Despite the remarkable progress made in agricultural industry in recent years, weed management is still a challenging and complex problem.
Even though mechanical weeding using inter-row cultivators is frequently used to remove the weeds between rows [1], the presence of weeds within crop rows and very closed to the main plant, makes it very difficult to eliminate intra-row weeds by mechanical cultivation implements where mechanical weed control in close proximity to the crop plant may damage the crop [2].
Traditional manual removal of intra-row weeds is still a common practice in the peanut fields. Hand hoeing is a very effective operation which is carried out properly. However, it is a tedious, extremely labor-intensive and time consuming operation with adverse health effects. Furthermore, manual weeding through hand tools can only be employed in small-scale farming or in home gardens and it is not a good practice in large scale cultivation [3, 4].
The most widespread method in weeds elimination is using herbicides. Conventionally, herbicides are uniformly applied to all parts of a farm even if there are no weeds in some certain parts of the farm. Increasing the cost of agriculture, environmental pollution, and negative effect on the human health are the major drawbacks associated with uniform application of herbicides [5, 6]. Site-specific weed management is an essential part of the progress towards economically and environmentally sustainable weed management [7]. Precision weed control within a field majorly requires information about weed distribution [8].
Mechanical destroying of intra-row weeds without damaging to crop plants, as well as variable application of chemicals based on the presence and intensity of weeds in different parts of fields, both primarily necessitate it to precisely detect and locate weeds in crop rows.
The acceptance and utilization of computer vision systems has been rapid and widespread. Along with other fields, the use of image processing techniques and image-extracted data for agro/food-informatics applications, have been widely investigated by researchers and reported in the literature [9–12]. Also, Studies on crop and weed detection using visible range image processing approaches has also been interested and conducted by many researchers [1, 13–16].
While there are literature reporting successful application of multi/hyper spectral image analysis for crop and weed detection [17–19], high quality non-visible spectrum imaging systems are generally more expensive and not-affordable to access – rather than visible ones – for researchers and farmers [20, 21]. Therefore, Optimizing the RGB image processing techniques for weed/crop detecting, is still a challenge.
Discrete Wavelet Transform (DWT) is a multiresolution image analysis tool which decomposes images into low and high frequency subbands by applying successive high-pass and low-pass filtering [22]. By implementing DWT on an image, approximation and details coefficients are obtained. Approximation coefficient subimages contain low frequencies which express the whole global trend of the image, whereas the detail coefficient subimages contain higher frequencies that express the local steep changes. Therefore, DWT is one of the powerful texture feature analysis techniques [23, 24]. Comprehensive information about wavelet transform and its applications in images, are provided by Vyas et al. [25] and Kolekar et al. [26]. Applications of DWT were reported for several agro-food related areas [11, 24, 27, 28].
Fuzzy logic theory, which was originally introduced by [29], is a problem solving tool that deals with approximate reasoning rather than fixed and exact reasoning [30]. Thus, Distinct from conventional techniques, fuzzy sets are capable of dealing with vague, ambiguous and imprecise data [31]. Fuzzy logic can help in the site-specific application of herbicides based on outputs from an image processing system, either in the real time or by the image-based weed maps [32].
Herrera et al. [33] applied a fuzzy decision-making method for discrimination between grasses and broad-leaved weeds based on shape descriptors. The best classification accuracy was reported to be 92.9%. An equal accuracy of fuzzy classifier was reported by Sujaritha et al. [34] for classification of sugarcane crop among nine different weed species based on leaf textures.
It should be noted that, in most cases, due to close similarity of weeds and main crop, very different image-based information must be investigated to find those set of features that satisfactorily differentiate between plant types. Furthermore, due to exponentially increasing the number of fuzzy logic rules by increasing the input variables, it is difficult to adjust fuzzy rules when there is a large amount of input variables [35]. Moreover, depending on the result of human adjustment, there is no guarantee to yield the optimal solution [36]. Therefore, if the use of fuzzy system for such a complex problem is intended, there is a need for invoking strategies to effectively incorporate numerous types of image-extracted features into the fuzzified weed detection system.
DTs are hierarchical classifiers that predict class membership by heuristically selecting the most relevant attributes and searching for if- then-else rules that best split the samples into correct classes [37, 38]. These automatically selected features and generated rules can be applied for constructing and tuning fuzzy classifier. DT-based fuzzy systems have been used for classification problems in agriculture-related studies [31, 39, 40].
Although good accuracies have been reported for weed/crop classification when applying several techniques [15, 41–43], however these researches don’t provide information about rules, membership functions, and antecedent, to be useful for developing a fuzzy weed detection system. Therefore, in the aim of this study was to investigate the capability of DT-based fuzzy logic system for classification of some broadleaf plants in peanut fields using image extracted colour and wavelet information.