1. Introduction
The global cropland area is predicted to decline by 1.8–2.4% by 2030 due to conversion of arable croplands to mostly built-up landcover, and 80% of this land cover change is expected to occur in Asia and Africa [
1]. Bengaluru is one of the megacities (over 10 million population) in southern India [
2], which has already lost 62% of the vegetated area, while the urban area increased by 125% between 2001 and 2011 [
3]. Agricultural production has intensified (i.e., high nitrogen (N) fertiliser usage, drip irrigation), and the cropping pattern has changed to meet the increasing food demand for the growing population. Between 2006 and 2012, the cropping pattern in Bengaluru changed from high water use paddy cultivation to dry land cereals and pulses (e.g., maize, finger millet, lablab). According to the state-level statistics, maize and finger millet crop yield increased by 4 to 6% annually, while pulse yield (including lablab) soared by 15% [
4].
Increasing crop production using available arable lands while sustainably managing resources (e.g., water, soil) and reducing climate change is challenging [
5]. Thus, near-real-time crop status monitoring could be a way forward to manage available resources and reduce inputs (i.e., precision agriculture). However, crop monitoring approaches need to be adapted to distinct crop types, in different growth stages (phenology), and under different agricultural practices. Remote sensing (RS) is one of the primary tools for crop monitoring [
6]. RS facilitates contactless data collection over a given crop area using reflected electromagnetic energy, enabling the characterisation of an area′s spatiotemporal information. The development of RS data collection and analysis techniques helps to achieve accurate models to estimate crop parameters.
Various sensor platforms (i.e., terrestrial, airborne, and spaceborne) have been employed to collect data about cropping areas and estimate crop growth and health parameters through different modelling approaches [
7]. Generally, the reflected electromagnetic energy from the plant changes according to the physiological and the structural condition of crops and the surrounding environment [
8]. Both multi- and hyperspectral sensors have been utilised from different platforms to capture these varying reflected energies. Hyperspectral sensors capture reflected energy at many narrow spectral bands (usually more than 30 bands). In comparison, multispectral sensor data contains fewer spectral bands with larger bandwidth [
9]. Due to the higher spectral sensitivity of the hyperspectral data, there is a significant potential to capture a wider variety of different physiological and structural crop traits [
8]. To make the clear comparison of the spectral resolution difference of the RS data for crop trait estimation, it is necessary to obtain RS data with similar spatial resolution. However, most studies which compared the spectral resolution sensitivity (hyperspectral vs. multispectral) for crop trait estimation were based on different spatial resolution; for example [
10] employed field spectroscopy data as hyperspectral data with point observation and satellite data as multispectral data with 10 m spatial resolution for estimation of maize crop traits.
Empirical (statistical) models (both parametric and non-parametric) or physical models (e.g., radiative transfer model inversion) have been employed to estimate crop parameters using spectral data [
11]. The empirical models inspect the association between in-situ measured target crop vegetation parameter and spectral reflectance data collected from RS. The reflectance data or their transformations (e.g., first derivative) or vegetation index (VI) developed from many wavebands were the inputs for the empirical models. A linear regression model is one of the standard parametric empirical modelling methods which estimates crop traits by utilising single waveband reflectance data or VI data as input [
12]. In contrast, all—or only the essential—waveband reflectance data (original and transformed) and a multitude of VI data can be used as inputs for non-parametric empirical modelling with, e.g., machine learning methods (i.e., random forest, Gaussian process) [
13]. Since both parametric and non-parametric models are data driven methods, a comparison of these methods for estimation of crop traits using RS spectral data can always provide capabilities of different modelling methods [
14].
Many crop vegetation parameters that indicate growth and health status have been estimated using RS spectral data, e.g., leaf area index (LAI), leaf chlorophyll content (LCC), and canopy water content (CWC) [
7]. LAI (m
2/m
2) is the leaf area per unit ground area, an essential plant biophysical variable to understand growth, health, and yield [
15]. When considering other photosynthetically-active plant parts besides the leaves, it is called the green area index or plant area index [
8]. Crop LAI estimation using RS reflectance data and empirical modelling approaches (both parametric and non-parametric) have shown promising results, but also considerable variation in prediction quality (coefficient of determination (R
2) ranges from 0.36 to 0.97) [
16].
