Winter Wheat Classification by Multi-Temporal Optimized Image Analysis Based on Random Forest Algorithm

Wheat is considered as a strategic product in the world, and in Iran it is considered as a major source of food, so that for many years, one of the goals of the government has been to achieve self-sufficiency in providing this product.
The estimation of the winter wheat crop area and its spatial distribution in the country, during the plantation and growth period, has a vital role in the value assessment, storage planning, as well as import-export planning. Remote sensing classification techniques, based on one image, are widely used to this task. Remote sensing is considered as a suitable solution to overcome the problems of common and traditional inventory methods. However, drawback of these classical methods is that the winter wheat crop’s spectral signature is very similar to spectral signature of some other crops growing simultaneously (e.g. barley, alfalfa), which limits the classification performance. To improve the classification results’ accuracy, multi-temporal imagery acquired during the growth period are instead used. Classification algorithms based on ensemble classifiers are suitable tools to decrease the classification problems of time series images. One appropriate technique for coping with multi-temporal imagery-based classification problems is random forest classifier. In researches performed to determine the area under wheat cultivation, in most cases due to high similarity between wheat and barley, these two products are considered as a same class.
In this research, ten Landsat satellite images from Marvdasht-Fars province with cloudiness less than 20% were selected. Then radiometric and atmospheric corrections were applied to all images and the time series analysis was performed as follows to increase the accuracy of wheat classification:- Analysis and selection of optimum images in time series instead of using all images and decreasing volume of calculations and process time
- Improving and upgrading the classification results by producing new features in time series and finally surveying the improvement rate of final accuracy and increasing rate of capability of proposed method in separation of wheat from barley.
-Determining the most important input features and most suitable dates for imaging by calculation of variable importance in random forest algorithm.
-Early classification of wheat and separating it from other classes at least two months before product harvest.
The results of this research are as follows: Although the use of spectral bands data of time series images in a random forest algorithm, as input features of the model, increased the accuracy of the results compared to the use of single-image, but the use of these features alone, will not have results with high precision for the separation of wheat and barley (due to the high similarity of these two products). If we consider these two products as a single class, by this method it is possible to separate these two classes from the others in the region with overall accuracy of 89.5% and Kappa Coefficient 93.1%.
Using time series of spectral bands of images and vegetation indices and their difference in average increased overall accuracy and Kappa coefficient and barley producer’s accuracy, so that in case of using only three optimized and selected images by algorithm, barley producer’s accuracy increased and improved by 47%.
By examining the variable importance of the random forest algorithm, optimal images and the most important vegetation indices were recognized as the most effective features. The results showed that among the vegetation indices and their differences, the difference of STVI3 index and then the EVI and MSAVI for the images of 2015.4.17 and 2015.5.3 dates are the most important features. In addition, by surveying the sum of features used in this research, it was determined that the images of 2015.1.11, 2015.4.17 and 2015.5.3 have the most importance among images of time series.
Finally, through analysis of the importance of the developed features, we found that different vegetation indices and spectral gradient of multi-temporal images’ bands are the most important features for improving the classification results.
Since in this method, the stages of plant growth are surveyed in different times in some way and all extracted information from the images are analyzed as a whole, the suggested way for classification of other products can be generalized considering different cultivation times, peak greenness, their harvest and also for different regions.