1. Introduction
Irrigation is one of the key agricultural management practices of crop cultivation in the world [
1,
2,
3,
4]. Irrigation reduces adverse effects of drought, increases crop yield, and finally maintains a good agricultural production profit. Irrigation consumes a lot of water resources and thus efficient water use management requires timely irrigation information in large regions [
5]. Irrigated crop land, irrigation events and irrigation water amount are provide important information in the support of sustainable water resource management. Studies on hydrology [
6], water availability and water use [
7], and their interaction with agricultural production and food security [
8], all require accurate information on the location and extent of irrigated croplands. Detailed knowledge about the timing and the amounts of water used for irrigation over large areas [
3] is also of importance for various studies and applications.
Irrigation practice is traceable on satellite images [
9]. A few global irrigation maps such as the Global Map of Irrigated Areas (GMIAs) [
10] and the Global Irrigated Area Map (GIAM) [
11] have become available. Recently, Wu [
12] retrieved a 30-m resolution global maximum irrigation extent (GMIE) using the Normalized Difference Vegetation Index (NDVI) and NDVI deviation (NDVIdev) thresholds in the dry and driest months. Zajac [
13] derived the European Irrigation Map for the year 2010 (EIM2010) underpinned by the agricultural census data. Siddiqui [
14] developed irrigated area maps for Asia and Africa regions using canonical correlation analysis and time lagged regression at 250 m resolution for the year 2000 and 2010. Zhang [
15] produced annual 500-m irrigated cropland maps across China for 2000–2019, using a two-step strategy that integrated statistics, remote sensing, and existing irrigation products into a hybrid irrigation dataset. Zhao [
16] developed crop class based irrigated area maps for India using net sown area and extent of irrigated crops from the census and land use land cover data at 500 m spatial resolution for the year 2005. Ambika [
17] developed annual irrigated area maps at a spatial resolution of 250 m for the period of 2000–2015 using data from the Moderate Resolution Imaging Spectroradiometer (MODIS) and 56 m high-resolution land use land cover (LULC) information in India. Gumma [
18] mapped irrigated agricultural areas for Ghana using remote-sensing methods and protocols with a fusion of 30 m and 250 m spatial resolution remote-sensing data. Xie [
19] mapped the extent of irrigated croplands across the conterminous U.S. (CONUS) for each year in the period of 1997–2017 at 30 m resolution, using the generated samples along with remote sensing features and environmental variables to train county-stratified random forest classifiers annually.
Most irrigated area mapping methods above-mentioned were based on time series of NDVI at a relatively low resolution of 250–1000 m. Disaggregating statistics data on the grid is another way to generate the irrigation maps. For example, the European irrigation map (EIM) [
20] was created by disaggregating regional-level statistics on irrigated cropland areas into a 100 × 100 m grid, using a land cover map and constrained by the Global Map of Irrigated Areas (GMIAs) [
10]. The remote sensing-based classification approach is also a great way to produce the irrigated crop maps. Salmon [
21] used supervised classification of remote sensing, climate, and agricultural inventory data to generate a global map of irrigated, rain-fed, and paddy croplands. Lu [
22] tried to use pixel-based random forest to map irrigated areas based on two scenes of GF-1 satellite images at 16 m in an irrigated district of China, during the winter-spring irrigation period of 2018. Magidi [
23] developed a cultivated areas dataset with the Google Earth Engine (GEE) and further used the NDVI to distinguish between irrigated and rainfed areas. A large variety of classification methods at different scales and showing various levels of accuracy can be found in the literature [
24,
25,
26,
27,
28,
29,
30,
31]. Many applications and the tool of cloud-based and open source in classification have been developed recently [
32,
33]. However, the cloudy contamination and revisit time of optical satellite creates a major limitation to accurately identifying irrigation signature on the imagery. SAR imagery is less impacted by the cloud and has the advantage of building a long time series data to detect the irrigation signature. A number of studies [
34,
35,
36] used timer series of SAR images to detect the irrigation event. The fusion of optical and SAR time series images for classification is also progressing well in recent years [
37,
38,
39] in order to reduce the cloudy issue on the optical image. One study [
40] assessed the value of satellite soil moisture for estimating irrigation timing and water amounts.
All these above-mentioned studies were designed to identify the irrigation signature mainly in growing season as crop develops. However, irrigation also happens out of season due to various reasons, such as sufficient water supply out of season, cheaper water prices, and lower energy prices as well as manpower availability. This kind of irrigation practice should be given more attention as the winter irrigation is dominating in this region. Therefore, this study aimed to develop a method to identify the irrigated fields and help irrigation authorities know the irrigation situation before the growing season arrives to improve their water supply capacity in the whole year so that the crop production may be stably maintained. In this study area, it found a great number of fields already irrigated in winter and in early spring, although fields are bare soil and large volumes of irrigation water was applied to the fields. This kind of irrigation practice aims to keep enough soil moisture for sowing crops at the beginning of growing season in spring to avoid irrigation water competition and in preparation for coping with spring drought. This case also complements the consideration from those researchers who are developing irrigation maps within growing season for a large area or at a global level.
5. Discussion
5.1. The Challenges of Identifying Irrigation Outside the Growing Season
Our purpose was to know how many and in which fields irrigation has applied before the sowing season in May in spring. Many irrigated fields were able to be retrieved in the classification as the training samples were able to be visually identified. This case study has achieved its original research purpose and may complement the existing methods of mapping irrigation fields in growing season.
