Towards a Machine Learning-based Model for Automated Crop Type Mapping

Crop type identification and mapping have been a classic issue essential for agriculture characterisation and food security as well as crop health monitoring. Remote sensing was proven to be an efficient tool to define and study agriculture. The complementarity between optical and radar data can hugely improve crop type identification and mapping. This document presents an overview of more than 40 recent works between 2013 and 2019, related to crop type mapping using satellite imagery, and points the general techniques used. Different classification algorithms was studied and compared, as well as the comparison of the inputs, optimal bands and polarization for radar data, and process giving the best results. It is often concluded that fusion improves results when compared with single data using. Nevertheless, it had been challenging to combine datasets from sensors that had different physical concepts, and progress in this domain require the development of systematic procedures.

Modeling time series of vegetation indices in tallgrass prairie using machine and deep learning algorithms

LiDAR data are becoming increasingly available, which has opened up many new applications. One such application is crop type mapping. Accurate crop type maps are critical for monitoring water use, estimating harvests and in precision agriculture. The traditional approach to obtaining maps of cultivated fields is by manually digitizing the fields from satellite or aerial imagery and then assigning crop type labels to each field – often informed by data collected during ground and aerial surveys. However, manual digitizing and labeling is time-consuming, expensive and subject to human error. Automated remote sensing methods is a cost-effective alternative, with machine learning gaining popularity for classifying crop types. This study evaluated the use of LiDAR data, Sentinel-2 imagery, aerial imagery and machine learning for differentiating five crop types in an intensively cultivated area. Different combinations of the three datasets were evaluated along with ten machine learning. The classification results were interpreted by comparing overall accuracies, kappa, standard deviation and f-score. It was found that LiDAR data successfully differentiated between different crop types, with XGBoost providing the highest overall accuracy of 87.8%. Furthermore, the crop type maps produced using the LiDAR data were in general agreement with those obtained by using Sentinel-2 data, with LiDAR obtaining a mean overall accuracy of 84.3% and Sentinel-2 a mean overall accuracy of 83.6%. However, the combination of all three datasets proved to be the most effective at differentiating between the crop types, with RF providing the highest overall accuracy of 94.4%. These findings provide a foundation for selecting the appropriate combination of remotely sensed data sources and machine learning algorithms for operational crop type mapping.

MODerate-resolution Imaging Spectroradiometer (MODIS) vegetation index (VI) time-series data are among the most widely utilized remote sensing datasets. To improve the quality of MODIS VI time-series data, most prior methods have focused on correcting negatively biased VI noise by approaching the upper envelope of the VI time series. Such treatment, however, may cause overcorrections on some true local low VI values, resulting in inaccurate simulations of vegetation phenological characteristics. In addition, another challenge in reconstructing MODIS VI time series is to fill temporally continuous gaps. The earlier spatiotemporal Savitzky-Golay (STSG) method tackled this problem by utilizing multiyear VI data, but its performance heavily relies on the consistency of data across different years. In this study, we proposed an improved STSG (iSTSG) method. The new method accounts for the autocorrelation within the VI time series and fills missing values in the VI time series by leveraging spatiotemporal VI data from the current year alone. Furthermore, iSTSG incorporates an indicator to quantify potential overcorrections in the VI time series, aiming to more accurately simulate phenological characteristics. The experiments to reconstruct MODIS normalized difference VI (NDVI) time-series product (MOD13A2) at four typical sites (a million square kilometers for each site) suggest two clear advantages in iSTSG over the iterative SG (called Chen-SG) and STSG methods. First, iSTSG more accurately reconstructs the annual NDVI time series, exhibiting the smallest mean absolute differences (MADs) between the smoothed and the simulated reference NDVI time series (0.012, 0.018, and 0.020 for iSTSG, STSG, and Chen-SG, respectively). Second, iSTSG more effectively simulates phenological characteristics in the NDVI time series, including the onset dates for vegetation greenup and dormancy, as well as the crop harvest period. The advantages of iSTSG were also demonstrated when applied to the successor of MODIS, Visible Infrared Imaging Radiometer Suite (VIIRS) VI time-series product (VNP13A1). iSTSG can be implemented on the Google Earth Engine (GEE), offering significant benefits for various applications, particularly in crop mapping and vegetation/crop phenology studies.

