1. Introduction
The increasing extreme weather in the context of global climate change has posed severe threats to the function of living environments and the health of humans [
1]. Among these threats, floods have become the most common type accounting for 44% of natural disasters and affecting 1.6 billion people globally from 2000 to 2019 [
2]. In addition to the increased proportion of the population exposed to floods [
3], agricultural production activities tend to be another relatively vulnerable sector in response to damaging floods. For example, the United Nations reported that 70% of global rain-fed agriculture and 1.3 billion people relying on arable lands are exposed to the threats of shifting rainfall patterns and larger precipitation variability [
2]. The impact of flooding on crop yield varies dependent on flood characteristics (i.e., frequency, duration, depth, seasonality), crop types (tolerance of excess water and anaerobic soil conditions), and emergency activities [
4]. Apart from fluvial flooding, flash floods primarily caused by heavy rain events are more dangerous as they occur suddenly and hinder early warning and emergency response. Coastal flooding may cause similar scale and immediate property damage, while the influence of salt deposition from seawater on agricultural soils could be persistent for years. In Asia, the primary crop growing season such as rice, highly overlap with the monsoon season, resulting in serious storm-related crop loss in extensive vulnerable floodplains [
5], for example, in Bangladesh [
6,
7], Cambodia [
8], Thailand [
9] and China [
10], etc. Without effective strategies to respond to the increasingly frequent flood disasters, the goal of poverty eradication and food safety becomes difficult to achieve.
Satellite remote sensing provides a spatially and temporally explicit framework for detecting flood extent [
6,
11,
12,
13,
14], measuring flooding severity [
15,
16,
17], and assessing flooding hazard and post-flood loss and recovery [
18,
19]. Among them, extensive studies have been working in remote sensing-based water body extraction [
14,
20,
21,
22,
23,
24,
25] since it is the basis of flood monitoring and management. Compared with other applications requiring water extraction techniques such as wetland mapping [
26,
27,
28] and coastline extraction [
29,
30,
31]), flood detection confronts more challenges, including cloudy and rainy weather conditions. Beyond that, a flood occurs suddenly in general, and the flood recession process varies under the control of rainfall and surface conditions. Observations with a long revisit period would overlook short-lived flooding and lead to underestimating the hazard level. Thus, the higher spatial and temporal resolution of satellite observations is pivotal in flood monitoring.
Two types of satellite data primarily map the inundation area, the synthetic aperture radar (SAR) images and optical sensor images [
3,
17]. The active microwave remote sensing displays its advantage with the character of working day and night and passing through clouds in flood monitoring. SAR-derived information including backscatter intensity, polarimetric parameters, and interferometric coherence, is commonly used for water inundation classification [
21]. Optical sensor is more often used in pre- or post-disaster flood extent extraction due to cloud contaminations [
3]. Besides, its ample spectral information enables identifying flood-affected land cover types. Methods of flood extent extraction can be categorized into pixel- and object-based, or from other perspectives, into supervised and unsupervised approaches [
21]. Traditional segmentation methods such as thresholding, edge detection, active contour model and region growing always use specific bands (i.e., VV and VH) or water indices (i.e., NDWI, MNDWI) to extract water body. Most of these methods require parameter selection which is particularly complicated and uncertain for time-serious monitoring of rapidly changing flood extent. In recent years, machine learning (ML) techniques (i.e., the Random Forest Classifier, the Support Vector Machine) have been applied in flood monitoring [
11,
32] with the benefit of combining multi-source features. Also, deep learning (DL, i.e., Convolutional Neural Network) is increasingly developed in this field to achieve rapid and more accurate flood mapping without data annotation [
33,
34]. However, the computation cost and efficiency remain concerns for large-scale flood detection. There are also amounts of research integrating digital elevation data into hydrodynamic or conceptual models to predict flood inundation [
35]. In a data-rich environment, the integration of high spatial resolution imagery with flood inundation modelling allows robust and temporally consistent flood propagation simulations. In addition, remote sensing products including land cover datasets, global surface water [
36], cropping intensity maps [
37], and social media data [
38] are also important auxiliary information for flood detection and subsequent hazard assessment. Adequately combing data from both Radar and optical systems and other “Big Data” systems is of great potential in future flood assessment [
3].