However, sometimes, it is not able to make the training samples inclusive. Shallow surface water or soil water in a few irrigated fields evaporates over time and the water in the fields gradually disappears as the air temperature goes up in spring. To distinguish this kind of dry up of irrigated field from other classes becomes indistinct due to the long interval between two satellite images. These kinds of irrigated fields will be omitted in the classified results as there are no training samples represented in this scenario.
This study was able to identify the irrigated fields but it did not answer which day irrigation was applied and how much water was applied. Both questions were not able to be answered in this study and they should be taken into consideration in the future research.
5.2. The Consideration in This Irrigation Mapping
In this study, only 7 scenes of GF-1 images out of growing season were valid and it witnessed the real capacity of GF-1 alone for identifying the irrigation fields. Optical satellite image is prone to cloud contamination. The better the results will be, the more multiple sources satellite images are available. Ideally, if daily and high-quality satellite images are available, it can identify the new irrigation event in time. In this sense, the integration of many more other high resolution satellite data, such as Sentinel-1/2 and Landsat8/9, should improve this study considerably.
In this study area, farmers conduct irrigation to the bare arable land as soon as the winter comes. It is easier to visually identify the irrigated field from bare land than vegetated fields. Two sets of irrigation scenarios in the fields were distinguished. Irrigation 1 represents the fields waterlogged or frozen in winter after the large volume flooding irrigation. Irrigation 2 represents the fields with the high soil moisture but without surface water. Due to the cold temperature and less evaporation in winter, no classification samples for Irrigation 2 were identified on the images of 17 December, 4 January, and 25 January while all irrigation samples represented Irrigation 1 as the irrigated fields were frozen on these dates. In the other four dates, the two kinds of irrigation conditions in the field were able to be identified.
5.3. The Winter Irrigation Impact on Ecosystem
The winter irrigation was a kind of cultivation management in the region in order to increase crop yield in the next year. Irrigation out of growing season has the advantage of protecting the ecosystem. It may help reduce the wind erosion due to wet soil in the field surface when the strong wind happens in spring. But a large volume of water applied also brings some adverse ecological effects on the farming system. Sowing in Spring 2023 had to be postponed due to wet soil in the field. On the image of 29 April 2023, it still found surface water on the fields. These fields were not able to be sowed in time. Therefore, the answer to the economic and minimum amount of water put into the field also needs to be further investigated. Soil salinization is another adverse effect induced by irrigation. Large volume of water speeds up evaporation in spring and brings the salt in deep soil back to the field surface. These effects on the ecological system imposed by irrigation out of season are worth further investigating in the near future.
6. Conclusions
This study explored the remote sensing-based classification approach to identify irrigated fields out of growing season in the winter season of 2022 to 2023. The proposed classification approach took four spectral bands and all NDVI like indices computed from any two of these four bands of GF-1 satellite data as the input features of the Random Forest algorithm. Regarding the two key parameters of RF, the number of features was set as the square root of the number of input bands of the image while the number of the tree was set to 100. The classification samples corresponding to each image were obtained by visual interpretation with the support of collected field data and then separated into training and validation sets by a ratio of 70% to 30%. Finally, the irrigated fields along with time in Jinzhong basin of Shanxi province, China were retrieved on the seven scenes of valid GF-1 satellite images, respectively.
The results show that the method developed in this study performed well and no overperformance and underperformance were found as the accuracies of classified image were not higher or far lower than that from models. The validations showed that the mean of the highest out-of-bag accuracies for seven RF models was 94.9% and the mean of the averaged out-of-bag accuracies in the plateau for seven RF models was 94.1%; the overall accuracy for all seven classified outputs was in the range of 86.8–92.5%, Kappa in the range of 84.0–91.0%, and F1-Score in the range of 82.1–90.1%. The lowest OA was 86.8% in comparison with the model accuracy of 92.9%, and the highest OA 92.5% in comparison with the model accuracy of 94.4%. The F1-Scores for irrigation 1 on 17 December, 4 January, and 25 January were very high and in the range of 92.2–97.2%. On the other four dates, the F1-Scores for Irrigation 1 decreased slightly and in the range of 86.0–91.7%, and the F1-Scores for Irrigation 2 were in a large range of 72.7 to 95.8%.
It also found that irrigation in the study area was carried out in early November but the quite few fields started to be irrigated, and the number of irrigated fields increased and suspended in December and January when the irrigated fields were covered by frozen ice and it was not able to apply irrigation due to low temperature. The irrigation was carried out again as the temperature went up in late February. The irrigation extended dramatically in March and April. The largest irrigation area was identified in later April as the sowing happened in May and the fields must dry up for sowing. The area of irrigated fields in the study area were increasing over time with sizes of 98.6, 166.9, 208.0, 292.8, 538.0, 623.1, 653.8 km2 from December to April, accounting for 6.1%, 10.4%, 12.9%,18.2%, 33.4%, 38.7%, and 40.6% of the total irrigatable land in the study area, respectively.
This case study shows that there is another window out of growing season to map the irrigated fields using Random Forest classification algorithm. This knowledge may complement the traditional consideration of retrieving irrigation maps only in growing season with remote sensing images for a large area. It also found too much water was applied in this study area and a few wet fields were not able to be sowed in time. The positive and adverse effect on the ecologic system imposed by irrigation out of season is worth being further investigated in the near future in order to support sustainable water resources management in the region. If the dense and even-distributed time series of valid satellite images may be made available, the irrigated fields over time may be well identified with the proposed approach. It frequently provides better irrigation information to the water resource authority and then the water resource authority may evaluate the excess water usage and its ecological consequences.