ABSTRACTCotton is the most important fibre culture in the world. In Brazil, cotton cultivation is concentrated in the Cerrado biome, the Brazilian savanna, and is one of the most important commodities in the country. As an annual crop, the updating frequency of the spatial distribution data of cotton fields is extremely important for crop monitoring systems. In order to provide fast and accurate information for crop monitoring, time series of remote- sensing data has been used in the development of several applications in agriculture, since the high temporal resolution of some orbital sensor allows monitoring targets with high spectral-temporal variations in the land surface. However, there are still some challenges to systematize the processing of such a large amount of data available by long time series of remote-sensing imagery. Thus, this study contributes to the construction of models to identify and separate specific crop types with similar spectral behaviour to other crops practised in the same period. The objective of this study was to develop a systematic methodology based on data mining of time series of vegetation indices (VI) to map cotton fields at the regional scale. Field reference data and time series of NDVI and EVI images, obtained from MODIS sensor products during four cropping seasons (from 2012–2013 to 2015–2016), were used to construct mapping models based on decision tree algorithms. Phenological metrics were calculated from the VI time series and used to build classification rules for mapping cotton fields. Our results demonstrate that the proposed method to map cotton fields achieve high accuracy when field data and visual interpretation of NDVI temporal profiles were used for validation (accuracy higher than 95% and 93%, respectively). Comparisons with the official statistics indicated an optimal fit, with linear correlation (r) and coefficient of determination (R2) above 0.93. Therefore, the proposed method was efficient to distinguish cotton fields from other crop types with similar spectral behaviour. In addition, this method can also be applied to other cotton-producing regions and other production seasons, by reusing the models generated through machine learning approaches.

In this paper, satellite images, machine learning algorithms and field measurements are combined to analyze the possibility of generating a model that predicts the yield of the legume Vicia villosa Roth before harvest. In a previous study, satellite information from different dates throughout the complete phenological cycle of the legume cultivated in several plots of the Guaminí district (Buenos Aires) was used and a close relationship was found between yield level and the time series of vegetation indices. Based on these results, the possibility of predicting yield is evaluated considering nine dates of the 2021-2022 campaign between sowing and prior to harvest. They are associated in different ways, determining their impact on the accuracy of the trained model.Results show that legume yield can be adequately classifed with remote monitoring of five dates. To have a prediction model could help in in-situ decisions by optimizing the use given to the crop (direct grazing, forage or seed production) based on its expected profit.

Timely and accurate crop type distribution maps are an important inputs for crop yield estimation and production forecasting as multi-temporal images can observe phenological differences among crops. Therefore, time series remote sensing data are essential for crop type mapping, and image composition has commonly been used to improve the quality of the image time series. However, the optimal composition period is unclear as long composition periods (such as compositions lasting half a year) are less informative and short composition periods lead to information redundancy and missing pixels. In this study, we initially acquired daily 30 m Normalized Difference Vegetation Index (NDVI) time series by fusing MODIS, Landsat, Gaofen and Huanjing (HJ) NDVI, and then composited the NDVI time series using four strategies (daily, 8-day, 16-day, and 32-day). We used Random Forest to identify crop types and evaluated the classification performances of the NDVI time series generated from four composition strategies in two studies regions from Xinjiang, China. Results indicated that crop classification performance improved as crop separabilities and classification accuracies increased, and classification uncertainties dropped in the green-up stage of the crops. When using daily NDVI time series, overall accuracies saturated at 113-day and 116-day in Bole and Luntai, and the saturated overall accuracies (OAs) were 86.13% and 91.89%, respectively. Cotton could be identified 40∼60 days and 35∼45 days earlier than the harvest in Bole and Luntai when using daily, 8-day and 16-day composition NDVI time series since both producer’s accuracies (PAs) and user’s accuracies (UAs) were higher than 85%. Among the four compositions, the daily NDVI time series generated the highest classification accuracies. Although the 8-day, 16-day and 32-day compositions had similar saturated overall accuracies (around 85% in Bole and 83% in Luntai), the 8-day and 16-day compositions achieved these accuracies around 155-day in Bole and 133-day in Luntai, which were earlier than the 32-day composition (170-day in both Bole and Luntai). Therefore, when the daily NDVI time series cannot be acquired, the 16-day composition is recommended in this study.