The LCC (both chlorophyll a and b) is a crop biochemical indicator for photosynthetic capacity, environmental stress, and N status of leaves [
17,
18]. LCC (µg/cm
2) is referred to as leaf-level quantification, while the multiplication of LCC with LAI is considered canopy chlorophyll content (CCC-g/cm
2). Spectral reflectance from the green to near-infrared region shows a strong relationship with LCC values [
8]. According to available literature, LCC can be estimated with a maximum relative error of less than 20 % from both multi- and hyperspectral sensors [
19,
20].
Quantification of CWC (g/m
2) attempts to identify crop water stress by estimating the quantity of water per unit area of the ground surface [
21]. Water absorption regions (970 nm and 1200 nm) of the spectral reflectance data have been employed to estimate CWC using RS spectral data [
21,
22,
23]. However, few studies were able to accurately estimate (R
2 > 0.7) maize crop CWC using linear regression models with VI derived from wavebands from the green, red-edge and near-infrared regions [
24,
25]. Conversely, the crop CWC has not yet been estimated using full spectral data to uncover the full potential of hyperspectral information.
Successful estimation of crop vegetation parameters with RS spectral data has been demonstrated for various crop types such as wheat, rice, barley, and maize [
7,
26,
27]. However, RS data application has not been examined for crops like finger millet and lablab, which are major monsoon crops in the tropical region (e.g., Bengaluru, Southern India). Furthermore, few studies have compared different remote sensing platforms (e.g., in-situ vs airborne vs spaceborne) and sensors (multispectral vs hyperspectral) for crop vegetation parameters estimation [
16,
28]. Thus, this study sought to fill the identified research and knowledge gap for RS for monsoon crop monitoring. The primary objective of this study is to evaluate two different RS spectral data types (420–970 nm) with a similar spatial resolution (~1 m), namely spaceborne multispectral (WorldView3–8 bands) and unmanned aerial vehicle (UAV) borne hyperspectral (Cubert–126 bands) for estimating three crop vegetation parameters (LAI, LCC, and CWC) from three crop types (finger millet, maize, and lablab) under different agricultural treatments (irrigation and fertiliser). The specific sub-objectives of this study were:
To build crop-specific parametric and non-parametric models to estimate crop vegetation parameters
To evaluate the developed vegetation parameter estimation models against (a) the spectral sensitivity of the RS data (multispectral vs hyperspectral), (b) modelling method (parametric and non-parametric), and (c) crop type (finger millet, maize, and lablab)
To explore how crop-wise vegetation parameter estimation is affected by agricultural treatment (irrigation and fertiliser)
4. Discussion
The main objective of this study was to evaluate two different spectral RS datasets (multispectral WV3 and hyperspectral CUB) for estimating three crop vegetation parameters (LAI, LCC, and CWC) of three major tropical crop types (finger millet, maize, and lablab). Considering the modelling method, out of the best nine (three vegetation parameters × three crop types) LR models based on VIs, CUB data provided six of the best models, while WV3 data provided three of the best models (
Table 4). In contrast, out of the best nine RFR models with selected WBs, five of the best models were based on CUB data, whereas the other four relied on WV3 data. Overall, these results did not show a definite pattern between the RS datasets and the vegetation parameter estimation model′s accuracy. Similarly, [
10] reported that maize LAI estimation accuracy did not significantly differ between data with two different spectral resolutions and two different modelling methods (LR vs machine learning regression). In contrast, [
54] detailed that narrow band VIs derived from hyperspectral data models yielded 20% higher R
2 values than multispectral data models for wheat and barley LAI estimation.
4.1. Finger Millet Vegetation Parameter Estimation
According to the authors′ knowledge, only a few studies have utilised RS data to estimate crop vegetation parameters of finger millet and lablab [
29,
55]. Finger millet is a small-grained cereal (C4 type) with similar crop characteristics as pearl millet, sorghum, and foxtail millet [
56]. This study revealed that the hyperspectral CUB data clearly showed the substantial potential to estimate finger millet vegetation parameters irrespective of the modelling method. For finger millet LAI estimation, NDVI
800_670 from CUB data showed the minimum error, which confirmed that NDVI has a closer relationship with LAI at lower LAI values (less than 3.2 m
2/m
2) [
8]. Similar to these results, NDVI showed the best estimation accuracy for sorghum LAI than other VIs (i.e., greenNDVI, EVI, and MTVI2) [
57].
