also reported that the total fish production increased from 4.99 Lac MT (100,000 metric tons) in 1998–1999 to 14.47 Lac MT in 2012–2013. Though rice is still the major food source for Bangladesh, the booming aquaculture is bringing diversity to the dietary structure of people in Bangladesh and gradually improving people’s health conditions [
3]. However, the growing aquaculture puts pressure on already limited croplands. In recent years, a great portion of croplands in Bangladesh has been gradually transforming to other land use types, such as fishponds, brickyards, and residential area [
1,
4]. While Bangladesh Statistical Bureau (BSB) publishes statistical yearbooks every year, it lacks information of the spatial distribution of the land use changes. Such information can better help decision-makers make land use policies for better resource distributions. With earth observation (EO) data, especially newly published Sentinel-2 Multispectral Instrument (MSI) 10 m resolution images, mapping and monitoring individual fishponds become feasible. In addition, the advent of Google Earth Engine (GEE) significantly reduced the workload and time of remote sensing data preprocessing, analysis, and visualization [
5]. Thus, it is important to investigate the potential of using Sentinel-2 images and GEE platform for fast, timely mapping of inland fishponds in Bangladesh.
There are many research works focus on mapping aquaculture ponds in coastal area of South Asia [
6,
7,
8]. However, inland fishponds differ from coastal aquaculture ponds in that inland fishponds are typically owned by individual families, which means they can have arbitrary shape and size and are not necessarily well-aligned as many coastal aquaculture ponds do. Some key features of fishponds in Bangladesh are that:
- (1)
they are usually filled with water all year round,
- (2)
they are small, and
- (3)
like many other man-made objects, they have regular boundaries and simple shapes, such as rectangles.
To address feature (1), multi-temporal and multi-spectral remote sensing images should be used. High-resolution images are the most suitable data to address feature (2). Specifically, based on the research conducted by Belton and Azad [
9], the average size of homestead fishponds in Bangladesh is between 0.08 to 0.1 ha, the median value can be even less due to the skewness towards a few large fishponds, which can go up to over 100 ha each. Fishponds with such small size are challenging to detect on medium-resolution (2–30 m) remote sensing images, and it is almost not feasible to do with low-resolution (>30 m) images. However, high-resolution images, such as SPOT (Satellite Pour l’Observation de la Terre) and IKONOS, are usually not available free of charge. Sentinel-2 MSI L1C data has become increasingly popular for land use and land cover (LULC) mapping in recent years mainly because of its finer spatial resolution (10 m) and temporal resolution (10 days before Sentinel-2B launches and 5 days after) [
10,
11,
12]. The significant improvement of both spatial resolution and temporal resolution offered great potentials of improving existing applications and enabling new missions, such as object detections [
11]. Therefore, this research uses Sentinel-2 MSI L1C data for fishpond mapping.
Water indexes (WI), such as Normalized Difference Water Index (NDWI) [
13], Modified Normalized Difference Water Index (MNDWI) [
14], and Automated Water Extraction Index (AWEI) [
15], have been developed to enhance water features on multi-spectral images. Most of the WIs utilize low reflectance of water in near-infrared (NIR) and shortwave-infrared (SWIR) spectrum [
12,
13,
16,
17]. NDWI takes the difference between the green band and the NIR band, which produce positive values for water and negative values for other LULC types [
13]. To address false positives from built-up using NDWI, Xu introduced MNDWI, which is calculated with green and SWIR bands [
14]. Previous research reported that MNDWI generally has a more stable threshold than NDWI [
16,
18,
19]. Aside from the NDWI and MNDWI that use 2 bands to compute, AWEI uses 5 bands to compute, and it aims to reduce false positives coming from shadow pixels [
15]. The AWEI consists of two formulas, AWEI
sh for areas that are contaminated by shadows, and AWEI
nsh for areas that are not [
20]. Feyisa et al. reported that AWEI has much more stable optimal thresholds than MNDWI [
15].
To address feature (3), shape-based metrics can be used to help characterize fishponds and non-fishponds. Object-based features (OBF), also referred to as geometrical features in Reference [
21] and as shape metric in Reference [
22], have been used in previous research as ancillary features in object-based image analysis. In a research conducted by van der Werff and van der Meer [
23], shape measures were extracted from Landsat image objects and were used to help classify spectrally identical objects, e.g., rivers and different shapes of lakes. Jiao et al. [
22] used 10 shape metrics to classify 8 LULC classes, including rivers and ponds, on SPOT-5 images. Their results showed that such metrics can well characterize all LULC classes quantitatively. For water features specifically, they characterized rivers as elongated, concave, and complex, while ponds were round, rectangular, convex, and simple [
22]. Both research works reported that OBFs significantly improved classification accuracy especially when objects are spectrally similar. However, previous research typically works with high-resolution images or objects that are large compared to the pixel sizes, so it is unclear how such OBFs will help with classifying small objects on relatively coarse resolution images.
Therefore, to investigate the performance of using WI and OBFs for inland fishpond mapping, this research aimed to develop a GEE-based workflow that incorporates spectral-based filtering with multiple WIs and spatial-based filtering with OBFs for fishpond mapping. To our knowledge, this is the first study that presents fully automated workflow for inland fishpond mapping. A case study in the Singra Upazila of Bangladesh was conducted to evaluate the performance of the workflow.
2. Study Area and Dataset
The study area of the case study we chose is the Singra Upazila (24°30′ N 89°08′ E) in Bangladesh, as shown in
Figure 1. It is a sub-district of Natore district in Northern Bangladesh that consists of 13 unions, namely Chamari, Chaugram, Chhatar Dighi, Dahia, Hatiandaha, Italy, Kalam, Lalore, Ramananda Khajura, Sherkole, Singra Paurashava, Sukash, and Tajpur. More than three hundred thousand people live in around 530 sq. km areas with a density of 607 persons per sq. km. Around 80% of people in this area are engaged with agriculture, more specifically rice crop farming. Since this area is located within one of the largest flood plains of the country, most of the agricultural fields are flooded in the rainy season every year. A recent trend of land use change from crop fields to fishponds has been found in this area because of the larger profit of fish culturing than growing rice.
In addition to the Singra upazila, where fishponds are located at, we also chose a subarea in Tibetan Plateau (upper-left corner: 84°52′ N 31°49′ E; lower-right corner: 87°49′ N 29°55′ E) to select non-fishpond water features to train a classifier for fishpond classification. Tibetan Plateau has an average elevation of over 4000 m, and the plateau is rich of water resources with over 1500 lakes scattered on the plateau [
24]. There are three reasons why we selected a region on Tibetan Plateau for non-fishpond sample selection. First, Bangladesh does not have many permanent water bodies for classification. Second, Tibetan Plateau has over 1200 lakes larger than 1 sq. km, as well as multiple river streams [
25]. Third, Tibetan Plateau has over 4000 m average altitude, which significantly limited human activities, such as fishery. Therefore, there are rarely any artificial water features in the area.
The dataset used in this study is the Sentinel-2 MSI Level-1C product hosted on GEE. It was preprocessed by radiometric and geometric corrections. As a result, the Sentinel-2 Level-1C is a Top-of-Atmosphere (TOA) reflectance dataset that consists of 100 km by 100 km image tiles projected in UTM/WGS84. The Multi-spectral instrument (MSI) sensors onboard Sentinel-2 satellites collect images with 13 spectral bands in the visible/near-infrared (VNIR) and SWIR spectrums, and the spatial resolutions of the spectral bands vary from 10 m to 60 m. The bands that were used in this study are listed in
Table 1. The Military Grid Reference System (MGRS) tile number for all images used in this study is 45RYH.