The aim of this study is to predict and map winter wheat yield in the Parvomay municipality, situated in the Upper Thracian Lowland of Bulgaria, utilizing satellite data from Sentinel-2. The main crops grown in the research area are winter wheat, rapeseed, sunflower, and maize. To distinguish winter wheat fields accurately, we evaluated classification methods such as Support Vector Machines (SVM) and Random Forest (RF). These methods were applied to satellite multispectral data acquired by the Sentinel-2 satellites during the growing season of 2020–2021. In accordance with their development cycles, temporal image composites were developed to identify suitable moments when each crop is most accurately distinguished from others. Ground truth data obtained from the integrated administration and control system (IACS) were used for training the classifiers and assessing the accuracy of the final maps. Winter wheat fields were masked using the crop mask created from the best-performing classification algorithm. Yields were predicted with regression models calibrated with in situ data collected in the Parvomay study area. Both SVM and RF algorithms performed well in classifying winter wheat fields, with SVM slightly outperforming RF. The produced crop maps enable the application of crop-specific yield models on a regional scale. The best predictor of yield was the green NDVI index (GNDVI) from the April monthly composite image.

Mapping and monitoring the distribution of croplands and crop types support policymakers and international organizations by reducing the risks to food security, notably from climate change and, for that purpose, remote sensing is routinely used. However, identifying specific crop types, cropland, and cropping patterns using space-based observations is challenging because different crop types and cropping patterns have similarity spectral signatures. This study applied a methodology to identify cropland and specific crop types, including tobacco, wheat, barley, and gram, as well as the following cropping patterns: wheat-tobacco, wheat-gram, wheat-barley, and wheat-maize, which are common in Gujranwala District, Pakistan, the study region. The methodology consists of combining optical remote sensing images from Sentinel-2 and Landsat-8 with Machine Learning (ML) methods, namely a Decision Tree Classifier (DTC) and a Random Forest (RF) algorithm. The best time-periods for differentiating cropland from other land cover types were identified, and then Sentinel-2 and Landsat 8 NDVI-based time-series were linked to phenological parameters to determine the different crop types and cropping patterns over the study region using their temporal indices and ML algorithms. The methodology was subsequently evaluated using Landsat images, crop statistical data for 2020 and 2021, and field data on cropping patterns. The results highlight the high level of accuracy of the methodological approach presented using Sentinel-2 and Landsat-8 images, together with ML techniques, for mapping not only the distribution of cropland, but also crop types and cropping patterns when validated at the county level. These results reveal that this methodology has benefits for monitoring and evaluating food security in Pakistan, adding to the evidence base of other studies on the use of remote sensing to identify crop types and cropping patterns in other countries.

Cropland mapping using remote sensing data is the basis for effective crop monitoring, crop rotation control, and the detection of irrational land use. Classification using Normalized Difference Vegetation Index (NDVI) time series from multi-year data requires additional time costs, especially when sentinel data are sparse. Approximation by nonlinear functions was proposed to solve this problem. Time series of weekly NDVI composites were plotted using multispectral Sentinel-2 (Level-2A) images at a resolution of 10 m for sites in Khabarovsk District from April to October in the years 2021 and 2022. Missing values due to the lack of suitable images for analysis were recovered using cubic polynomial, Fourier series, and double sinusoidal function approximation. The classes that were considered included crops, namely, soybean, buckwheat, oat, and perennial grasses, and fallow. The mean absolute percentage error (MAPE) of each class fitting was calculated. It was found that Fourier series fitting showed the highest accuracy, with a mean error of 8.2%. Different classifiers, such as the support vector machine (SVM), random forest (RF), and gradient boosting (GB), were comparatively evaluated. The overall accuracy (OA) for the site pixels during the cross-validation (Fourier series restored) was 67.3%, 87.2%, and 85.9% for the SVM, RF, and GB classifiers, respectively. Thus, it was established that the best result in terms of combined accuracy, performance, and limitations in cropland mapping was achieved by composite construction using Fourier series and machine learning using GB. Similar results should be expected in regions with similar cropland structures and crop phenological cycles, including other regions of the Far East.