DATT4 is a VI for leaf chlorophyll a and chlorophyll a+b content estimation [
36] and, when derived from CUB data, showed the strongest correlation with finger millet LCC (
Table 4). However, DATT4 from WV3 was the least correlated VI (
Figure A1). The central wavelengths of the WV3 bands do not match with the exact wavelengths of the DATT4′s formula, which may have reduced the sensitivity of the index. In contrast, Two NDVIs (NDVI
800_670 and NDVI
750_550) from CUB and WV3 data also showed a strong correlation with finger millet LCC (
Figure A1). However, sorghum′s LCC showed the highest correlation with hyperspectral data NDVI [
58] and indicated a lower correlation with multispectral data NDVI [
59].
Models with VIs showed better finger millet CWC estimation results for both RS datasets. NDVI
750_550 was the best correlated VI from both datasets, which predicts CWC indirectly [
25] and contained green and near-infrared bands. CWC estimation with VI derived from green and near-infrared bands (
) also showed the best results among other VIs that predict CWC indirectly (i.e., NDVI, NDVIrededge, and CIrededge) [
24]. When it comes to RFR modelling with selected WBs, WBs above 750 nm were not selected for finger millet CWC estimation. Nevertheless, some of the identified vital WBs were comparable with important WBs for finger millet fresh biomass estimation using multi-temporal terrestrial CUB data (e.g., 694 nm) [
29].
4.2. Lablab Vegetation Parameter Estimation
Lablab is a legume crop similar to pea, beans, and lentils [
60]. The lablab LAI values showed a strong correlation with NDVI values, but the LAI estimation error with NDVI was higher than the error from RFR models with selected WBs. The higher LAI values (>3.0) from lablab may impede accurately estimating LAI with NDVI due to the saturation effect, which also demonstrated by [
39] with pea LAI values. In comparison to lablab LAI estimation, LR models with VI showed improved results for lablab LCC estimation. NDVI
750_550, which contains the green band with the near-infrared band instead of the red band, was the most highly correlated VI with lablab LCC. NDVI
750_550 is also known as ‘Green NDVI’, and according to [
38], shows a strong relationship with Chlorophyll a.
NDVI and REIP, respectively, from CUB and WV3 data, delivered the lowest error for lablab CWC estimation. Even though these VI do not directly relate to the leaf water content, they could determine CWC because they are linked to crop biomass [
25]. Furthermore, the identified best WBs from CUB data for lablab CWC estimation (
Table 5 and
Figure A2) were similar to the critical WBs for lablab fresh biomass estimation [
29].
4.3. Maize Crop Vegetation Parameter Estimation
As opposed to finger millet and lablab, maize has been frequently explored with RS data for its vegetation parameter estimation. LR modelling with hyperspectral (CUB) data to calculate NDVI showed a lower error than NDVI from multispectral (WV3) data for maize LAI estimation. [
10] also revealed the same pattern for maize LAI estimation using VI from hyperspectral (field spectrometer) and multispectral (Sentinel-2) data. RFR models with essential WBs showed similar relative errors for maize LAI estimation using both RS datasets. Likewise, maize LAI estimation models from hyperspectral data and multispectral data also demonstrated similar cross-validation error (nRMSE
cv = 14.9 %) with a support vector machine algorithm [
10].
VI derived from green, red-edge, and near-infrared bands were usually better for LCC estimation [
61,
62] Logically, VI containing those bands (i.e., NDIV
800_670, DATT4) were strongly correlated with maize LCC values. However, RFR models with WV3 data had > 31% relative error, although the centre wavelength of the red band from WV3 data is 660.1 nm, which is the region absorbed by leaf chlorophyll a [
63]. In comparison, RFR models with CUB data obtained slightly lower error, but all the essential WBs were between 682–702 nm (red-edge region) (
Table 5 and
Figure A2). This contrasts with results from another study using the same hyperspectral sensor (CUB) data, which reported the usefulness of WBs from blue, red, red-edge, and near-infrared regions for maize LCC estimation [
64].
Indirectly linked VIs could estimate maize CWC in this study, while WI, which is a directly sensitive VI for CWC, showed the weakest relationship with CWC for all crops. This could be because crop parameters were highly correlated, and the variation of CWC somehow directly linked with the crop LAI and biomass values [
25]. Nevertheless, water absorption at 970 nm due to O-H bonds in liquid canopy water [
65] was one of the key WBs for maize CWC estimation by CUB data only (
Table 5 and
Figure A2).
