Rice Crop Yield Prediction from Sentinel-2 Imagery Using Phenological Metric ^†

Abstract

Crop yield prediction at plot scale is a vitally important magnitude for farmers at the socio-economic level. This study aims to quantify rice yield using phenological metrics from a normalized difference vegetation index (NDVI) time series derived from Sentinel-2 imagery, with yield data collected from 32 plots with an area of 36 ha in the Ferreñafe District of the Lambayeque region, Peru. Three different rice yield models were obtained, the best linear regression models were obtained for the SVM classification, with R² of 0.69, MAE = 1.01 and RMSE = 1.23 t ha⁻¹; and MRL with R² of 0.61, MAE = 1.10 and RMSE = 1.38 t ha⁻¹; RF with R² of 0.44, MAE = 1.23 and RMSE = 1.66 t ha⁻¹. The models obtained open the possibility to generate more robust models using a larger number of samples, which would be useful for farmers as well as for management and planning decisions for food and economic security.

1. Introduction

One of the world’s major cereal crops is rice [1], which meets the nutritional demand and food security of 50% of the world’s population [2]. The importance of yield knowledge has been a challenge for future food production due to climatic factors and water scarcity. Remote sensing tools allow yield prediction using satellite imagery (MODIS, LandSat, Sentinel) [3]. However, being Sentinel-2 optical imagery, with higher spatial and temporal resolution it allows to decrease cloud-induced noise. Therefore, it allows the use of different methodologies, including Machine Learning for yield prediction [4].

Northern Peru’s contribution to national rice production is distributed in three regions: San Martin with 24%, Piura with 15% and Lambayeque with 13% of the national production. In the Lambayeque region, with a per capita consumption of 63.5 kg, approximately 417,597 ha are planted, representing 40% of the total sown area [5,6], playing a fundamental role from an economic and food security point of view for its inhabitants.

Therefore, and in line with similar studies [3,4,7], this study aims to quantify rice yield using phenological metrics from a Normalized Difference Vegetation Index (NDVI) time series derived from Sentinel-2 images.

2. Materials and Methods

2.1. Study Area

The study area of this work is the district of Ferreñafe, located in the region of Lambayeque, Peru (79°47′09.73″ W; 6°35′36.68″ S; 46 m above sea level); with a mean annual rainfall of 22 mm and mean annual temperatures with a minimum average of 15.4 °C and a maximum average of 28.8 °C, with clay soils. The tests were carried out on a total area of 36 ha in 5 different zones of approximately 5–12 ha each, with a total of 32 plots of rice (Tinajones variety) (Figure 1). Yield samples were taken from each study plot at the end of the growing season. See

Table 1. Information rice fields.

A-1		A-2		A-3		A-4		A-5
Area (Ha)	Yield (t ha−1)	Area (Ha)	Yield (t ha−1)	Area (Ha)	Yield (t ha−1)	Area (Ha)	Yield (t ha−1)	Area (Ha)	Yield (t ha−1)
1.98	10.41	1.11	10.45	0.64	11.63	0.77	11.91	1.99	8.81
1.74	9.66	0.96	10.51	1.14	8.11	0.99	13.26	1.42	7.36
1.58	9.72	0.76	10.90	1.11	10.00	1.38	14.31	1.11	7.87
		0.75	10.75	1.03	12.04	1.49	12.29	0.80	6.28
		1.16	9.91	0.85	10.21	0.95	13.88	1.58	4.98
		1.13	11.06					0.66	5.06
		0.90	10.18					0.99	7.57
		0.78	11.53
		1.13	10.71
		1.03	10.81
		1.02	10.61
		1.15	8.69

.

2.2. Data Processing

The Google Earth Engine (GEE) platform was used to extract the NDVI time series from ten Sentinel-2 images, with a spatial resolution of 10 m, considering a filtering of <30% cloud cover. In addition, the mean and maximum NDVI value at a 15-day interval was extracted for the entire rice growing season (December 2021 to June 2022). Subsequently, the extracted values of the NDVI time series were input to generate the phenological metrics in the R programming language with the “CropPhenology” library [8]. In addition, “Weka” software [9] was used for variable selection, model generation and cross-validation, as shown in Figure 2.

2.3. Phenological Metrics

The phenological metrics were extracted considering studies conducted by Araya et al. [8]. They are defined as shown in Figure 3.

2.4. Statistical Analysis

Data analysis was performed by applying Support Vector Machine (SVM), Linear Regression (LR) and Random Forest (RF) for the selection of the metrics, and Multiple Linear Regression (MRL) to obtain rice yield models, which were evaluated by Leave-One-Out Cross-Validation (LOOCV) [10], obtaining the coefficients of determination (R²), the root mean square error (RMSE) and the Mean Absolute Error (MAE).

3. Results and Discussion

Figure 4 shows the r-Pearson correlation coefficients between the phenological metric variables extracted from the NDVI time series of the rice crop. Presenting collinearity between the variables “GreenUpSolpe” and “MaxV”, and in the other hand, between “TINDVI” and “TINDVIAfterMax”, withan r-Pearson correlation ≥ 0.90.

Normally, rice yield models are based on NDVI analysis during different phenological stages of the crop, which are reflected in the phenological metrics according to [8]. This allows to generate models by automatic learning, thus obtaining reasonable errors even with few images available due to weather conditions. Therefore, a higher frequency of images could potentially allow a better accuracy in determining the key date for model creation.

The machine learning models selected the following metrics: LR (OffsetT and TINDVIBeforeMax), SVM (OnsetV, MaxT, OffsetT and BrownDownSolpe) and RF (MaxT, TINDVIAfterMax and TINDVI).

Table 2. Statistical summary of yield prediction models evaluated using cross-validation (leave-one-out), coefficient of determination, root mean square error and mean absolute error.

Method of Selection		Models	R2	RMSE (t ha−1)	MAE	N-Value
LR	Multiple Linear Regression	2.3369 × OffsetT + 8.6715 × TINDVIBeforeMax − 21.0089	0.61	1.38	1.10	32
SVM		36.4304 × OnsetV + 1.4759 × MaxT + 3.0196 × OffsetT + 74.65 × BrownDownSlope − 34.6315	0.69	1.23	1.01	32
RF		2.8828 × MaxT − 4.9856 × TINDVIAfeterMax + 7.2392 × TINDVI − 15.0098	0.44	1.66	1.23	32

shows the summary statistics of the models generated by the multiple linear regression of the variables selected for each automatic selection method. The best result was obtained using SVM for metrics selection, with R² of 0.69, MAE = 1.01 and RMSE = 1.23 t ha⁻¹; and LR with R² of 0.61, MAE = 1.10 and RMSE = 1.38 t ha⁻¹; RF with R² of 0.44, MAE = 1.23 and RMSE = 1.66 t ha⁻¹.

4. Conclusions

The present study evaluated three rice yield models, based on the analysis of sentinel-2 NDVI time series, with the extraction of 15 phenological metrics. Three machine learning methods (LR, SVM and RF) were used for variable selection, then multiple linear regressions were generated for each machine learning model. The best models were generated using the following phenological metrics of OnsetV, MaxT, OffsetT, brownDownSlopw, TINDVIBeforeMax, TINDVIAfterMax and TINDVI. It is expected to obtain data from future seasons at the same points in order to develop a more robust model applicable to plots without yield information.