With the latest development and increasing availability of high spatial resolution sensors, earth observation technology offers a viable solution for crop identification and management. There is a strong need to produce accurate, reliable and up to date crop type maps for sustainable agriculture monitoring and management. In this study, RapidEye, the first high-resolution multi-spectral satellite system that operationally provides a Red-edge channel, was used to test the potential of the data for crop type mapping. This study was investigated at a selected region mostly covered with agricultural fields locates in the low lands of Menemen (İzmir) Plain, TURKEY. The potential of the three classification algorithms such as Maximum Likelihood Classification, Support Vector Machine and Object Based Image Analysis is tested. Accuracy assessment of land cover maps has been performed through error matrix and kappa indexes. The results highlighted that all selected classifiers as highly useful (over 90%) in mapping of crop types in the study region however the object-based approach slightly outperforming the Support Vector Machine classification by both overall accuracy and Kappa statistics. The success of selected methods also underlines the potential of RapidEye data for mapping crop types in Aegean region.

Climate change is of great significance to vegetation coverage. However, the long time series of climatic factors and vegetation index is non-stationary and non-linear, containing different information in frequency and time scales. The study innovatively integrated Support Vector Machine (SVM) and Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), and analyzed the relationship between climatic factors and the Normalized Difference Vegetation Index (NDVI) at three-time scales in Loess Plateau (LP). The results indicate that: (1) The mode mixing phenomenon of EMD can be better solved by using ICEEMDAN, and the end effect can be mitigated by using SVM to expand both ends of the data before decomposition. (2) The residual NDVI stripped out using ICEEMDAN-SVM can represent the long-term trend of the original data. The results show that the NDVI and climate factors performed noticeable spatial and temporal differences under different climate zones. (3) The relationship between vegetation and climatic factors revealed obvious spatial heterogeneity at different time scales, and the climate change in the overall trend explained the vegetation change to the greatest extent, about 95%. The study has shown that temperature is a limiting factor affecting the growth of vegetation on the Loess Plateau. We propose ICEEMDAN-SVM to study the relationship between climate factors and vegetation indices at multiple timescales., revealing some hidden information in long time series and providing a new method to quantify the impact of climate change on vegetation dynamics.

Recently, there has been a remarkable growth in Artificial Intelligence (AI) with the development of efficient AI models and high-power computational resources for processing complex datasets. There has been a growing number of applications of machine learning in satellite remote sensing image data processing. In this work, machine learning methods were applied for crop classification of temporal multi- spectral satellite image to achieve better prediction of crop-wise area statistics. In India, agriculture has a huge impact on the national economy and most of the critical decisions are dependent on agricultural statistics. Sentinel-2 satellite image data for the Guntur district region of Andhra Pradesh, India has been used as the study area. The main reason for selecting this region is the diversity of agricultural crops and the availability of ground truth. As a baseline, the performance of machine learning models like Support Vector Machine (SVM), Random Forest (RF), Decision Tree (DT), and K-Nearest Neighbor (KNN) on crop classification was evaluated on 2016 single time multi-spectral satellite image. SVM performed well with an F1 score of 0.94. Further, the performance of SVM, RF, 1D and 2D Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) and RNN with Gated Recurrent Unit (GRU) on 2018 Kharif season temporal multi-spectral satellite image has been evaluated for estimation of crop areas. The results show that SVM has the best F1 score of 0.99 and achieved a 95.9% agreement with the ground surveyed crop areas.

Crop type mapping is relevant to a wide range of food security applications. Supervised classification methods commonly generate these data from satellite image time-series. Yet, their successful implementation is hindered by the lack of training samples. Solutions like transfer learning, development of temporal-spectral signatures of the target classes, re-utilization of existing inventories, or crowdsourcing initiatives are commonly applied to generate samples for thematically coarser classifications. These methods are rarely used for generating crop types samples. In this study, we leverage the phenology information of existing data inventories using Time-Weighted Dynamic Time Warping (TWDTW) to address the problem of automatic crop sample generation in two target areas. Resulting labeled samples are refined using proximity measures obtained from Random Forests (RF). Sentinel-2 time-series are used to obtain phenology information from two study areas. The proposed methodology achieved promising results for classes with a reduced inter-classes similarity such as sugar beets (user’s accuracy, UA, of 98% and producer’s accuracy, PA, of 100%) or grains (UA of 98% and PA of 90%). The crops with a high inter-classes similarity yielded less satisfactory results. Potatoes, for example, obtained a high PA of 95%, but a UA of only 36% because of the spectral-temporal similarity with maize. The methodology works well for areas with balanced crop samples. Yet, it favors prevalent classes in areas with imbalanced crops at the expense of a low accuracy for the minority crops. Despite these shortcomings, the proposed methodology offers a viable option to generate crop samples in regions with few ground labels.

MODIS phenology-derived, multi-year distribution of conterminous U.S. crop types