4.4. Overall Discussion
This study could not conclude which RS data (spaceborne multispectral or UAV-borne hyperspectral) is better for the evaluated crop parameters for three crop types. Nevertheless, it is worth to mention the pros and cons of the two RS systems in terms of practical aspects of general crop monitoring. The spaceborne multispectral WV3 data hugely affected by cloud coverage in tropical regions, especially in the rainy season. Proper atmospheric corrections are needed to obtain accurate surface reflectance data from WV3 images to relate spectral values with crop vegetation parameters, which might not be easy to achieve. Additionally, the WV3 data cannot be acquired whenever it is needed because of its revisit frequency of one to five days, depending on the latitude. However, applying WV3 data to estimate crop parameter in the entire crop field can be efficiently performed because of the large spatial coverage of each satellite scene.
On the other hand, the UAV-borne CUB data can be collected whenever the data is needed, and there is no effect on the data due to cloud cover (when a proper radiometric correction is applied). However, coverage of a larger field needs to done using several UAV flight sessions, which could be a disadvantage over the WV3 data. Additionally, UAV-borne data is also challenging to collect in extreme weather conditions such as rain and wind, typical of the tropical region′s monsoon seasons.
This study’s third sub-objective explored how the crop parameter estimation accuracy was affected by the crop′s water and fertiliser treatments. The collected field data showed a significant positive effect due to irrigation in all three crops. However, finger millet (inflorescence emergence) and maize (development of fruit) were in similar phenological stages in both water treatments, while lablab showed two different phenological stages for irrigated and rainfed crops. (
Table 1). The results clearly showed that the prediction accuracy of crop vegetation parameters did not significantly affect irrigation, and only finger millet′s LAI and CWC prediction error had a significant difference due to fertiliser treatments (
Figure A3). Confirming these findings, [
29] also reported no significant difference for biomass prediction error between two water treatments and fertiliser treatments for the same three crops with three-year data using in-situ hyperspectral data with machine learning methods.
This study utilised only a few (n = 24) samples for model building for vegetation parameter estimation. For this reason, separate models for the irrigation treatments were not employed, even though the data showed a significant difference between treatments. Therefore, the CV was applied to build unbiased models, which facilitated evaluating models with a limited number of data points from both treatments. However, the number of sample points for both training (n = 22) and validation (n = 2) in the CV was not enough to capture the dataset′s total variability. For example, when the model was trained with a unique range of dataset and the validation data points were out of the range from the trained model, then the model tends to under or overestimates the prediction value. It is necessary to have more data points to increase the model sensitivity to the dataset′s total variability. However, having many sample points is always challenging for RS-based crop parameter estimation for many reasons, including human and physical resource availability.
The two RS datasets used in this study were sensitive from the visible to the near-infrared region. According to published studies, usage of the spectral region until the shortwave infrared (2500 nm) could increase crop parameter estimation potential [
12,
24]. The two RS datasets utilised in this study could accurately estimate three crop vegetation parameters from three crop types with different agriculture treatments. Hence these results could be utilised as a starting point to an in-depth examination of how to use RS data without shortwave infrared spectral data for modelling LAI, LCC, and specifically CWC. Additionally, these research findings could be employed to monitor monsoon crops using the currently available spaceborne and UAV-borne high spatial resolution remote sensors with similar spectral sensitivity (e.g., Parrot Sequoia, Micasense RedEdge, and microsatellite constellations such as Planet).
5. Conclusions
This study focused on uncovering how two different spectral resolution RS data can be utilised for estimating crop vegetation parameters from three crops (finger millet, maize, and lablab) prominently grown in Southern India. This study evaluated two different very high spatial resolution (>1.5 m) RS spectral datasets (UAV-borne hyperspectral Cubert–CUB, spaceborne multispectral WorldView3–WV3) for estimating LAI, LCC, and CWC for the three target crops. Two distinct modelling methods, namely linear regression with best-correlated vegetation index and random forest regression with important wavebands, were also evaluated. According to the results, irrespective of the RS datatype, crop type, and modelling method, the average relative estimation error was less than 16%, 25%, and 22%, respectively, for LAI, LCC, and CWC estimation. However, there was no clear evidence to identify the best RS dataset or the best modelling method to estimate the examined crop parameters. Nevertheless, there was a trend that hyperspectral (CUB) data was better for estimation of vegetation parameters of finger millet while multispectral (WV3) data was better for both lablab and maize vegetation parameter estimation. Overall, vegetation indices derived from the combination of either green, red, red-edge, and near-infrared wavebands showed clear potential from either multi or hyperspectral data for an accurate estimation of the investigated vegetation parameters regardless of the crop type.