There are a wide range of remote sensing applications centered on flooding mapping. At a global scale, the UNU-INWEH developed the World Flood Mapping Tool based on the Google Earth Engine with decades of Landsat data since 1985, which provides the first comprehensive historical flooding information globally [
39]. The automatic near real-time flood detection software using Suomi-NPP/VIIRS data has been running routinely since 2014 [
40]. These map products including daily products and post-processed 5-day composite flood extent datasets at the 375 m resolution, are updated routinely, supporting monitoring global outburst flood. Nardi et al. [
41] published a global high-resolution dataset of Earth’s flood plain at a 250 m resolution with the Shuttle Radar Topography Mission (SRTM) digital terrain model using the geomorphic algorithm GFPLAIN [
42,
43]. Such maps can support large-scale risk management studies, including assessing human exposure to potential flooding [
41] and categorizing flood insurance levels for different human settlement areas. Several attempts have been made to reveal large-scale annual and interannual flooding patterns for some flood-prone areas such as the Mekong Delta and the Ganges Delta [
6,
7,
44]. In addition, many studies focus on specific extreme flood events at local scales [
16,
17,
32]. With abundant flooding products derived from space observations, understanding detailed flooding progress and human reactions to natural disasters is insufficient and more pressing.
In rural areas which inhabit more poor households and exist large-scale fertile farmlands, the flooding condition is highly correlated with dwellers’ livelihood and food security. Although remote sensing data have been applied on crop loss assessment [
45], the contributions of adaptation measures such as rush planting have been limitedly considered. At present, flood crop loss assessment could be categorized into the flood-intensity-based approach and crop-condition-based approach. For the flood-based methods, most researches stimulate flood conditions (i.e., inundation depth) and evaluate crop loss using flood damage functions [
45]. In this way, the natural ability of crops to withstand floods is the preoccupation. The crop-condition-based approach is mainly achieved by comparing pre- and post-flood vegetation conditions and establishing relations between vegetation index and crop yield. Similarly, researchers hardly take account of the phenomenon that Vegetation Index (VI) profiles decrease followed by the subsequent increase in the disaster year. To date, few researchers apply remote sensing technology to highlight the effects of rapid post-flood recovery activities initiatively conducted by local residents, which is indispensable for comprehensive damage evaluation.
In the 2020 summer monsoon season, the unusually heavy and continuous precipitation caused extensive flooding in many Asian countries, resulting in serious losses of lives and properties [
46]. According to the Ministry of Emergency Management of the People’s Republic of China, the 2020 flood season in southern China has been recognized as the most severe flood situation since 1998. July’s heavy rainfall and flooding event in the Yangtze and Huai River basins, which affected 34.173 million people in 11 provinces and caused 132.2 billion yuan economic losses, has been officially selected as one of the top 10 natural disasters in China 2020. As important grain-cultivating areas in southern China, Yangtze-Huai River basin were exposed to immense pressure in agricultural production. Nevertheless, local residents and agricultural technicians implemented active and timely remedial measures adjusting to different types of flooded croplands. Hence the flooding disaster losses were successfully mitigated [
47].
In this paper, we proposed a method by combining SAR and optical imaging to detect flood-affected cropland and monitor spontaneous agricultural recovery during and after the flooding disasters. Specifically, we chose three experimental sites in Jiangxi and Anhui Province which were seriously affected by the 2020 summer flood to test the proposed methodology pipeline, with the aim to answer the following scientific questions: (1) What is the spatial distribution of flood-affected croplands during the 2020 monsoon season? (2) how to determine the timing and extent of recovered agricultural productions? and (3) what are differences in local characteristics between recovered and unrecovered agricultural productions?